A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs
Open accessAbstract
The era of large astronomical surveys generates massive image catalogs requiring efficient and secure access, particularly during pre-publication periods where data confidentiality and integrity are paramount. While Findable, Accessible, Interoperable, and Reusable (FAIR) principles guide the eventual public dissemination of data, traditional security methods for restricted phases often lack granularity or incur prohibitive performance penalties. To address this, we present a framework that integrates a flexible policy engine for fine-grained access control with a novel GPU-accelerated implementation of the AES-GCM authenticated encryption protocol. The novelty of this work lies in the adaptation and optimization of a parallel tree-reduction strategy to overcome the main performance bottleneck in authenticated encryption on GPUs: the inherently sequential Galois/Counter Mode (GCM) authentication hash (GHASH). We present both the algorithmic adaptation and its efficient execution on GPU architectures. Building on optimized GPU AES kernels from recent work in cryptographic acceleration, this work presents the first integration of these techniques into a high-throughput, FITS-aware encryption framework specifically designed for large-scale astronomical data, combining cryptographic authentication, dual-key access control, and direct compatibility with the standard astronomical Python ecosystem. Our implementation transforms the sequential GHASH computation into a highly parallelizable, logarithmic-time process, achieving authenticated encryption throughput suitable for petabyte-scale image analysis. Our solution provides a robust mechanism for data providers to enforce access policies, ensuring both confidentiality and integrity without hindering research workflows, thereby facilitating a secure and managed transition of data to public, FAIR archives.
AI snapshot
Written from the paper’s full text.
TL;DR
This paper presents GpuFitsCrypt, a C++/CUDA framework that combines policy-based access control with GPU-accelerated AES-128-GCM for massive FITS astronomical catalogs. It adapts a parallel tree-reduction strategy to overcome the sequential GHASH bottleneck, achieving high authenticated-encryption throughput while preserving compatibility with Astropy and cfitsio.
Key findings
- The parallel GHASH reduction transforms the authentication hash from O(N) to O(log N), enabling kernel times for a ~600 MB FITS image of 0.026 s on H100 PCIe, 0.042 s on A100, 0.039 s on L40S, 0.070 s on RTX 3090, 0.236 s on RTX 3060, 0.369 s on RTX 3050 Ti, and 0.542 s on Jetson Orin.
- GCM authentication adds a net slowdown of 1.72×–4.89× over unauthenticated CTR; for ~3.4 GB Extra-Large files, H100 PCIe shows Net GCM Cost 6.541 s versus Net CTR Cost 1.338 s (4.89×), A100 6.208 s versus 1.612 s (3.85×), L40S 6.230 s versus 1.595 s (3.91×), and RTX 3090 6.167 s versus 2.038 s (3.03×).
- GHASH-only throughput reaches 664 Gbps on L40S, 541 Gbps on H100 PCIe, 472 Gbps on RTX 3090, and 436 Gbps on A100-SXM4; on RTX 3090 this is ~1.91× higher than the 247 Gbps reported by Lee et al. (2025), and L40S exceeds the 536 Gbps reported for RTX 4090 by ~24%.
- Photometric validation on a real TTT exposure found 0 sources in the encrypted payload and an identical source catalog after decryption for all 182 detected sources, with FWHM ≈ 6.84 px and zero residuals.
Methods
- Design: a modular architecture separating a Control Plane (Policy Engine, Access Control Module, external KMS interface) from a Data Plane (GpuFitsCrypt C++/CUDA library) implementing AES-128-GCM, a two-HDU FITS-compatible authenticated container, dual Header/Data keys, AAD context binding, and verify-then-decrypt semantics.
- Data and sample: real and synthetic FITS images ranging from 16 MB to ~600 MB (Large) and 3.4 GB (Extra-Large); synthetic images were generated via Astropy with float32 pixel data drawn from a Gaussian distribution (µ = 1500, σ = 200) and standard WCS metadata. Hardware included H100 PCIe, A100-SXM4, L40S, RTX 3090, RTX 3060, RTX 3050 Ti, and Jetson Orin AGX.
- Analysis: benchmarks used 5 discarded warm-up runs followed by 200 timed repetitions, reporting medians and standard deviations; kernel times were measured with CUDA events, end-to-end times with monotonic host timers, and Net Cryptographic Cost was defined as E2E_GCM minus Astropy load time. Photometric validation used GPUPHOT on a Two-meter Twin Telescope exposure.
Limitations
- Stated in the paper: the implementation depends on NVIDIA CUDA, excluding AMD and Intel GPUs; pinned memory management requires careful host integration to avoid memory exhaustion in highly concurrent multi-user environments.
- Evident from the text: Extra-Large benchmarks exclude consumer and edge devices (RTX 3060, RTX 3050 Ti, Jetson Orin) because the 3.4 GB dataset exceeds their VRAM; the Large regime is latency-bound, so system overheads can mask datacenter GPU compute advantages; and the comparison with Lee et al. (2025) uses different input sizes (16 KB pre-staged messages versus ~600 MB real FITS payloads), as the authors note.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
15 of 37 cited sources checked · 6 supported · 1 flagged · accuracy not yet measured
- Premisecitation§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3
FAIR principles primarily apply once data enters the public domain.
“Alignment with Open Science Goals: While the community strongly moves towards Open Science and FAIR data principles (Wilkinson et al., 2016; Chen et al., 2022), facilitated by public archives (Green et al., 2024), these principles primarily apply once data enters the public domain.”
Quote in the paperSubstantial claim2 of 3 sources checked- cites Wilkinson et al., 2016 source
In “The FAIR Guiding Principles in detail” (page unavailable), the paper says sensitive data can achieve a high degree of FAIRness through published metadata even when the data itself is not published. It also says the principles can be adopted incrementally, contrary to the claim that they primarily apply only once data enters the public domain. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measured - cites Chen et al., 2022 source
The Introduction (p. 2) says the paper’s data-publication recommendations reflect FAIR principles, and the Data Archiving and Access section (p. 10) addresses making published data available. But the paper does not say FAIR principles primarily apply once data is public; its Summary (p. 13) expressly recommends related practices during preparation, submission, and review before publication. (LLM, conf=0.84)
Experimental — accuracy not yet measured - cites Green et al., 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation§ 1. Introduction · p. 1
Modern astronomy is generating petabyte-scale image catalogs.
“Modern astronomy, driven by large-scale surveys such as the Legacy Survey of Space and Time (LSST), is generating petabyte-scale image catalogs (Ivezi´ c et al., 2019; Berriman and Groom, 2011).”
Quote in the paperSubstantial claim1 of 2 sources checked- cites Ivezi´ c et al., 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Berriman and Groom, 2011 source
The paper reports that astronomy already had 1 PB of accessible public data and discusses discovering images across PB-sized data sets (Innovations in Serving and Discovering Data). But it describes LSST as a project yet to move into operation, and its examples of implemented image and catalog databases are TB-scale; it does not establish that LSST was already generating petabyte-scale image catalogs. (LLM, conf=0.86)
Experimental — accuracy not yet measured
- Premisecitation§ 1. Introduction · p. 2
AES-GCM is widely recommended for high-bandwidth applications because it combines the parallelizability of Counter Mode with a built-in integrity check.
“AES-GCM is widely recommended for such high-bandwidth applications because it combines the parallelizability of Counter Mode, essential for handling large payloads efficiently, with a built-in integrity check that detects data corruption without the overhead of a separate HMAC pass (McGrew and Viega, 2004; An and Seo, 2020).”
Quote in the paperSubstantial claim1 of 2 sources checked- cites McGrew and Viega, 2004
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites An and Seo, 2020 source
Section 3.2.4 (page unavailable) says CTR mode encrypts blocks in parallel and enables efficient GPU encryption; it also mentions Galois/Counter mode as a use of CTR. It does not say AES-GCM is widely recommended for high-bandwidth applications or describe a built-in integrity check that avoids a separate HMAC pass. (LLM, conf=0.86)
Experimental — accuracy not yet measured
- Findingcitation§ 3.2.1. Cryptographic Protocol: AES-GCM · p. 4
The framework implements the AES-128-GCM standard.
“The framework implements the AES-128-GCM standard (Dworkin, 2007).”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Dworkin, 2007 source
- Findingcitation§ 3.2.2. Authenticated FITS File Format · p. 5
To eliminate the risk of nonce reuse, the framework employs a stochastic generation approach compliant with NIST SP 800-38D guidelines.
“To eliminate the risk of nonce reuse, the framework employs a stochastic generation approach compliant with NIST SP 800-38D guidelines (Dworkin, 2007).”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Dworkin, 2007 source
Section 8.2 describes an approved random-bit-generator construction for GCM IVs, supporting the use of a stochastic approach under NIST SP 800-38D. But Section 8 defines the uniqueness requirement as a probability of nonce reuse no greater than 2^-32, not zero; the source does not support the claim that this approach eliminates the risk. (LLM, conf=0.90)
Experimental — accuracy not yet measured
- Premisecitation§ 3.2.2. Authenticated FITS File Format · p. 5
The security of AES-GCM mode relies strictly on the uniqueness of the initialization vector for a given key.
“The security of the AES-GCM mode relies strictly on the uniqueness of the initialization vector for a given key.”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Dworkin, 2007 source
- Interpretationcitation§ 5.1. Choice of Cryptographic Mode · p. 8
AES-128 is preferred over AES-256 for the threat model of scientific archives with embargo periods of months to a few years.
“AES-128 is preferred over AES-256 for the threat model of scientific archives with embargo periods of months to a few years: the best known cryptanalysis, the biclique attack of Bogdanov et al. (2011), requires approximately 2 126 operations, which remains computationally infeasible by orders of magnitude.”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Bogdanov et al. (2011) source
- Premisecitation§ 5.1. Choice of Cryptographic Mode · p. 8
Recent results (Malal and Tezcan, 2026) indicate that migration to WAES-256 / Rijndael-256 produces no observable slowdown (3,053.5 Gbps for WAES-256 versus 3,062.5 Gbps for AES-256 on RTX 4090, a difference of approximately 0.3%).
“The architectural choice does not preclude future migration to a wider block cipher: GpuFitsCrypt isolates the AES core behind a clear interface, and recent results (Malal and Tezcan, 2026) indicate that migration to WAES-256 / Rijndael-256 produces no observable slowdown (3,053.5 Gbps for WAES-256 versus 3,062.5 Gbps for AES-256 on RTX 4090, a difference of approximately 0.3%).”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Malal and Tezcan, 2026 source
- Findingreasoning§ 1. Introduction · p. 2
Python bindings integrate GpuFitsCrypt with Astropy's open() workflow, allowing users to access encrypted FITS files with the same idiomatic patterns used for unencrypted files when valid keys are available.
“Python bindings integrate GpuFitsCrypt with Astropy's open() workflow, allowing users to access encrypted FITS files with the same idiomatic patterns used for unencrypted files when valid keys are available (Table 1).”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
GpuFitsCrypt introduces a FITS-compatible authenticated container that is syntactically valid FITS.
“GpuFitsCrypt introduces a FITS-compatible authenticated container (Section 3) that is syntactically valid FITS, opens cleanly in Astropy and cfitsio without modification, and provides both cryptographic authentication through AES-GCM authentication tags and accidental-corruption detection through the FITS-standard CHECKSUM / DATASUM mechanism.”
Quote in the paperSubstantial claim - Findingreasoning§ Abstract · p. 1
The implementation transforms the sequential GHASH computation into a highly parallelizable, logarithmic-time process.
“Our implementation transforms the sequential GHASH computation into a highly parallelizable, logarithmic-time process, achieving authenticated encryption throughput suitable for petabyte-scale image analysis.”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
The encrypted pixel payload is encoded as a single-row, single-column BinTableHDU regardless of the original image dimensions.
“The encrypted pixel payload is encoded as a single-row, single-column BinTableHDU regardless of the original image dimensions, concealing geometry from unauthorized accessors while preserving scientific invariance.”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
The container provides accidental-corruption detection through the FITS-standard CHECKSUM/DATASUM mechanism.
“GpuFitsCrypt introduces a FITS-compatible authenticated container (Section 3) that is syntactically valid FITS, opens cleanly in Astropy and cfitsio without modification, and provides both cryptographic authentication through AES-GCM authentication tags and accidental-corruption detection through the FITS-standard CHECKSUM / DATASUM mechanism.”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
Critical metadata is cryptographically bound to the ciphertext via AES-GCM's Associated Authenticated Data field, preventing context-swapping attacks.
“Critical metadata is cryptographically bound to the ciphertext via AES-GCM's Associated Authenticated Data field, preventing context-swapping attacks; the framework further enforces verify-then-decrypt semantics, rejecting authentication failures before any plaintext is exposed.”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
The container opens cleanly in Astropy and cfitsio without modification.
“GpuFitsCrypt introduces a FITS-compatible authenticated container (Section 3) that is syntactically valid FITS, opens cleanly in Astropy and cfitsio without modification, and provides both cryptographic authentication through AES-GCM authentication tags and accidental-corruption detection through the FITS-standard CHECKSUM / DATASUM mechanism.”
Quote in the paperSubstantial claim - Findingreasoning§ 1. Introduction · p. 2
The container provides cryptographic authentication through AES-GCM authentication tags.
“GpuFitsCrypt introduces a FITS-compatible authenticated container (Section 3) that is syntactically valid FITS, opens cleanly in Astropy and cfitsio without modification, and provides both cryptographic authentication through AES-GCM authentication tags and accidental-corruption detection through the FITS-standard CHECKSUM / DATASUM mechanism.”
Quote in the paperSubstantial claim - Findingcitation§ 1. Introduction · p. 2
Following the strategy described by Lee et al. (2025), we transform the linear O(N) GHASH computation into a scalable, logarithmic-time O(log N) process.
“Following the strategy described by Lee et al. (2025), we transform the linear O ( N ) GHASH computation into a scalable, logarithmictime O (log N ) process.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Lee et al., 2025 source
- Findingreasoning§ 4.2. Policy Enforcement and Granular Access · p. 7
Abstract policy rules are translated into concrete cryptographic operations through a conditional key release mechanism.
“As depicted in the workflow diagram (Figure 1 in Section 3), abstract policy rules are translated into concrete cryptographic operations through a conditional key release mechanism.”
Quote in the paperSubstantial claim - Findingcitation§ 3.3.1. Parallel GHASH Reduction · p. 6
The implementation utilizes a parallel reduction strategy operating in GF(2^128), inspired by recent optimizations for parallel architectures.
“To address this, the implementation utilizes a parallel reduction strategy operating in the Galois Field GF (2 128 ), inspired by recent optimizations for parallel architectures (Lee et al., 2025).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Lee et al., 2025 source
- Findingnone given§ 3.2.3. Context Binding via Associated Data (AAD) · p. 6
A unique feature of the implementation is the use of GCM's Associated Data (AAD) capability to cryptographically bind the plaintext metadata to the encrypted pixels.
“A unique feature of the implementation is the use of GCM's Associated Data (AAD) capability to cryptographically bind the plaintext metadata to the encrypted pixels.”
Quote in the paperNot a substantial claim - Findingreasoning§ 3.2.2. Authenticated FITS File Format · p. 5
For larger archives reaching N ∼ 10^9 files per key, P ≈ 6.3 × 10^-12, providing an acceptable security margin for operational use.
“For larger archives reaching N ∼ 10 9 files per key, P ≈ 6 . 3 × 10 -12 , providing an acceptable security margin for operational use.”
Quote in the paperSubstantial claim - Findingown data§ 4.2. Policy Enforcement and Granular Access · p. 8
A 'Public User' is initially denied access.
“As shown in Table 2, a 'Public User' is initially denied access.”
Quote in the paperNot a substantial claim - Findingreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6
Instead of a linear dependency chain, GPU threads compute partial hashes concurrently, multiplying by the appropriate power of H to 'jump' gaps in the sequence.
“Instead of a linear dependency chain, GPU threads compute partial hashes concurrently, multiplying by the appropriate power of H to”
Quote in the paperSubstantial claim - Findingnone given§ 3.3.2. Asynchronous Streams Pipeline · p. 7
The library implements an asynchronous pipeline using NVIDIA CUDA streams to overlap host-device communication with computation.
“The library implements an asynchronous pipeline using NVIDIA CUDA streams to overlap host-device communication with computation:”
Quote in the paperSubstantial claim - Findingnone given§ 3.2.2. Authenticated FITS File Format · p. 5
The 96-bit nonces are generated utilizing the host system's entropy pool (e.g., via /dev/urandom).
“The 96-bit nonces are generated utilizing the host system's entropy pool (e.g., via /dev/urandom ).”
Quote in the paperSubstantial claim - Findingreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8
The Policy Engine supports time-based rules, essential for automating data embargoes.
“A critical feature of the Policy Engine is its support for time-based rules, essential for automating data embargoes.”
Quote in the paperNot a substantial claim- Rests on an assumption: Missions and observatories such as JWST, Euclid, ALMA, and LSST enforce specific data rights policies that include proprietary periods ranging from months to years.
- Findingcitation§ 3.2.2. Authenticated FITS File Format · p. 5
Given the 96-bit namespace, the probability of a collision is bounded by the birthday formula P ≈ N^2 / (2 · 2^96).
“Given the 96bit namespace, the probability of a collision is bounded by the birthday formula P ≈ N 2 / (2 · 2 96 ), where N is the number of nonces generated under a single key and the 96-bit nonce length is specified by NIST SP 800-38D for AES-GCM.”
Quote in the paperSubstantial claim - Findingreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6
The library precomputes a table of powers of the hash key H (H^{2^0}, H^{2^1}, H^{2^2}, ..., H^{2^k}).
“The library precomputes a table of powers of the hash key H ( H 2 0 , H 2 1 , H 2 2 , . . . , H 2 k ).”
Quote in the paperSubstantial claim - Findingreasoning§ 3.2.2. Authenticated FITS File Format · p. 5
For LSST-scale workloads (N ∼ 10^7 images per year, single key), this gives P ≈ 6.3 × 10^-16 per year, negligible by any operational standard.
“For LSSTscale workloads ( N ∼ 10 7 images per year, single key), this gives P ≈ 6 . 3 × 10 -16 per year, negligible by any operational standard.”
Quote in the paperSubstantial claim - Findingreasoning§ 3.2.2. Authenticated FITS File Format · p. 4
The encrypted pixel payload is stored as a single-row, single-column binary blob (TFORM='<N>B') in a BinTableHDU extension.
“The encrypted pixel payload is stored as a single-row, singlecolumn binary blob ( TFORM='<N>B' ) in a BinTableHDU extension; this is the standard FITS mechanism for embedding arbitrary binary content, and is semantically correct because the bytes are not, by construction, an image.”
Quote in the paperSubstantial claim - Findingreasoning§ 5.4. Results and Analysis · p. 10
Across all architectures, GCM introduces a slowdown factor ranging from 1.72x to 4.89x compared to unauthenticated encryption.
“Across all architectures, GCM introduces a slowdown factor ranging from 1 . 72x to 4 . 89x compared to unauthenticated encryption.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
The GPU kernel execution time is a fraction of the total End-to-End time, which is dominated by system overheads and I/O.
“The GPU kernel execution time (blue) is a fraction of the total End-to-End time, which is dominated by system overheads and I / O (red).”
Quote in the paperSubstantial claim - Findingown data§ 4.2. Policy Enforcement and Granular Access · p. 8
Empirical validation of all four access scenarios on a synthetic 256 × 256 float32 FITS file with representative WCS metadata confirms that each scenario produces a syntactically valid FITS output (full content, zero-filled payload, or minimal header).
“Empirical validation of all four access scenarios on a synthetic 256 × 256 float32 FITS file with representative WCS metadata confirms that each scenario produces a syntactically valid FITS output (full content, zero-filled payload, or minimal header), opens cleanly in Astropy without warnings, and is bit-exact to the original in the Full Access case.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
In the latency-bound Large regime, the variance in the Astropy I/O baseline dominates the Net GCM Cost.
“In the latency-bound Large regime, the variance in the Astropy I / O baseline dominates the Net GCM Cost ( E2E -Astropy ), which can occasionally mask the raw compute advantage of datacenter architectures like the H100 when compared to high-bandwidth consumer cards like the RTX 3090; Table 3 confirms the expected hardware ordering in pure-kernel work.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 12
When applied to the encrypted payload, the source extraction algorithm failed to converge, detecting zero sources due to the complete destruction of spatial structure.
“When applied to the encrypted payload (Panel b), the source extraction algorithm failed to converge, detecting zero sources due to the complete destruction of spatial structure.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
The Net GCM Cost values cluster within a narrow range (6.2-6.5 s).
“In the Extra-Large regime, the expected hardware ordering is confirmed for unauthenticated throughput: the H100 PCIe achieves the lowest Net CTR Cost (1.34 s) across all benchmarked GPUs, and the Net GCM Cost values cluster within a narrow range (6.2-6.5 s), confirming that GpuFitsCrypt scales effectively with data volume.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
For files in the hundreds of megabytes range, the process is latency-bound.
“The results highlight a critical architectural insight: for files in the hundreds of megabytes range, the process is latencybound.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
Table 3 confirms the expected hardware ordering in pure-kernel work.
“In the latency-bound Large regime, the variance in the Astropy I / O baseline dominates the Net GCM Cost ( E2E -Astropy ), which can occasionally mask the raw compute advantage of datacenter architectures like the H100 when compared to high-bandwidth consumer cards like the RTX 3090; Table 3 confirms the expected hardware ordering in pure-kernel work.”
Quote in the paperNot a substantial claim - Findingown data§ 5.4. Results and Analysis · p. 11
In the Extra-Large regime, the expected hardware ordering is confirmed for unauthenticated throughput.
“In the Extra-Large regime, the expected hardware ordering is confirmed for unauthenticated throughput: the H100 PCIe achieves the lowest Net CTR Cost (1.34 s) across all benchmarked GPUs, and the Net GCM Cost values cluster within a narrow range (6.2-6.5 s), confirming that GpuFitsCrypt scales effectively with data volume.”
Quote in the paperNot a substantial claim - Findingown data§ 4.2. Policy Enforcement and Granular Access · p. 8
Each of the four access scenarios opens cleanly in Astropy without warnings.
“Empirical validation of all four access scenarios on a synthetic 256 × 256 float32 FITS file with representative WCS metadata confirms that each scenario produces a syntactically valid FITS output (full content, zero-filled payload, or minimal header), opens cleanly in Astropy without warnings, and is bit-exact to the original in the Full Access case.”
Quote in the paperSubstantial claim - Findingreasoning§ 5.4. Results and Analysis · p. 10
The H100 PCIe ratio increases from 3.40x in the Large regime to 4.89x in the Extra-Large regime.
“The H100 PCIe ratio increases from 3 . 40x in the Large regime to 4 . 89x in the Extra-Large regime, reflecting the transition from an I / O-dominated to a computedominated regime as the dataset size approaches the cryptographic kernel's saturation point.”
Quote in the paperNot a substantial claim - Findingown data§ 4.2. Policy Enforcement and Granular Access · p. 8
The Full Access case is bit-exact to the original.
“Empirical validation of all four access scenarios on a synthetic 256 × 256 float32 FITS file with representative WCS metadata confirms that each scenario produces a syntactically valid FITS output (full content, zero-filled payload, or minimal header), opens cleanly in Astropy without warnings, and is bit-exact to the original in the Full Access case.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 10
Performance is sensitive to the Bit-Sliced AES parameters: Thread Size (TS) and Repeat Blocks (R).
“Performance is sensitive to the Bit-Sliced AES parameters: Thread Size (TS) and Repeat Blocks (R).”
Quote in the paperNot a substantial claim - Findingown data§ 6.1. Latency vs. Throughput in Astronomical Pipelines · p. 12
For interactive analysis of individual images (e.g., a 600 MB file), consumer-grade GPUs like the RTX 3090 or RTX 3060 provide performance indistinguishable from datacenter hardware, as the process is bound by I/O latency.
“For interactive analysis of individual images (e.g., a researcher opening a 600 MB file), consumer-grade GPUs like the RTX 3090 or even the RTX 3060 provide performance indistinguishable from datacenter hardware, as the process is bound by I / O latency.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 10
There is parity between high-end datacenter GPUs (H100) and consumer hardware (RTX 3060) for the 600 MB dataset.
“Acounter-intuitive observation from Table 3 is the parity between high-end datacenter GPUs (H100) and consumer hardware (RTX 3060) for the 600 MB dataset.”
Quote in the paperSubstantial claim - Findingown data§ 5.4. Results and Analysis · p. 12
The decrypted output yielded a source catalog identical to the original input.
“Conversely, the decrypted output (Panel c) yielded a source catalog identical to the original input.”
Quote in the paperSubstantial claim - Premisecitation§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2
The vast majority of astronomical image data adheres to the FITS standard.
“The vast majority of astronomical image data, from groundbased telescopes to space missions, adheres to the Flexible Image Transport System (FITS) standard (Greisen et al., 1980; Wells et al., 1981; Ponz et al., 1994; Iau Fits Working Group, 2016).”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites Greisen et al., 1980 source
- cites Wells et al., 1981
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Ponz et al., 1994
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation§ 1. Introduction · p. 2
The standard GHASH algorithm is inherently sequential, making it ill-suited for the massive parallelism of GPUs.
“Consequently, the standard GHASH algorithm is inherently sequential, making it ill-suited for the massive parallelism of Graphics Processing Units (GPUs) (Manavski, 2007; Lee et al., 2025).”
Quote in the paperSubstantial claim0 of 2 sources checked - Interpretationreasoning§ Abstract · p. 1
The novelty of this work lies in adapting and optimizing a parallel tree-reduction strategy to overcome the main performance bottleneck in authenticated encryption on GPUs: the sequential GHASH.
“The novelty of this work lies in the adaptation and optimization of a parallel tree-reduction strategy to overcome the main performance bottleneck in authenticated encryption on GPUs: the inherently sequential Galois / Counter Mode (GCM) authentication hash (GHASH).”
Quote in the paperSubstantial claim - Interpretationreasoning§ 1. Introduction · p. 2
This enables authenticated encryption at throughputs matching the I/O capabilities of modern storage systems.
“This enables authenticated encryption at throughputs matching the I / O capabilities of modern storage systems.”
Quote in the paperSubstantial claim- Rests on an assumption: Traditional cryptographic measures impose significant computational overhead (latency and reduced bandwidth), creating unacceptable bottlenecks for data-intensive analysis.
- Interpretationreasoning§ Abstract · p. 1
The implementation achieves authenticated encryption throughput suitable for petabyte-scale image analysis.
“Our implementation transforms the sequential GHASH computation into a highly parallelizable, logarithmic-time process, achieving authenticated encryption throughput suitable for petabyte-scale image analysis.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 1. Introduction · p. 2
This encoding conceals geometry from unauthorized accessors.
“The encrypted pixel payload is encoded as a single-row, single-column BinTableHDU regardless of the original image dimensions, concealing geometry from unauthorized accessors while preserving scientific invariance.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 4.2. Policy Enforcement and Granular Access · p. 7
The framework's ability to enforce fine-grained access control is predicated on the dual-key design of the encrypted FITS format, orchestrated by the architecture's Control Plane.
“The framework's ability to enforce fine-grained access control is predicated on the dual-key design of the encrypted FITS format, orchestrated by the architecture's Control Plane.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8
A collaborator granted the Header Key can decrypt the ENCHDR keyword, enabling analysis of observational telemetry without accessing the scientific pixel data.
“A collaborator may be granted the Header Key to decrypt the ENCHDR keyword, enabling analysis of observational telemetry without accessing the scientific pixel data.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8
The library implements an automatic fallback mechanism for blind analysis scenarios where metadata is restricted.
“Conversely, to support blind analysis scenarios where metadata is restricted, the library implements an automatic fallback mechanism.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8
The model allows decoupling metadata access from pixel data access.
“Furthermore, the model allows decoupling metadata access from pixel data access.”
Quote in the paperSubstantial claim - Study design and conductnone given§ 3.3.2. Asynchronous Streams Pipeline · p. 7
The library enforces a strict validation gate: the output buffer is only marked as valid if the locally computed GHASH matches the AUTHTAG stored in the file.
“Validation Gate: The library enforces a strict validation gate. The output buffer is only marked as valid if the locally computed GHASH matches the AUTHTAG stored in the file.”
Quote in the paperSubstantial claim - Interpretationnone given§ 3.2.2. Authenticated FITS File Format · p. 5
This dual-key design allows a Policy Engine to grant access to the header without releasing the key for the scientific pixel data.
“This dual-key design allows a Policy Engine to grant access to the header (e.g., for query validation) without releasing the key for the scientific pixel data.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6
This enables the input data to be split into independent blocks and combined in a hierarchical manner.
“As illustrated in Figure 3-b, this enables the input data to be split into independent blocks and combined in a hierarchical manner.”
Quote in the paperSubstantial claim - Study design and conductreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8
A mismatch in the Data Key or integrity failure triggers a payload suppression mechanism: the library returns a structurally valid FITS object where the scientific pixel array is zero-filled while the authenticated header remains fully accessible.
“Consequently, a mismatch in the Data Key (or integrity failure) triggers a payload suppression mechanism: the library returns a structurally valid FITS object where the scientific pixel array is zero-filled (physically suppressed via memset ), while the authenticated header remains fully accessible.”
Quote in the paperSubstantial claim- Rests on an assumption: AES-GCM ensures data integrity through compartmentalized verification.
- Interpretationreasoning§ 7. Conclusions · p. 13
This approach successfully overcomes the sequential bottleneck of the GHASH function.
“This approach successfully overcomes the sequential bottleneck of the GHASH function, achieving end-to-end decryption throughputs that saturate standard storage interfaces.”
Quote in the paperNot a substantial claim - Interpretationnone given§ 6.3. Limitations · p. 13
The primary limitation of the current implementation is its dependency on the NVIDIA CUDA platform.
“The primary limitation of the current implementation is its dependency on the NVIDIA CUDA platform.”
Quote in the paperNot a substantial claim - Interpretationnone given§ 5.1. Choice of Cryptographic Mode · p. 8
The primary challenge addressed in this work is the efficient parallelization of GCM's authentication component, GHASH, which is inherently sequential.
“The primary challenge addressed in this work is the efficient parallelization of GCM's authentication component, GHASH, which is inherently sequential.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12
It decouples the security policy from the data itself.
“It decouples the security policy from the data itself.”
Quote in the paperSubstantial claim - Interpretationcitation§ 5.2. GPU Implementation: Parallel GHASH Architecture · p. 8
Building upon the high-performance bit-sliced AES-CTR core architecture described by Lee et al. (2024), we introduced a critical modification to ensure cryptographic interoperability across heterogeneous hardware.
“Building upon the high-performance bit-sliced AESCTR core architecture described by Lee et al. (2024), we introduced a critical modification to ensure cryptographic interoperability across heterogeneous hardware.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Lee et al. (2024) source
- Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12
The framework's reliance on an external Key Management Service (KMS) is a deliberate architectural choice.
“Our framework's reliance on an external Key Management Service (KMS) is a deliberate architectural choice.”
Quote in the paperSubstantial claim - Premisecitation§ 5.4. Results and Analysis · p. 11
The most direct comparison is with Lee et al. (2025), who measure GHASH-only throughput using 16 KB pre-staged message blocks, a scope equivalent to our GHashKernelOnly measurement.
“The most direct comparison is with Lee et al. (2025), who measure GHASH-only throughput using 16 KB pre-staged message blocks, a scope equivalent to our GHashKernelOnly measurement.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Lee et al. (2025) source
- Interpretationreasoning§ 6.1. Latency vs. Throughput in Astronomical Pipelines · p. 12
The value of Datacenter architectures (H100, A100) lies in their capacity for massive concurrency.
“However, the value of Datacenter architectures (H100, A100) lies in their capacity for massive concurrency.”
Quote in the paperSubstantial claim - Interpretationcitation§ 5.1. Choice of Cryptographic Mode · p. 8
The architectural choice does not preclude future migration to a wider block cipher: GpuFitsCrypt isolates the AES core behind a clear interface.
“The architectural choice does not preclude future migration to a wider block cipher: GpuFitsCrypt isolates the AES core behind a clear interface, and recent results (Malal and Tezcan, 2026) indicate that migration to WAES-256 / Rijndael-256 produces no observable slowdown (3,053.5 Gbps for WAES-256 versus 3,062.5 Gbps for AES-256 on RTX 4090, a difference of approximately 0.3%).”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12
Access to a dataset can be revoked instantly by updating the policy in the control plane, without needing to re-encrypt the petabytes of stored data.
“Unlike passwordprotected ZIP files or static GPG keys, access to a dataset can be revoked instantly by updating the policy in the control plane, without needing to re-encrypt the petabytes of stored data.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 6.1. Latency vs. Throughput in Astronomical Pipelines · p. 12
This lowers the barrier to entry for secure data access on standard workstations.
“This is a positive finding, as it lowers the barrier to entry for secure data access on standard workstations.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 5.1. Choice of Cryptographic Mode · p. 8
We chose AES-GCM over CTR + HMAC for three reasons specific to the GpuFitsCrypt deployment context.
“An alternative encrypt-then-MAC construction would combine AES-CTR with HMAC-SHA256; we chose AES-GCM over this option for three reasons specific to the GpuFitsCrypt deployment context.”
Quote in the paperSubstantial claim- Rests on an assumption: AES-GCM ensures data integrity through compartmentalized verification.
- Rests on an assumption: CTR mode lacks intrinsic integrity verification, leaving scientific data vulnerable to ciphertext manipulation unless paired with a separate MAC.
- Interpretationreasoning§ 5.2.1. Parallel Tree-Reduction Strategy · p. 9
In each step of the tree reduction, threads combine partial results using the precomputed powers, reducing the complexity from linear O(N) to logarithmic O(log N).
“In each step, threads combine partial results using the precomputed powers, reducing the complexity from linear O ( N ) to logarithmic O (log N ).”
Quote in the paperSubstantial claim - Interpretationown data§ 5.4. Results and Analysis · p. 11
GpuFitsCrypt scales effectively with data volume.
“In the Extra-Large regime, the expected hardware ordering is confirmed for unauthenticated throughput: the H100 PCIe achieves the lowest Net CTR Cost (1.34 s) across all benchmarked GPUs, and the Net GCM Cost values cluster within a narrow range (6.2-6.5 s), confirming that GpuFitsCrypt scales effectively with data volume.”
Quote in the paperSubstantial claim - Interpretationown data§ 5.4. Results and Analysis · p. 11
This variance can occasionally mask the raw compute advantage of datacenter architectures like the H100 when compared to high-bandwidth consumer cards like the RTX 3090.
“In the latency-bound Large regime, the variance in the Astropy I / O baseline dominates the Net GCM Cost ( E2E -Astropy ), which can occasionally mask the raw compute advantage of datacenter architectures like the H100 when compared to high-bandwidth consumer cards like the RTX 3090; Table 3 confirms the expected hardware ordering in pure-kernel work.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 5.4. Results and Analysis · p. 10
This slowdown reflects the unavoidable arithmetic cost of the finite field multiplications required for GHASH.
“This reflects the unavoidable arithmetic cost of the finite field multiplications required for GHASH.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12
The two mechanisms therefore protect orthogonal threat surfaces.
“The two mechanisms therefore protect orthogonal threat surfaces, consistent with the layered security model outlined in Section 2.”
Quote in the paperSubstantial claim- Rests on an assumption: Transport-layer protocols such as SSH and TLS protect data in transit but not after the file has been downloaded, replicated, or staged on cold storage.
- Interpretationreasoning§ 7. Conclusions · p. 13
It achieves end-to-end decryption throughputs that saturate standard storage interfaces.
“This approach successfully overcomes the sequential bottleneck of the GHASH function, achieving end-to-end decryption throughputs that saturate standard storage interfaces.”
Quote in the paperNot a substantial claim - Interpretationnone given§ 7. Conclusions · p. 13
GpuFitsCrypt implements a novel parallel tree-reduction algorithm to accelerate AES-128-GCM on GPUs.
“At its core is GpuFitsCrypt , a library that implements a novel parallel treereduction algorithm to accelerate AES-128-GCM on GPUs.”
Quote in the paperSubstantial claim - Interpretationown data§ 5.4. Results and Analysis · p. 11
For moderate file sizes, the workflow is dominated by system overheads rather than compute capacity.
“This observation explains why a consumer-grade RTX 3060 can appear competitive with an H100 for single-image processing: for moderate file sizes, the workflow is dominated by system overheads rather than compute capacity.”
Quote in the paperSubstantial claim - Interpretationnone given§ 5.2. GPU Implementation: Parallel GHASH Architecture · p. 8
One of the core innovations of this work lies in transforming the serial GHASH dependency chain into a parallelizable workload suitable for massive datasets.
“One of the core innovations of this work lies in transforming the serial GHASH dependency chain (Equation 1) into a parallelizable workload suitable for massive datasets.”
Quote in the paperNot a substantial claim - Interpretationown data§ 6. Discussion and Future Work · p. 12
The historical trade-off between rigorous security and high-performance data access in astronomy is no longer an insurmountable barrier.
“The results presented in this study demonstrate that the historical trade-off between rigorous security and highperformance data access in astronomy is no longer an insurmountable barrier.”
Quote in the paperSubstantial claim - Interpretationown data§ 5.4. Results and Analysis · p. 10
This increase reflects the transition from an I/O-dominated to a compute-dominated regime as the dataset size approaches the cryptographic kernel's saturation point.
“The H100 PCIe ratio increases from 3 . 40x in the Large regime to 4 . 89x in the Extra-Large regime, reflecting the transition from an I / O-dominated to a computedominated regime as the dataset size approaches the cryptographic kernel's saturation point.”
Quote in the paperNot a substantial claim - Study design and conductreasoning§ 5.2. GPU Implementation: Parallel GHASH Architecture · p. 8
We refactored the kernel to enforce a deterministic, geometry-agnostic mapping between data blocks and AES counters.
“We refactored the kernel to enforce a deterministic, geometry-agnostic mapping between data blocks and AES counters.”
Quote in the paperSubstantial claim - Interpretationown data§ 5.4. Results and Analysis · p. 10
For single-image processing of this size, the workload operates in a latency-bound regime.
“This phenomenon indicates that for single-image processing of this size, the workload operates in a latency-bound regime.”
Quote in the paperSubstantial claim - Study design and conductnone given§ 4.2. Policy Enforcement and Granular Access · p. 8
In the implementation, the integrity checks for metadata and pixel data are decoupled.
“In our implementation, the integrity checks for metadata and pixel data are decoupled.”
Quote in the paperSubstantial claim - Interpretationown data§ 6.1. Latency vs. Throughput in Astronomical Pipelines · p. 12
The scalability demonstrated in the Extra-Large benchmarks suggests that these architectures can handle petabyte-scale encryption without becoming a bottleneck.
“The scalability demonstrated in the Extra-Large benchmarks suggests that these architectures can handle petabyte-scale encryption without becoming a bottleneck.”
Quote in the paperNot a substantial claim - Premisecitation§ 5.2. GPU Implementation: Parallel GHASH Architecture · p. 8
The original algorithm's counter generation logic was coupled to the GPU grid geometry (total thread count), causing the keystream to vary with hardware configuration.
“The original algorithm's counter generation logic was coupled to the GPU grid geometry (total thread count), causing the keystream to vary with hardware configuration.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Lee et al. (2024) source
- Interpretationreasoning§ 5.2.1. Parallel Tree-Reduction Strategy · p. 9
The precalculated table of hash key powers allows any data block in the sequence to be processed independently, effectively bridging the positional gap without serial dependencies.
“This precalculated table allows any data block in the sequence to be processed independently, effectively bridging the positional gap without serial dependencies.”
Quote in the paperSubstantial claim - Interpretationown data§ 6. Discussion and Future Work · p. 12
It is possible to implement authenticated encryption (AES-GCM) at speeds that match the I/O capabilities of standard storage infrastructure.
“By leveraging the parallel processing capabilities of modern GPUs, we have shown that it is possible to implement authenticated encryption (AES-GCM) at speeds that match the I / O capabilities of standard storage infrastructure.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 5.1. Choice of Cryptographic Mode · p. 8
We selected AES-128-GCM as the cryptographic standard.
“We selected AES-128-GCM (Galois / Counter Mode) as the cryptographic standard.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12
Object-layer encryption complements rather than replaces transport-layer protection.
“Object-layer encryption complements rather than replaces transport-layer protection.”
Quote in the paperNot a substantial claim
- Classified as probably not a substantial claim. 58
- May combine two claims. 64
- The quote only partly states this claim. 1
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2Header sections store essential metadata as keyword/value pairs, including observational parameters, processing history, WCS information, and flux calibration.
“Critically, the header sections store essential metadata as keyword / value pairs. This includes observational parameters (telescope telemetry, instrument configuration, exposure time, filters), processing history (calibration steps, software versions), and arguably most importantly, the World Coordinate System (WCS) information (Greisen and Calabretta, 2002; Calabretta and Greisen, 2002) and flux”
Quote in the paperNot a substantial claim2 of 2 sources checked- cites Greisen and Calabretta, 2002 source
The Introduction describes FITS headers as keyword-equals-value records and says this paper defines WCS keywords. Section 5.4 gives an exposure-time keyword as an example. The cited paper does not establish the full list of header contents attributed to it, particularly processing history, software versions, and flux calibration. Page numbers are unavailable in our copy. (LLM, conf=0.88)
Experimental — accuracy not yet measured - cites Calabretta and Greisen, 2002 source
Classified as probably not a substantial claim. May combine two claims.
Premisecitation§ 2.1. Characteristics of Astronomical Image Catalogs · p. 3Archives routinely host millions of such files, pushing total data volumes into the petabyte and exabyte scales.
“Archives routinely host millions of such files, pushing total data volumes into the petabyte and exabyte scales (Berriman and Groom, 2011).”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Berriman and Groom, 2011 source
The opening discussion (page not specified) says that 1 PB of public astronomy data was accessible and projects more than 60 PB of archived data by 2020, supporting the petabyte-scale part. Our copy does not say that archives routinely host millions of files or that their holdings reach exabyte scale; its reference to millions in the Innovations in Serving and Discovering Data section (p. 55) counts queries per year, not files. (LLM, conf=0.88)
Experimental — accuracy not yet measured
May combine two claims.
Findingcitation§ 3.2.2. Authenticated FITS File Format · p. 5Archives approaching N ∼ 2^32 files per key fall into the regime where NIST SP 800-38D §8.3 recommends key rotation; the framework supports this through the explicit key reference required by every encryption operation.
“Archives approaching N ∼ 2 32 files per key fall into the regime where NIST SP 800-38D §8.3 recommends key rotation; our framework supports this through the explicit key reference required by every encryption operation.”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Dworkin, 2007 source
Section 8.3 limits authenticated-encryption invocations to 2^32 per key for the specified IV constructions, so changing keys before that limit is a reasonable operational implication. It does not state a limit in files, and the 2^32 cap does not apply to implementations using only deterministically generated 96-bit IVs. The statement about the citing authors’ framework is their own, not a claim attributed to NIST. (LLM, conf=0.88)
Experimental — accuracy not yet measured
May combine two claims.
Premisecitation§ 5.1. Choice of Cryptographic Mode · p. 8AES-GCM is specified by NIST SP 800-38D (Dworkin, 2007) as an approved AEAD mode, whereas CTR + HMAC, although built from approved primitives, is not itself a standardized AEAD construction.
“Third, standardization: AES-GCM is specified by NIST SP 800-38D (Dworkin, 2007) as an approved AEAD mode, whereas CTR + HMAC, although built from approved primitives, is not itself a standardized AEAD construction.”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Dworkin, 2007 source
The Introduction specifies GCM for authenticated encryption with associated data and identifies AES as a block cipher on which it can be built, supporting the AES-GCM portion of the claim. The paper does not address whether CTR + HMAC is a standardized AEAD construction or establish the approval status of both of its primitives. (LLM, conf=0.88)
Experimental — accuracy not yet measured
May combine two claims.
Premisecitation§ 5.1. Choice of Cryptographic Mode · p. 8The best known cryptanalysis, the biclique attack of Bogdanov et al. (2011), requires approximately 2^126 operations, which remains computationally infeasible by orders of magnitude.
“AES-128 is preferred over AES-256 for the threat model of scientific archives with embargo periods of months to a few years: the best known cryptanalysis, the biclique attack of Bogdanov et al. (2011), requires approximately 2 126 operations, which remains computationally infeasible by orders of magnitude.”
Quote in the paperSubstantial claim1 of 1 sources checked- cites Bogdanov et al. (2011) source
May combine two claims.
Findingnone given§ 1. Introduction · p. 2A dual-key access policy mechanism manages Header Keys and Data Keys independently, enabling four distinct access scenarios that decouple metadata visibility from data confidentiality and support embargo-period workflows.
“A dual-key access policy mechanism manages Header Keys and Data Keys independently, enabling four distinct access scenarios that decouple metadata visibility from data confidentiality and support embargo-period workflows (Section 4, Table 2).”
Quote in the paperSubstantial claim- Rests on an assumption: Missions and observatories such as JWST, Euclid, ALMA, and LSST enforce specific data rights policies that include proprietary periods ranging from months to years.
Classified as probably not a substantial claim. May combine two claims.
Findingreasoning§ 3.3.2. Asynchronous Streams Pipeline · p. 7streamD manages the heavy lifting for the pixel data, reading the encrypted BLOB directly into pinned memory and executing the GCM pipeline.
“Data Pipeline: streamD manages the heavy lifting for the pixel data. The host reads the encrypted BLOB directly into pinned memory. The stream then executes the GCM pipeline:”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Findingnone given§ 3. Framework Architecture · p. 4The architecture strictly separates the concerns of policy enforcement (Control Plane) from high-performance cryptographic execution (Data Plane).
“The architecture strictly separates the concerns of policy enforcement (Control Plane) from high-performance cryptographic execution (Data Plane).”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Findingnone given§ 3.2.3. Context Binding via Associated Data (AAD) · p. 6The library constructs the AAD vector for the data payload by concatenating the data Nonce with the string representation of the critical structural keywords.
“The library constructs the AAD vector for the data payload by concatenating the data Nonce with the string representation of the critical structural keywords:”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based).
Findingnone given§ 4.1. User Interaction and Transparency Mechanism · p. 7This memory buffer can be passed directly to standard scientific libraries like Astropy using memory-efficient methods such as HDUList.fromstring(), treating the decrypted data as a virtual file.
“This memory buffer can be passed directly to standard scientific libraries like Astropy using memory-efficient methods such as HDUList.fromstring() , treating the decrypted data as a virtual file.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ 4.1. User Interaction and Transparency Mechanism · p. 7Upon success, the library returns a FitsOperationResult structure containing a pointer to a fully decrypted FITS file in memory.
“Upon success, the library returns a FitsOperationResult structure containing a pointer to a fully decrypted FITS file in memory.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Findingnone given§ 3.3.2. Asynchronous Streams Pipeline · p. 7The library maintains a persistent context with two concurrent CUDA streams: streamH (Header) and streamD (Data), along with Pinned Memory host buffers to maximize PCIe bandwidth.
“Dual-Stream Context: The library maintains a persistent context with two concurrent CUDA streams: streamH (Header) and streamD (Data), along with Pinned Memory (page-locked) host buffers to maximize PCIe bandwidth.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ AAD Data = NONCE D ∥ Serialize(ORIG\_BITPIX) ∥ Serialize(ORIG\_NAXIS) ∥ . . . · p. 6The library exposes a C-style API designed for integration with high-level languages like Python.
“The library exposes a Cstyle API designed for integration with high-level languages like Python.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ 4.1. User Interaction and Transparency Mechanism · p. 7The implementation provides a C-style API designed to be wrapped by higher-level languages commonly used in astronomy, such as Python.
“The implementation provides a C-style API (refer to Table 1) designed to be wrapped by higher-level languages commonly used in astronomy, such as Python.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6This reduces the computational complexity from linear time O(N) to logarithmic time O(log N), effectively matching the throughput of the counter-mode encryption component.
“This reduces the computational complexity from linear time O ( N ) to logarithmic time O (log N ), effectively matching the throughput of the countermode encryption component.”
Quote in the paperSubstantial claimMay combine two claims.
Findingnone given§ 3.2.2. Authenticated FITS File Format · p. 4Both the encrypted container and all decryption outputs are opened by Astropy and cfitsio without warnings or errors, recognizing the structure as syntactically valid FITS files.
“Both the encrypted container and all decryption outputs are opened by Astropy and cfitsio without warnings or errors, recognizing the structure as syntactically valid FITS files.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim. May combine two claims.
Findingnone given§ 3.1. Conceptual Components: The Control Plane · p. 4The ACM serves as the API gateway and orchestrator, handling authentication, requesting policy decisions, and interfacing with the KMS, and isolates keys from the end-user.
“Access Control Module: The ACM serves as the Application Programming Interface (API) gateway and orchestrator. It handles authentication, requests policy decisions, and interfaces with the KMS. Crucially, it isolates the keys from the end-user; the user never sees the raw key, only the resulting data stream.”
Quote in the paperSubstantial claim- The authors state this as an assumption: The framework assumes the existence of a secure, external Key Management Service (KMS) for key provisioning.
Classified as probably not a substantial claim.
Findingnone given§ 3.1. Conceptual Components: The Control Plane · p. 4The framework relies on three conceptual components to manage authorization and key distribution: Policy Engine, Access Control Module, and Key Management Service Interface.
“The framework relies on three conceptual components to manage authorization and key distribution, ensuring that cryptographic operations are policy-driven rather than ad-hoc.”
Quote in the paperSubstantial claim- The authors state this as an assumption: The framework assumes the existence of a secure, external Key Management Service (KMS) for key provisioning.
May combine two claims.
Findingreasoning§ 3.3.2. Asynchronous Streams Pipeline · p. 7streamH handles the ENCHDR decryption, providing low-latency access to metadata.
“Header Processing: streamH handles the ENCHDR decryption. Since headers are typically small ( < 80 kB), this provides low-latency access to metadata.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Findingreasoning§ 3.2.2. Authenticated FITS File Format · p. 4The original FITS file is transformed into a two-HDU structure: a Primary HDU metadata container and a BINTABLE extension data container.
“To support granular access while maintaining the integrity of the file structure, a FITS-compatible authenticated container is employed (see Figure 2). The original FITS file is transformed into a two-HDU structure:”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Findingnone given§ AAD Data = NONCE D ∥ Serialize(ORIG\_BITPIX) ∥ Serialize(ORIG\_NAXIS) ∥ . . . · p. 6This AAD is fed into the GHASH computation during both encryption and decryption.
“This AAD is fed into the GHASH computation during both encryption and decryption.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 5.4. Results and Analysis · p. 12We compared critical parameters for all 182 detected sources, including centroid positions, FWHM ≈ 6.84 px, and instrumental flux, finding zero residuals.
“We compared critical parameters for all 182 detected sources, including centroid positions, Full Width at Half Maximum (FWHM ≈ 6 . 84 px), and instrumental flux, finding zero residuals.”
Quote in the paperSubstantial claimMay combine two claims.
Findingcitation§ 5.4. Results and Analysis · p. 11On Ada Lovelace hardware, the L40S reaches 664 Gbps, exceeding the 536 Gbps reported for the RTX 4090 by approximately 24%.
“On Ada Lovelace hardware, the L40S reaches 664 Gbps, exceeding the 536 Gbps reported for the RTX 4090 by approximately 24%.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Lee et al. (2025) source
May combine two claims.
Findingown data§ 5.4. Results and Analysis · p. 11On the RTX 3090, GpuFitsCrypt achieves 472 Gbps versus 247 Gbps reported by Lee et al., a factor of approximately 1.91× higher.
“On the RTX 3090, GpuFitsCrypt achieves 472 Gbps versus 247 Gbps reported by Lee et al., a factor of approximately 1.91 × higher.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Lee et al. (2025) source
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Findingnone given§ Appendix A.1.1. Integration Workflow · p. 14In partial-access mode the pixel data is zero-filled: it is safe to read but contains no signal.
“# Pixel data is zero-filled; safe to read but contains no signal”
Quote in the paperNot a substantial claimThe quote only partly states this claim.
Findingown data§ 5.4. Results and Analysis · p. 11The H100 PCIe achieves the lowest Net CTR Cost (1.34 s) across all benchmarked GPUs.
“In the Extra-Large regime, the expected hardware ordering is confirmed for unauthenticated throughput: the H100 PCIe achieves the lowest Net CTR Cost (1.34 s) across all benchmarked GPUs, and the Net GCM Cost values cluster within a narrow range (6.2-6.5 s), confirming that GpuFitsCrypt scales effectively with data volume.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingnone given§ Appendix A.1.1. Integration Workflow · p. 14The solution is transparent to the scientist: the complex GPU operations are abstracted and the decrypted data is handed off to Astropy via zero-copy memory buffers.
“It highlights the transparency of the solution: the complex GPU operations are abstracted, and the decrypted data is handed off to Astropy via zero-copy memory buffers.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisenone given§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2The WCS maps pixel coordinates to celestial coordinates, while flux calibration enables conversion of pixel values to physical units.
“The WCS maps pixel coordinates to celestial coordinates (e.g., Right Ascension and Declination), while flux calibration enables the conversion of pixel values to physical units.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Premisenone given§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Operating system filesystem permissions are insufficiently expressive to capture complex scientific access policies, and managing them across petabyte-scale archives is administratively unscalable.
“Operating System Filesystem Permissions: Standard Portable Operating System Interface permissions are insufficiently expressive to capture complex scientific access policies (e.g., time-based embargoes, role-based access tied to collaboration status). Managing these permissions across petabyte-scale archives with millions of files is administratively unscalable.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationnone given§ Abstract · p. 1The framework integrates a flexible policy engine for fine-grained access control with a GPU-accelerated implementation of the AES-GCM authenticated encryption protocol.
“we present a framework that integrates a flexible policy engine for fine-grained access control with a novel GPU-accelerated implementation of the AES-GCM authenticated encryption protocol.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based).
Premisenone given§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Due to perceived performance overheads and complexity of key management, end-to-end encryption of large scientific data archives is not yet common practice.
“Lack of Widespread Encryption: Due to perceived performance overheads and the complexity of key management, end-to-end encryption of large scientific data archives is not yet common practice.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 1. Introduction · p. 1The BICEP2/Planck controversy highlighted risks of utilizing unvalidated, preliminary data products.
“Similarly, the risks of utilizing unvalidated, preliminary data products were highlighted by the BICEP2 / Planck controversy (Ade et al., 2014, 2015).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Ade et al., 2014, 2015
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisenone given§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2FITS provides a structure based on Header/Data Units (HDUs), recognizing distinct roles of metadata and pixel arrays.
“FITS provides a crucial structure based on Header / -Data Units (HDUs), recognizing the distinct yet complementary roles of metadata and pixel arrays in scientific analysis.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisenone given§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2Typically, a FITS file contains a primary HDU holding the main image array and its associated descriptive header.
“Typically, a FITS file contains a primary HDU holding the main image array (the core scientific data, e.g., pixel values) and its associated descriptive header.”
Quote in the paperNot a substantial claimMay combine two claims.
Premisereasoning§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Perimeter security (VPNs, firewalls) offers coarse-grained access and does not provide fine-grained control over specific datasets or protect data once downloaded.
“Perimeter Security (VPNs, Firewalls): While useful for network-level protection, these offer coarse-grained (”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ Abstract · p. 1This work presents the first integration of optimized GPU AES kernels into a high-throughput, FITS-aware encryption framework for large-scale astronomical data, combining cryptographic authentication, dual-key access control, and compatibility with the standard astronomical Python ecosystem.
“Building on optimized GPU AES kernels from recent work in cryptographic acceleration, this work presents the first integration of these techniques into a high-throughput, FITS-aware encryption framework specifically designed for large-scale astronomical data, combining cryptographic authentication, dual-key access control, and direct compatibility with the standard astronomical Python ecosystem.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Premisereasoning§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Data is often stored unencrypted, leaving it vulnerable if storage systems are compromised or physical media is lost.
“Data is often stored unencrypted, leaving it vulnerable if storage systems are compromised or physical media is lost.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationnone given§ 1. Introduction · p. 2This encoding preserves scientific invariance.
“The encrypted pixel payload is encoded as a single-row, single-column BinTableHDU regardless of the original image dimensions, concealing geometry from unauthorized accessors while preserving scientific invariance.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable).
Premisereasoning§ 2.1. Characteristics of Astronomical Image Catalogs · p. 2This rich metadata is often as scientifically valuable as the pixel data itself.
“This rich metadata is often as scientifically valuable as the pixel data itself, enabling data discovery and context, thus necessitating careful consideration in any security framework regarding how both components are accessed and protected.”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductnone given§ 1. Introduction · p. 2The framework enforces verify-then-decrypt semantics, rejecting authentication failures before any plaintext is exposed.
“Critical metadata is cryptographically bound to the ciphertext via AES-GCM's Associated Authenticated Data field, preventing context-swapping attacks; the framework further enforces verify-then-decrypt semantics, rejecting authentication failures before any plaintext is exposed.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 1. Introduction · p. 1The 2022 cyberattack on ALMA forced a suspension of scientific operations and data delivery.
“More recently, the vulnerability of astronomical infrastructure was underscored by the 2022 cyberattack on the Atacama Large Millimeter / submillimeter Array (ALMA), which forced a suspension of scientific operations and data delivery, demonstrating that major facilities are targets for direct cyber threats (National Radio Astronomy Observatory, 2022).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites National Radio Astronomy Observatory, 2022
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
May combine two claims.
Interpretationreasoning§ Abstract · p. 1The solution provides a robust mechanism for data providers to enforce access policies, ensuring both confidentiality and integrity without hindering research workflows.
“Our solution provides a robust mechanism for data providers to enforce access policies, ensuring both confidentiality and integrity without hindering research workflows, thereby facilitating a secure and managed transition of data to public, FAIR archives.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Premisenone given§ 1. Introduction · p. 2Unlike CTR mode, AES-GCM provides AEAD, ensuring both confidentiality and data integrity.
“Unlike simpler modes, such as Counter (CTR) that provide only confidentiality , AES-GCM provides Authenticated Encryption with Associated Data (AEAD), ensuring both confidentiality and data integrity.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Premisenone given§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Web portal authentication often lacks granular authorization based on data properties or user roles, and data downloaded through such portals is typically unprotected thereafter.
“Web Portal Authentication: Basic username / password access to data portals often lacks granular authorization based on data properties or user roles beyond the initial login. Furthermore, data downloaded through such portals is typically unprotected thereafter (”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 1. Introduction · p. 2The primary technical challenge of this approach, and the core contribution of our work, is overcoming the performance bottleneck inherent in AES-GCM's authentication component: the GHASH function.
“The primary technical challenge of this approach, and the core contribution of our work, is overcoming the performance bottleneck inherent in AES-GCM's authentication component: the GHASH function.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 4.1. User Interaction and Transparency Mechanism · p. 7This approach achieves a high degree of transparency; the analysis code interacts with standard data objects, remaining agnostic to the underlying cryptographic operations.
“This approach achieves a high degree of transparency; the analysis code interacts with standard data objects, remaining agnostic to the underlying cryptographic operations.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationreasoning§ 4. Secure and Transparent Data Access · p. 7The framework provides robust security and granular access control without imposing a significant burden on the end-user.
“The primary goal of the framework is to provide robust security and granular access control without imposing a significant burden on the end-user.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Study design and conductnone given§ 3. Framework Architecture · p. 4The data access workflow follows a strict 'verify-then-decrypt' orchestration.
“The data access workflow follows a strict 'verify-thendecrypt' orchestration.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ AAD Data = NONCE D ∥ Serialize(ORIG\_BITPIX) ∥ Serialize(ORIG\_NAXIS) ∥ . . . · p. 6This prevents 'context swapping attacks,' ensuring the scientific data cannot be presented with falsified telemetry.
“This prevents 'context swapping attacks,' ensuring the scientific data cannot be presented with falsified telemetry.”
Quote in the paperSubstantial claimMay combine two claims.
Study design and conductreasoning§ 3.3.2. Asynchronous Streams Pipeline · p. 7If verification fails, the operation returns a distinct integrity error code, ensuring no corrupted or manipulated data is released to the user.
“If verification fails, the operation returns a distinct integrity error code, ensuring no corrupted or manipulated data is released to the user.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Interpretationreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6This sequence corresponds to the stride length at each level of a binary reduction tree.
“This sequence corresponds to the stride length at each level of a binary reduction tree.”
Quote in the paperSubstantial claimMay combine two claims.
Premisereasoning§ 3.2.1. Cryptographic Protocol: AES-GCM · p. 4GCM is an AEAD mode that generates a 128-bit Authentication Tag guaranteeing both confidentiality of pixel data and integrity of the file.
“Unlike the CTR mode used in early prototypes which provided only confidentiality, GCM is an AEAD mode. It generates a 128-bit Authentication Tag that guarantees both the confidentiality of the pixel data and the integrity of the file.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 3.2.1. Cryptographic Protocol: AES-GCM · p. 4Any unauthorized modification to the encrypted payload or metadata is immediately detected, resulting in a decryption failure.
“This ensures that any unauthorized modification to the encrypted payload or the metadata is immediately detected (resulting in a decryption failure), a critical requirement for scientific data provenance and validation.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based).
Premisenone given§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3Existing mechanisms often lack dedicated solutions to robustly manage transition phases, embargo periods, collaborative analysis, and pre-publication review.
“Existing mechanisms often lack dedicated solutions to robustly manage the transition phases , embargo periods, collaborative analysis, pre-publication review, where controlled, secure access is temporarily necessary before full public release.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Premisenone given§ 3.3.1. Parallel GHASH Reduction · p. 6The standard GHASH algorithm operates via a serial chain defined by the recurrence relation.
“The standard GHASH algorithm operates via a serial chain defined by the recurrence relation:”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationnone given§ 4. Secure and Transparent Data Access · p. 7This is achieved through a combination of a transparency mechanism at the application layer and a strict policy enforcement workflow.
“This is achieved through a combination of a transparency mechanism at the application layer and a strict policy enforcement workflow.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationreasoning§ 2.3. Limitations of Existing Security Approaches and the Role of FAIR Principles · p. 3These limitations highlight the need for a dedicated framework that bridges the gap, integrating strong, policy-driven access control directly with performant cryptographic protection at the data level.
“These limitations highlight the need for a dedicated framework that bridges the gap, integrating strong, policy-driven access control directly with performant cryptographic protection at the data level.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 3.3.1. Parallel GHASH Reduction · p. 6This serial dependency implies that the processing of block i cannot strictly begin until block i-1 is complete, making it typically ill-suited for the massive parallelism of GPUs.
“This serial dependency implies that the processing of block i cannot strictly begin until block i -1 is complete (see Figure 3a), making it typically ill-suited for the massive parallelism of GPUs.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationnone given§ 3.3. GPU Acceleration Strategy · p. 6Achieving high throughput with AES-GCM on large files requires masking the latency of memory transfers and overcoming the inherently serial nature of the authentication phase.
“Achieving high throughput with AES-GCM on large files requires masking the latency of memory transfers and overcoming the inherently serial nature of the authentication phase.”
Quote in the paperNot a substantial claim- Rests on an assumption: Traditional cryptographic measures impose significant computational overhead (latency and reduced bandwidth), creating unacceptable bottlenecks for data-intensive analysis.
Classified as probably not a substantial claim. May combine two claims.
Study design and conductreasoning§ 3. Framework Architecture · p. 4The Cryptographic Module executes high-throughput, GPU-accelerated decryption and integrity verification, returning plaintext FITS data only if both identity authorization and cryptographic signature verification succeed.
“This module then executes the high-throughput, GPUaccelerated decryption and integrity verification, returning the plaintext FITS data to the application only if both the identity authorization and the cryptographic signature verification succeed.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Interpretationnone given§ 3.2. The Cryptographic Module: GpuFitsCrypt · p. 4The core technical contribution of this work is the implementation of the data plane: a high-performance C++/CUDA library named GpuFitsCrypt.
“The core technical contribution of this work is the implementation of the data plane: a high-performance C ++/ CUDA library named GpuFitsCrypt .”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationcitation§ 5.4. Results and Analysis · p. 10Our results are consistent with this performance gap, validating the architectural choice of accelerator-based cryptography for massive datasets.
“Our results are consistent with this performance gap, validating the architectural choice of accelerator-based cryptography for massive datasets.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 7. Conclusions · p. 13Our architecture enables fine-grained, policy-based access control while guaranteeing both confidentiality and integrity of scientific data.
“Our architecture enables fine-grained, policy-based access control (e.g., role-based, temporal embargoes) while guaranteeing both the confidentiality and integrity of scientific data.”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductreasoning§ 5.2.3. Optimized Memory Pipeline · p. 9Encryption/decryption and authentication are managed within a stream pipeline that overlaps with disk I/O operations.
“Dual-Stream Overlap: As detailed in Section 3, encryption / decryption (AES-CTR) and authentication (GHASH) are managed within a stream pipeline that overlaps with disk I / O operations.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 6.1. Latency vs. Throughput in Astronomical Pipelines · p. 12While a single file does not saturate these cards, their massive VRAM and compute throughput allow for the processing of multiple streams simultaneously.
“While a single file does not saturate these cards, their massive VRAM and compute throughput allow for the processing of multiple streams simultaneously, a scenario typical of observatory pipelines reducing thousands of exposures per night.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Interpretationreasoning§ 5.4. Results and Analysis · p. 11These differences reflect our bit-sliced implementation, which operates on ≈600 MB real FITS payloads rather than synthetic 16 KB blocks, and may also reflect architectural and clock-speed differences between GPU models.
“These differences reflect our bit-sliced implementation, which operates on ≈ 600 MB real FITS payloads rather than synthetic 16 KB blocks, and may also reflect architectural and clock-speed differences between GPU models.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 5.2.2. Scalable MapReduce for Massive Files · p. 9Since a single CUDA block cannot process an arbitrary number of data blocks due to shared memory limits, we employ a two-phase MapReduce architecture to handle large FITS files.
“Since a single CUDA block cannot process an arbitrary number of data blocks due to shared memory limits, we employ a two-phase MapReduce architecture to handle large FITS files (e.g., 600 MB to 3 GB):”
Quote in the paperSubstantial claim- Rests on an assumption: A single CUDA block cannot process an arbitrary number of data blocks due to shared memory limits.
May combine two claims.
Interpretationreasoning§ 5.2. GPU Implementation: Parallel GHASH Architecture · p. 8This ensures that a FITS file encrypted on an edge device (e.g., Jetson Nano) can be correctly decrypted on a datacenter GPU (e.g., NVIDIA A100), regardless of the parallelism degree, without compromising the high-throughput performance.
“This ensures that a FITS file encrypted on an edge device (e.g., Jetson Nano) can be correctly decrypted on a datacenter GPU (e.g., NVIDIA A100), regardless of the parallelism degree, without compromising the highthroughput performance.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 5.4. Results and Analysis · p. 10Recent state-of-the-art benchmarks for fully parallelized AES-GCM on GPUs (Lee et al., 2025) demonstrate a speedup factor of ≈15× compared to highly optimized multi-threaded CPU implementations.
“Recent state-ofthe-art benchmarks for fully parallelized AES-GCM on GPUs (Lee et al., 2025) demonstrate a speedup factor of ≈ 15 × compared to highly optimized multi-threaded CPU implementations (Intel Core i7 with AES-NI instructions using OpenSSL).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Lee et al., 2025 source
Classified as probably not a substantial claim.
Study design and conductnone given§ Appendix A.1.2. Benchmarking Methodology · p. 15The net cryptographic cost is isolated by subtracting the baseline I/O latency from the end-to-end median latency.
“net_crypto_cost = median_e2e - baseline_io_time”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductreasoning§ Appendix A.1.2. Benchmarking Methodology · p. 15Discarded warm-up runs are executed before timing to stabilize the GPU state.
“# Stabilizes GPU clocks and JIT compilation.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 5.1. Choice of Cryptographic Mode · p. 8CTR + HMAC has no native AAD construct, and replicating this guarantee with a custom HMAC scope would require an ad-hoc, non-standardized construction.
“Second, native AAD binding: GpuFitsCrypt cryptographically binds the FITS metadata (size, geometry, exposure parameters) to the encrypted pixel payload through GCM's Associated Authenticated Data field, preventing context-swapping attacks; CTR + HMAC has no native AAD construct, and replicating this guarantee with a custom HMAC scope would require an ad-hoc, non-standardized construction.”
Quote in the paperSubstantial claim- Rests on an assumption: CTR + HMAC has no native AAD construct.
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning§ 7. Conclusions · p. 13This work provides a path forward for data centers to enforce rigorous security during proprietary periods without hindering the pace of discovery.
“This work provides a path forward for data centers to enforce rigorous security during proprietary periods without hindering the pace of discovery, ultimately supporting a trusted and robust transition of data into the public domain.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone given§ Appendix A.1.2. Benchmarking Methodology · p. 15Benchmarks were executed on a diverse array of physical hosts, ranging from edge devices to HPC nodes.
“Benchmarks were executed on a diverse array of physical hosts, ranging from edge devices to HPC nodes.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 4.2. Policy Enforcement and Granular Access · p. 8The fallback ensures that even with restricted permissions, the user receives a parsable FITS file, preserving compatibility with standard tools while keeping the core scientific content secure.
“This ensures that even with restricted permissions, the user receives a parsable FITS file, preserving compatibility with standard tools while keeping the core scientific content secure.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationnone given§ 5.4. Results and Analysis · p. 11GpuFitsCrypt extends beyond standalone cryptographic kernels by integrating full AES-GCM authenticated encryption with AAD binding and direct integration with the standard astronomical Python ecosystem, producing FITS-compatible encrypted containers that open cleanly in unmodified scientific tools.
“While these raw throughput metrics provide an essential baseline, GpuFitsCrypt extends beyond standalone cryptographic kernels by integrating full AESGCM authenticated encryption with Associated Authenticated Data binding and direct integration with the standard astronomical Python ecosystem (Astropy / cfitsio), producing FITScompatible encrypted containers that open cleanly in unmodified scienti”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Interpretationreasoning§ 6.3. Limitations · p. 13This dependency covers a vast majority of HPC and workstation environments in astronomy, but excludes hardware from other vendors (AMD, Intel GPUs).
“While this covers a vast majority of HPC and workstation environments in astronomy, it excludes hardware from other vendors (AMD, Intel GPUs).”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 5.2.2. Scalable MapReduce for Massive Files · p. 9The two-phase MapReduce architecture prevents the latency of transferring intermediate results back to the CPU, keeping the entire authentication pipeline on the device.
“This architecture prevents the latency of transferring intermediate results back to the CPU, keeping the entire authentication pipeline on the device.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12The CHECKSUM/DATASUM convention and the AES-GCM authentication tags address orthogonal threats: the former detects accidental corruption with broad ecosystem compatibility, while the latter detects intentional manipulation with cryptographic strength.
“The CHECKSUM / DATASUM convention written by the framework (see Section 3) and the AES-GCM authentication tags address orthogonal threats: the former detects accidental corruption (bit rot, transmission errors, partial-write failures) with broad ecosystem compatibility, while the latter detects intentional manipulation with cryptographic strength.”
Quote in the paperSubstantial claim- Rests on an assumption: The CHECKSUM/DATASUM convention detects accidental corruption (bit rot, transmission errors, partial-write failures) with broad ecosystem compatibility.
- Rests on an assumption: AES-GCM authentication tags detect intentional manipulation with cryptographic strength.
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone given§ 6.3. Limitations · p. 13The requirement for pinned memory management requires careful integration into host applications to avoid memory exhaustion in highly concurrent multi-user environments.
“Additionally, the requirement for pinned memory management, while essential for performance, requires careful integration into host applications to avoid memory exhaustion in highly concurrent multi-user environments.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 5.4. Results and Analysis · p. 10The same study highlights a power efficiency advantage of over 20× for the GPU architecture.
“Furthermore, the same study highlights a power efficiency advantage of over 20 × for the GPU architecture.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Lee et al., 2025 source
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation§ 5.1. Choice of Cryptographic Mode · p. 8This is also the choice made by Lee et al. (2024) and Lee et al. (2025), whose implementation we adopt as our cryptographic core.
“This is also the choice made by Lee et al. (2024) and Lee et al. (2025), whose implementation we adopt as our cryptographic core.”
Quote in the paperNot a substantial claim0 of 2 sources checkedProbably not a substantial claim: may not pass S4 (based).
Interpretationnone given§ 7. Conclusions · p. 13By binding metadata context to the encrypted payload, we prevent tampering attacks that could compromise scientific validity.
“By binding metadata context to the encrypted payload, we prevent tampering attacks that could compromise scientific validity.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 5.4. Results and Analysis · p. 10Prior optimized AES-CTR implementations have achieved approximately 878 Gbps on consumer hardware (Tezcan, 2021).
“Within the GPU tier, prior optimized AES-CTR implementations have achieved approximately 878 Gbps on consumer hardware (Tezcan, 2021), providing additional context for the GPU cryptographic performance envelope demonstrated here.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Tezcan, 2021 source
Classified as probably not a substantial claim. May combine two claims.
Study design and conductown data§ 5.3.1. Experimental Protocol · p. 10For every hardware and file size configuration, 5 warm-up runs were discarded followed by 200 timed repetitions, and reported values represent the median with standard deviations.
“Sampling Strategy: For every hardware and file size configuration, we performed 5 warm-up runs (discarded) to stabilize GPU clock states and JIT compilation effects, followed by 200 timed repetitions. Reported values represent the median of these executions, accompanied by standard deviations to quantify variability.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12GpuFitsCrypt operates at the object layer: each FITS file is independently encrypted and authenticated, and remains so for the lifetime of the file regardless of where or how it is stored.
“GpuFitsCrypt operates at the object layer: each FITS file is independently encrypted and authenticated, and remains so for the lifetime of the file regardless of where or how it is stored.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductnone given§ Appendix A.1.2. Benchmarking Methodology · p. 15The baseline I/O measurement uses memmap=False to force actual disk I/O, ensuring a fair comparison against the decryption pipeline.
“# memmap=False forces actual disk I/O, ensuring a fair”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 5.4. Results and Analysis · p. 10The massive parallelism of the H100 is underutilized, and the total time is dominated by constant system overheads rather than raw arithmetic throughput.
“The massive parallelism of the H100 is underutilized, and the total time is dominated by constant system overheads (PCIe latency, kernel launch time, and I / O initialization) rather than raw arithmetic throughput.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductown data§ Appendix A.1.2. Benchmarking Methodology · p. 15Benchmarks were conducted using a suite of real and synthetic FITS images, with synthetic images generated via Astropy to mimic real scientific observations where specific sizes were needed.
“Benchmarks were conducted using a suite of real and synthetic FITS images. Synthetic images were generated via Astropy to mimic real scientific observations where specific sizes were needed.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationreasoning§ 5.1. Choice of Cryptographic Mode · p. 8We consider the migration path straightforward when NIST completes the standardization of Rijndael-256 GCM / GMAC currently under public comment.
“We therefore consider the migration path straightforward when NIST completes the standardization of Rijndael-256 GCM / GMAC currently under public comment.”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductreasoning§ 5.1. Choice of Cryptographic Mode · p. 8GpuFitsCrypt cryptographically binds the FITS metadata (size, geometry, exposure parameters) to the encrypted pixel payload through GCM's Associated Authenticated Data field, preventing context-swapping attacks.
“Second, native AAD binding: GpuFitsCrypt cryptographically binds the FITS metadata (size, geometry, exposure parameters) to the encrypted pixel payload through GCM's Associated Authenticated Data field, preventing context-swapping attacks; CTR + HMAC has no native AAD construct, and replicating this guarantee with a custom HMAC scope would require an ad-hoc, non-standardized construction.”
Quote in the paperSubstantial claim- Rests on an assumption: GCM's Associated Authenticated Data field prevents context-swapping attacks.
Classified as probably not a substantial claim. May combine two claims.
Study design and conductreasoning§ 5.3.1. Experimental Protocol · p. 9The CUDA context was initialized once per session and reused across all runs, ensuring that measurements reflect the steady-state performance of the pipeline, excluding the one-time overhead of driver initialization and context creation.
“Execution Context: The CUDA context, including pinned memory allocation and stream creation, was initialized once per session and reused across all runs. This ensures that measurements reflect the steady-state performance of the pipeline, excluding the one-time overhead of driver initialization and context creation.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim. May combine two claims.
Study design and conductnone given§ 5.2.3. Optimized Memory Pipeline · p. 9All host buffers use page-locked memory to maximize PCIe bandwidth and enable Direct Memory Access.
“Pinned Memory I / O: All host buffers use page-locked memory ( cudaMallocHost ) to maximize PCIe bandwidth and enable Direct Memory Access.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 6.2. Security Implications of the Architecture · p. 12GpuFitsCrypt integrates both mechanisms simultaneously, ensuring comprehensive integrity verification.
“To provide robust protection against both threat vectors, GpuFitsCrypt integrates both mechanisms simultaneously, ensuring comprehensive integrity verification.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim. May combine two claims.
Study design and conductnone given§ Appendix A.1.2. Benchmarking Methodology · p. 15Context initialization is performed once per session and its overhead is excluded from the measured latency.
“# Allocates pinned memory and CUDA streams. Overhead excluded.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductown data§ 5.3.1. Experimental Protocol · p. 10GPU kernel execution times were measured using high-precision cudaEventRecord pairs placed strictly around the cryptographic kernels, excluding host-device memory transfers and auxiliary operations.
“Timing Mechanisms: GPU kernel execution times were measured using high-precision cudaEventRecord pairs on the device stream. To isolate raw computational throughput accurately, these markers were placed strictly around the cryptographic kernels, explicitly excluding host-device memory transfers and auxiliary operations.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationown data§ 4.2. Policy Enforcement and Granular Access · p. 8The payload suppression mechanism combined with the automatic header fallback ensures that granular access policies are enforced without crashing the read operation or exposing the application to corrupted content, maintaining compatibility with standard FITS viewers even in restricted access scenarios.
“This mechanism, combined with the automatic header fallback described above, ensures that granular access policies are enforced without crashing the read operation or exposing the application to corrupted content, maintaining compatibility with standard FITS viewers even in restricted access scenarios.”
Quote in the paperSubstantial claim- Rests on an assumption: AES-GCM ensures data integrity through compartmentalized verification.
May combine two claims.
Interpretationreasoning§ 5.4. Results and Analysis · p. 10This transition is the expected behavior for an architecture optimized for high-throughput compute, and it is consistent across the four GPUs benchmarked in the Extra-Large regime (range: 3.03x to 4.89x).
“This transition is the expected behavior for an architecture optimized for high-throughput compute, and it is consistent across the four GPUs benchmarked in the Extra-Large regime (range: 3 . 03x to 4 . 89x).”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 5.1. Choice of Cryptographic Mode · p. 8CTR + HMAC requires two sequential passes over the data (encryption then MAC), which doubles memory traffic on memory-bandwidth-bound GPU architectures, whereas AES-GCM performs encryption and authentication in a single fused pass.
“First, single-pass throughput: CTR + HMAC requires two sequential passes over the data (encryption then MAC), which doubles memory traffic on memory-bandwidth-bound GPU architectures, whereas AES-GCM performs encryption and authentication in a single fused pass.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationown data§ 5.4. Results and Analysis · p. 12The framework introduces no numerical artifacts, preserving the exact radiometric properties required for high-precision astronomy.
“This confirms that the framework introduces no numerical artifacts, preserving the exact radiometric properties required for high-precision astronomy.”
Quote in the paperSubstantial claim
The authors state this as an assumption.
The framework assumes the existence of a secure, external Key Management Service (KMS) for key provisioning.
“Our framework assumes the existence of a secure, external Key Management Service (KMS) for key provisioning, focusing strictly on the cryptographic performance and policy enforcement layer.”
“The architecture assumes the existence of a secure, industry-standard KMS (e.g., HashiCorp Vault, AWS KMS), accessed over a TLS / mTLS-secured channel for all ACM-to-KMS interactions.”
The paper relies on this without support.
Traditional cryptographic measures impose significant computational overhead (latency and reduced bandwidth), creating unacceptable bottlenecks for data-intensive analysis.
“Traditional cryptographic measures can impose significant computational overhead (latency and reduced bandwidth), creating unacceptable bottlenecks for dataintensive analysis.”
The paper relies on this without support.
Missions and observatories such as JWST, Euclid, ALMA, and LSST enforce specific data rights policies that include proprietary periods ranging from months to years.
“Missions and observatories such as the James Webb Space Telescope, Euclid, ALMA, and LSST enforce specific data rights policies that include proprietary periods ranging from months to years.”
The paper relies on this without support.
AES-GCM ensures data integrity through compartmentalized verification.
“Crucially, unlike confidentiality-only modes such as AESCTR, the use of AES-GCM ensures data integrity through compartmentalized verification.”
The paper relies on this without support.
CTR mode lacks intrinsic integrity verification, leaving scientific data vulnerable to ciphertext manipulation unless paired with a separate MAC.
“While Counter (CTR) mode offers high parallelizability and confidentiality, it lacks intrinsic integrity verification, leaving scientific data vulnerable to ciphertext manipulation (e.g., bitflipping attacks) unless paired with a separate Message Authentication Code (MAC).”
The paper relies on this without support.
Transport-layer protocols such as SSH and TLS protect data in transit but not after the file has been downloaded, replicated, or staged on cold storage.
“Transport-layer protocols such as SSH and TLS protect data in transit between an archive endpoint and a user, but not after the file has been downloaded, replicated to a backup, or staged on a cold-storage tier.”
The paper relies on this without support.
A single CUDA block cannot process an arbitrary number of data blocks due to shared memory limits.
“Since a single CUDA block cannot process an arbitrary number of data blocks due to shared memory limits, we employ a two-phase MapReduce architecture to handle large FITS files (e.g., 600 MB to 3 GB):”
The paper relies on this without support.
The CHECKSUM/DATASUM convention detects accidental corruption (bit rot, transmission errors, partial-write failures) with broad ecosystem compatibility.
“The CHECKSUM / DATASUM convention written by the framework (see Section 3) and the AES-GCM authentication tags address orthogonal threats: the former detects accidental corruption (bit rot, transmission errors, partial-write failures) with broad ecosystem compatibility, while the latter detects intentional manipulation with cryptographic strength.”
The paper relies on this without support.
AES-GCM authentication tags detect intentional manipulation with cryptographic strength.
“The CHECKSUM / DATASUM convention written by the framework (see Section 3) and the AES-GCM authentication tags address orthogonal threats: the former detects accidental corruption (bit rot, transmission errors, partial-write failures) with broad ecosystem compatibility, while the latter detects intentional manipulation with cryptographic strength.”
The paper relies on this without support.
GCM's Associated Authenticated Data field prevents context-swapping attacks.
“Second, native AAD binding: GpuFitsCrypt cryptographically binds the FITS metadata (size, geometry, exposure parameters) to the encrypted pixel payload through GCM's Associated Authenticated Data field, preventing context-swapping attacks; CTR + HMAC has no native AAD construct, and replicating this guarantee with a custom HMAC scope would require an ad-hoc, non-standardized construction.”
The paper relies on this without support.
CTR + HMAC has no native AAD construct.
“Second, native AAD binding: GpuFitsCrypt cryptographically binds the FITS metadata (size, geometry, exposure parameters) to the encrypted pixel payload through GCM's Associated Authenticated Data field, preventing context-swapping attacks; CTR + HMAC has no native AAD construct, and replicating this guarantee with a custom HMAC scope would require an ad-hoc, non-standardized construction.”
The paper relies on this without support.
Storage technology is a primary factor in I/O throughput bottlenecks during FITS processing.
“Storage technology is a primary factor in I / O throughput bottlenecks during FITS processing.”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Claimed H100-RTX 3060 kernel-time parity conflicts with Table 3
MajorThe text describes parity between these GPUs for the 600 MB workload, but Table 3 reports 0.026 s for the H100 and 0.236 s for the RTX 3060-a roughly ninefold difference in the kernel-only metric being discussed.
“Acounter-intuitive observation from Table 3 is the parity between high-end datacenter GPUs (H100) and consumer hardware (RTX 3060) for the 600 MB dataset.”
“| GeForce RTX 3060 | 128 | 4 | 0.236 ± 0.001 |”
Extra-Large file is said to exceed RTX 3060 VRAM
MinorThe stated Extra-Large dataset is approximately 3.4 GB, while the hardware table lists 12 GB of VRAM for the RTX 3060. The dataset size therefore does not itself exceed that GPU's listed memory capacity, although additional buffers could affect whether the benchmark runs.
“Note: Consumer-grade and edge devices with limited VRAM (RTX 3060, RTX 3050 Ti, and Jetson Orin) are excluded from the Extra-Large benchmark as the dataset size exceeds their available memory capacity.”
“| CCU | Desktop | GeForce RTX 3060 (12GB) | Intel Core i5-10400F | 62 GB | NVMe SSD |”
Cites
The FAIR Guiding Principles for scientific data management and stewardship
Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., Appleton, G., Axton, M., Baak, A., Blomberg, N., Boiten, J.-W., da Silva Santos, L. B., Bourne, P. E., et al. (2016) · Scientific Data · doi:10.1038/sdata.2016.18 · Cited by 18,423 · cited 2× · see citing works
“• Alignment with Open Science Goals: While the community strongly moves towards Open Science and FAIR data principles (Wilkinson et al., 2016; Chen et al., 2022), facilitated by public archives (Green et al., 2024), these principles primarily apply once data enters the public domain.”
LSST: From Science Drivers to Reference Design and Anticipated Data Products
Ivezić, Ž., Kahn, S. M., Tyson, J. A., Abel, B., Acosta, E., Allsman, R., Alonso, D., AlSayyad, Y., Anderson, S. F., Andrew, J., et al. (2019) · The Astrophysical Journal · doi:10.3847/1538-4357/ab042c · Cited by 4,357 · see citing works
The Zwicky Transient Facility: System Overview, Performance, and First Results
Bellm, E. C., Kulkarni, S. R., Graham, M. J., Dekany, R., Smith, R. M., Riddle, R., Masci, F. J., Hélou, G., Prince, T. A., Adams, S. M., et al. (2018) · Publications of the Astronomical Society of the Pacific · doi:10.1088/1538-3873/aaecbe · Cited by 2,077 · see citing works
Euclid Definition Study Report
Laureijs, R., Amiaux, J., Arduini, S., Auguères, J., Brinchmann, J., Cole, R., Cropper, M., Dabin, C., Duvet, L., Ealet, A., et al. (2011) · CERN Document Server (European Organization for Nuclear Research) · doi:10.48550/arxiv.1110.3193 · Cited by 1,767 · see citing works
Detection of B -Mode Polarization at Degree Angular Scales by BICEP2
Ade, P. A. R., Aikin, R. W., Barkats, D., Benton, S. J., Bischoff, C. A., Bock, J. J., Brevik, J. A., Buder, I., Bullock, E., Dowell, C. D., et al. (2014) · Physical Review Letters · doi:10.1103/physrevlett.112.241101 · Cited by 1,572 · see citing works
The Pan-STARRS1 SurveysPreprint
Chambers, K. C., Magnier, E. A., Metcalfe, N., Flewelling, H., Huber, M. E., Waters, C. Z., Denneau, L., Draper, P. W., Farrow, D. J., Finkbeiner, D. P., et al. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1612.05560 · Cited by 1,320 · see citing works
Joint Analysis of BICEP2/ <i>Keck Array</i> and <i>Planck</i> Data
Ade, P. A. R., Aghanim, N., Ahmed, Z., Aikin, R. W., Alexander, K. D., Arnaud, M., Aumont, J., Baccigalupi, C., Banday, A. J., Barkats, D., et al. (2015) · Physical Review Letters · doi:10.1103/physrevlett.114.101301 · Cited by 1,013 · see citing works
Biclique Cryptanalysis of the Full AES
Bogdanov, A., Khovratovich, D., & Rechberger, C. (2011) · Lecture Notes in Computer Science · doi:10.1007/978-3-642-25385-0_19 · Cited by 546 · see citing works
Recommendation for block cipher modes of operation :
Dworkin, M. J. (2007) · doi:10.6028/nist.sp.800-38d · Cited by 463 · see citing works
Representations of celestial coordinates in FITS
Calabretta, M. R., & Greisen, E. W. (2002) · Astronomy & Astrophysics · doi:10.1051/0004-6361:20021327 · Cited by 431 · see citing works
CUDA Compatible GPU as an Efficient Hardware Accelerator for AES Cryptography
Manavski, S. A. (2007) · 2007 IEEE International Conference on Signal Processing and Communications · doi:10.1109/icspc.2007.4728256 · Cited by 307 · see citing works
A collisional family of icy objects in the Kuiper belt
Brown, M. E., Barkume, K. M., Ragozzine, D., & Schaller, E. L. (2007) · Nature · doi:10.1038/nature05619 · Cited by 234 · see citing works
Representations of world coordinates in FITS
Greisen, E. W., & Calabretta, M. R. (2002) · Astronomy & Astrophysics · doi:10.1051/0004-6361:20021326 · Cited by 206 · see citing works
Discovery of a Planetary-sized Object in the Scattered Kuiper Belt
Brown, M. E., Trujillo, C. A., & Rabinowitz, D. L. (2005) · The Astrophysical Journal · doi:10.1086/499336 · Cited by 179 · see citing works
AES Galois Counter Mode (GCM) Cipher Suites for TLS
Salowey, J., Choudhury, A., & McGrew, D. A. (2008) · doi:10.17487/rfc5288 · Cited by 108 · see citing works
ZTF Early Observations of Type Ia Supernovae. I. Properties of the 2018 Sample
Yao, Y., Miller, A. A., Kulkarni, S. R., Bulla, M., Masci, F. J., Goldstein, D. A., Goobar, A., Nugent, P., Dugas, A., Blagorodnova, N., et al. (2019) · The Astrophysical Journal · doi:10.3847/1538-4357/ab4cf5 · Cited by 103 · see citing works
Lossless Astronomical Image Compression and the Effects of Noise
Pence, W. D., Seaman, R. L., & White, R. L. (2009) · Publications of the Astronomical Society of the Pacific · doi:10.1086/599023 · Cited by 50 · see citing works
How will astronomy archives survive the data tsunami?
Berriman, G. B., & Groom, S. L. (2011) · Communications of the ACM · doi:10.1145/2043174.2043190 · Cited by 50 · see citing works
Optimization of advanced encryption standard on graphics processing units
Tezcan, C. (2021) · IEEE Access · doi:10.1109/access.2021.3077551 · Cited by 45 · see citing works
An, S., & Seo, S. C. (2020) · Applied Sciences · doi:10.3390/app10113711 · Cited by 26 · see citing works
Best practices for data publication in the astronomical literature
Chen, T. X., Schmitz, M., Mazzarella, J. M., Wu, X., van Eyken, J. C., Accomazzi, A., Akeson, R. L., Allen, M., Beaton, R., Berriman, G. B., et al. (2022) · The Astrophysical Journal Supplement Series · doi:10.3847/1538-4365/ac6268 · Cited by 20 · see citing works
<title>The FITS Tape Formats: Flexible Image Transport Systems</title>
Greisen, E. W., Wells, D. C., & Harten, R. H. (1980) · SPIE Proceedings · doi:10.1117/12.959819 · Cited by 15 · see citing works
Parallel implementation of GCM on GPUs
Lee, J., Kim, D., & Seo, S. C. (2025) · ICT Express · doi:10.1016/j.icte.2025.01.006 · Cited by 7 · see citing works
Speed record of AES-CTR and AES-ECB bit-sliced implementation on GPUs
Lee, W.-K., Seo, S. C., Seo, H., Kim, D. C., & Hwang, S. O. (2024) · IEEE Embedded Systems Letters · doi:10.1109/les.2024.3409725 · Cited by 4 · see citing works
Malal, A., & Tezcan, C. (2026) · Journal of Cryptographic Engineering · doi:10.1007/s13389-025-00388-2 · Cited by 3 · see citing works
Postigo, A. D. U., Fatkhullin, T. A., Jóhannesson, G., et al. (2007) · Max Planck Digital Library · doi:10.1051/0004-6361
Сучасний підхід до проблеми виникнення університету як соціокультурного феномену
Aslan (2018) · Electronic Volyn National University Institutional Repository · doi:10.5281/zenodo
Flagged first, then by citations
Research works printed but not linked: 2
- Ponz, 1994, The FITS image extension, Astron. Astrophys. Suppl. Ser., 105, 53
- Wells, 1981, FITS - a flexible image transport system, Astron. Astrophys. Suppl. Ser., 44, 363
Cited by
Sorted by citations
Hardware Implementation of AES-256 Encryption on FPGA for Real-Time Video Stream Security
Akilandeswari, A. (2026) · International Conference on Trends in Electronics and Informatics (ICOEI) · doi:10.1109/icoei68323.2026.11541675
- Authors
- Lemes-Perera, S., R. Alarcon, M., Caballero-Gil, P., & Serra-Ricart, M.
- Affiliations
- Universidad de La Laguna (ES)
- Corresponding author
- Lemes-Perera, Samuel
- Venue
- Astronomy and Computing
- Publisher
- Elsevier BV
- Published
- 2026-10
- Volume
- 57
- Pages
- 101153
- ISSN
- 2213-1337 (ISSN-L)
- OpenAlex
- W7164763788
- Semantic Scholar
- 530b95375c532370073829978c24c026e8322335
- Cited by
- 1
- References
- 20
- Citations by year
- 2026: 1
- FWCI
- 3.46
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- Astronomy and Computing (publishedVersion, cc-by)
- SSRN Electronic Journal (acceptedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Funders
- España Ministerio de Ciencia e Innovación
- CajaCanarias Foundation
- Universidad de La Laguna
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Cryptographic Implementations and Security
- Physical Sciences › Computer Science › Artificial Intelligence › Cryptography and Data Security
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Keywords
- AES-GCM, authenticated encryption, GPU acceleration, cryptographic acceleration, tree reduction, access control, data confidentiality, data integrity, FAIR principles
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-30.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-30.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
