Exploring Bitslicing Architectures for Enabling FHE-Assisted Machine Learning
The DOI matches its Crossref record; 29 of the 29 references it cites were found in a registry or an index.
Abstract
Homomorphic encryption (HE) is the ultimate tool for performing secure computations even in untrusted environments. Application of HE for deep learning (DL) inference is an active area of research, given the fact that DL models are often deployed in untrusted environments (e.g., third-party servers) yet inferring on private data. However, existing HE libraries [somewhat (SWHE), leveled (LHE) or fully homomorphic (FHE)] suffer from extensive computational and memory overhead. Few performance optimized high-speed homomorphic libraries are either suffering from certain approximation issues leading to decryption errors or proven to be insecure according to recent published attacks. In this article, we propose architectural tricks to achieve performance speedup for encrypted DL inference developed with exact HE schemes without any approximation or decryption error in homomorphic computations. The main idea is to apply quantization and suitable data packing in the form of bitslicing to reduce the costly noise handling operation, Bootstrapping while achieving a functionally correct and highly parallel DL pipeline with a moderate memory footprint. Experimental evaluation on the MNIST dataset shows a significant ( $37\times$ ) speedup over the nonbitsliced versions of the same architecture. Low memory bandwidths (700 MB) of our design pipelines further highlight their promise toward scaling over larger gamut of Edge-AI analytics use cases.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“However, existing HE implementations suffer from the burden of extensive computational and memory overhead.”
From Confidential Computing Across Edge‐To‐Cloud for Machine Learning: A Survey Study (2025) · cites this paper as [144]
Citations
Cites
Advances in neural information processing systems 7
(1996) · Computers & Mathematics with Applications · doi:10.1016/s0898-1221(96)90463-0 · Cited by 14,281
Homomorphic Encryption for Arithmetic of Approximate Numbers
Cheon, J. H., Kim, A., Kim, M., & Song, Y. (2017) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-70694-8_15 · Cited by 2,522 · cited 2×
“It has been found that the secret key of certain LHE schemes (such as CKKS [11]) can be extracted with the knowledge of a few decrypted ciphertexts [12], [13].”
BackgroundA fully homomorphic encryption scheme
Boneh, D., & Gentry, C. (2009) · Cited by 2,341
On lattices, learning with errors, random linear codes, and cryptography
Regev, O. (2009) · Journal of the ACM · doi:10.1145/1568318.1568324 · Cited by 2,315
(1978) · Cited by 1,875
(Leveled) fully homomorphic encryption without bootstrapping
Brakerski, Z., Gentry, C., & Vaikuntanathan, V. (2012) · Proceedings of the 3rd Innovations in Theoretical Computer Science Conference · doi:10.1145/2090236.2090262 · Cited by 1,782
A Public Key Cryptosystem and a Signature Scheme Based on Discrete Logarithms
ElGamal, T. (2007) · Lecture Notes in Computer Science · doi:10.1007/3-540-39568-7_2 · Cited by 1,604
Gentry, C., Sahai, A., & Waters, B. (2013) · Lecture Notes in Computer Science · doi:10.1007/978-3-642-40041-4_5 · Cited by 1,492
(2016) · International Conference on Machine Learning · Cited by 1,270
TFHE: Fast Fully Homomorphic Encryption Over the Torus
Chillotti, I., Gama, N., Georgieva, M., & Izabachène, M. (2020) · Journal of Cryptology · doi:10.1007/s00145-019-09319-x · Cited by 990 · cited 3×
“1 is implemented as based on python library NuFHE [16], which is an extension of that TFHE scheme [10].”
BackgroundMethodFHEW: Bootstrapping Homomorphic Encryption in Less Than a Second
Ducas, L., & Micciancio, D. (2015) · Lecture Notes in Computer Science · doi:10.1007/978-3-662-46800-5_24 · Cited by 745 · cited 1×
“Ducas and Micciancio [9] proposed an FHE bootstrapping mechanism which takes less than 1 s.”
BackgroundPrivacy-Preserving Machine Learning With Fully Homomorphic Encryption for Deep Neural Network
Lee, J.-W., Kang, H., Lee, Y., Choi, W., Eom, J., Deryabin, M., Lee, E., Lee, J., Yoo, D., Kim, Y.-S., et al. (2022) · IEEE Access · doi:10.1109/access.2022.3159694 · Cited by 414 · cited 2×
“4) We perform an in-depth memory profiling and try to benchmark our inference accelerator with other state-of-the-art approaches reported in [4], [5], [7], and [19].”
BackgroundA fast new DES implementation in software
Biham, E. (1997) · Lecture Notes in Computer Science · doi:10.1007/bfb0052352 · Cited by 328 · cited 2×
“Originally proposed by Biham [18], for accelerating the computation of DES [32] in software, bitslicing is a parallelization technique widely utilized for cryptographic implementations.”
BackgroundMethodCryptoDL: Deep Neural Networks over Encrypted Data
Hesamifard, E., Takabi, H., & Ghasemi, M. (2017) · arXiv · doi:10.48550/arxiv.1711.05189 · Cited by 306 · cited 3×
“Inference frameworks proposed in [4] and [24] build up on same ideas and also produce good accuracy.”
BackgroundFast Homomorphic Evaluation of Deep Discretized Neural Networks
Bourse, F., Minelli, M., Minihold, M., & Paillier, P. (2018) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-96878-0_17 · Cited by 300
Low Latency Privacy Preserving InferencePreprint
Brutzkus, A., Elisha, O., & Gilad-Bachrach, R. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1812.10659 · Cited by 297 · cited 5×
“4) We perform an in-depth memory profiling and try to benchmark our inference accelerator with other state-of-the-art approaches reported in [4], [5], [7], and [19].”
BackgroundMethodInfluentialAl Badawi, A., Jin, C., Lin, J., Mun, C. F., Jie, S. J., Tan, B. H. M., Nan, X., Aung, K. M. M., & Chandrasekhar, V. R. (2021) · IEEE Transactions on Emerging Topics in Computing · doi:10.1109/tetc.2020.3014636 · Cited by 167 · cited 1×
“and [23] suffer from liberty of selecting the activation functions.”
BackgroundProgrammable Bootstrapping Enables Efficient Homomorphic Inference of Deep Neural Networks
Chillotti, I., Jóye, M., & Paillier, P. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-78086-9_1 · Cited by 164
SHE: A Fast and Accurate Deep Neural Network for Encrypted DataPreprint
Lou, Q., & Jiang, L. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1906.00148 · Cited by 160
On the Security of Homomorphic Encryption on Approximate Numbers
Li, B., & Micciancio, D. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-77870-5_23 · Cited by 159 · cited 2×
“It has been found that the secret key of certain LHE schemes (such as CKKS [11]) can be extracted with the knowledge of a few decrypted ciphertexts [12], [13].”
BackgroundConcrete Crack Images for Classification
Özgenel, Ç. F. (2019) · Data Archiving and Networked Services (DANS) · doi:10.17632/5y9wdsg2zt.2 · Cited by 134 · cited 11×
“the Mendeley Bridge dataset [36] as shown in Table IX.”
MethodInfluentialMoons, B., Bankman, D. J., & Verhelst, M. (2018) · Springer eBooks · doi:10.1007/978-3-319-99223-5 · Cited by 71
Translating Algorithms to Handle Fully Homomorphic Encrypted Data on the Cloud
Chatterjee, A., & Sengupta, I. (2015) · IEEE Transactions on Cloud Computing · doi:10.1109/tcc.2015.2481416 · Cited by 55 · cited 1×
“Realization of maximum function in homomorphic domain requires encrypted decision making, that is not feasible with LHE computations [39], [40].”
BackgroundSelf-citationSEAL-Embedded: A Homomorphic Encryption Library for the Internet of Things
Natarajan, D., & Dai, W. (2021) · IACR Transactions on Cryptographic Hardware and Embedded Systems · doi:10.46586/tches.v2021.i3.756-779 · Cited by 44
Accelerating Sorting of Fully Homomorphic Encrypted Data
Chatterjee, A., Kaushal, M., & Sengupta, I. (2013) · Lecture notes in computer science · doi:10.1007/978-3-319-03515-4_17 · Cited by 41 · cited 1×
“Realization of maximum function in homomorphic domain requires encrypted decision making, that is not feasible with LHE computations [39], [40].”
BackgroundSelf-citationGentry, C. (2009) · Lecture notes in computer science · doi:10.1007/978-3-642-10433-6_32 · Cited by 15
Efficient CNN Building Blocks for Encrypted DataPreprint
Jain, N., Nandakumar, K., Ratha, N., Pankanti, S., & Kumar, U. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2102.00319 · Cited by 15 · cited 3×
“For example, [6] is such an approach which approximates nonlinear activations with arithmetic polynomials.”
BackgroundMethodAppendix 1: Data Encryption Standard (DES)
Boura, C. (2023) · Wiley eBooks · doi:10.1002/9781394256358.app1 · Cited by 6
Encrypted Operator Design with Domain Aware Practical Performance Improvement
Ghosh, A., Raj, A., & Chatterjee, A. (2022) · Advances in intelligent systems and computing · doi:10.1007/978-981-16-6890-6_8 · Cited by 1 · cited 2×
“However, this limitation can be outdone in ML processing with prior domain knowledge, where the maximum value of such input vectors is known [41].”
BackgroundSelf-citation
Research works printed but not linked: 14
- Root, 2022, Cryptominers Threaten GCP Virtual Servers
- Crypto, 2021, CKKS key recovery
- 2021, ZAMA.AI/Concrete
- NuFHE, A GPU-Powered Torus FHE Implementation
- GAMA, TFHE
- 2022, ZAMA.Ai/concrete
- Goodfellow, 2016, Deep Learning
- 2022, Tensorflow model optimization
- 2021, NVIDIA/FasterTransformer
- LeCun, 2010, MNIST Handwritten Digit Database
- 2020, Quantized TFLite model gives better accuracy then TF model
- Praveen, 2020, Improving INT8 accuracy using quantization aware training and the NVIDIA TAO toolkit
- 2020, pyFHE Github Repository
- Krizhevsky, 2009, The Cifar10 Dataset
Cited by
Optimizing Edge AI: A Comprehensive Survey on Data, Model, and System StrategiesPreprint
Wang, X., & Jia, W. (2025) · Qeios · doi:10.32388/izohch · Cited by 56
Confidential Computing Across Edge‐To‐Cloud for Machine Learning: A Survey Study
Zobaed, S., & Amini Salehi, M. (2025) · Software Practice and Experience · doi:10.1002/spe.3398 · Cited by 21
“Recent research studies [144] have been undertaken to propose architectural optimizations and algorithmic enhancements that address these challenges.”
FHEDGE: Encrypted Inference on Lightweight Edge Devices
Sinha, S., Saha, S., Chatterjee, A., & Mukhopadhyay, D. (2024) · IEEE Asia Pacific Conference on Circuits and Systems (APCCAS) · doi:10.1109/apccas62602.2024.10808548 · Cited by 1
Self-citationVinko, D., Köhler, M., Miličević, K., & Lukić, I. (2026) · Telecom · doi:10.3390/telecom7010014 · Cited by 1
Metadata
- Authors
- Sinha, S., Saha, S., Alam, M., Agarwal, V., Chatterjee, A., Mishra, A., Khazanchi, D., & Mukhopadhyay, D.
- Affiliations
- Indian Institute of Technology Kharagpur (IN)
- Nanyang Technological University (SG)
- New York University Abu Dhabi (AE)
- Delhi Technological University (IN)
- University of Nebraska at Omaha (US)
- ORCID
- Venue
- IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Published
- 2022-11
- Volume
- 41
- Issue
- 11
- Pages
- 4004-4015
- ISSN
- 0278-0070 (ISSN-L)
- 1937-4151
- OpenAlex
- W4308479834
- Cited by
- 4
- References
- 29
- Citations by year
- 2026: 1 · 2025: 2 · 2024: 1
- FWCI
- 0.25
- Type
- journal article
- Language
- en
- Open access
- closed
- Licence
- Locations
- IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (publishedVersion)
- DR-NTU (Nanyang Technological University) (publishedVersion)
- Funders
- SERB SUPRA DST Government of India
- U.S.–India Cooperation in Higher Education With University of Nebraska at Omaha
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › Cryptography and Data Security
- Physical Sciences › Computer Science › Artificial Intelligence › Cryptographic Implementations and Security
- Physical Sciences › Computer Science › Hardware and Architecture › Security and Verification in Computing
- Keywords
- fully homomorphic encryption, bootstrapping, noise handling, quantization, data packing, hardware architecture optimization, memory bandwidth, edge AI, privacy-preserving computation
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.
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
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.