CoChat

A Study on Broker-Assisted Blockchain Trust Chains for Provenance and Integrity Verification of Generative Media Using Watermarking, Semantic Fingerprinting, and C2PA

Open access
Yang, Chaelin, Kim, Minchul · 2026 · Applied Sciences

The DOI matches its Crossref record; 23 of the 30 references it cites were found in a registry or an index, and 2 are not checked yet; 7 of the 17 sources its claims cite were checked for support.

Abstract

This research proposes a trust chain architecture that treats authenticity verification as an evidence registration and validation workflow rather than a single-signal decision.

Verified 2026-11-19 · 2 references not checked yet

Overview

CoChat Summary

TL;DR

The paper proposes a broker-assisted trust chain architecture that seals media with robust watermarking, embedding-based semantic fingerprints, and C2PA-compatible provenance manifests, storing sealed media off-chain via IPFS and anchoring compact evidence on an EVM ledger. On a COCO 2017 validation subset (N = 200), retrieval achieved Recall@1 = 0.9988 and Recall@5/10 = 1.000 under JPEG compression, resizing, and center cropping, while a representative registration transaction consumed 25,380 gas.

Key findings

  • Retrieval was near-perfect on the COCO 2017 validation subset (N = 200): overall Recall@1 = 0.9988, Recall@5 = 1.0000, Recall@10 = 1.0000 under JPEG compression (Q = 50 and Q = 80), 0.5× bicubic resizing, and 10% center cropping. Per-transformation Recall@1 was 100.0% for JPEG Q = 50, 100.0% for JPEG Q = 80, 99.5% for resize, and 100.0% for center crop.
  • On-chain anchoring overhead was low in the EVM proof-of-concept: a representative registration transaction consumed 25,380 gas for fixed-size compact evidence fields (CID hash, semantic fingerprint evidence hash, provenance manifest hash, timestamp; broker address as transaction sender). Annualized fee estimates ranged from 0.001506 ETH / $0.64 in 2020 to 0.000097 ETH / $0.31 in 2025, and 0.00000173 ETH / $0.0035 in March 2026.
  • The architecture separates off-chain storage (IPFS content identifier), off-chain FAISS similarity search, and on-chain compact evidence; verification uses similarity search to retrieve candidates, then validates ledger inclusion and broker signatures, with optional watermark detection and sealed-media retrieval, issuing signed verdicts such as Verified, Derived, Unverified, and Suspected Tampered.
  • The threat model treats watermarking as a complementary signal rather than a standalone guarantee; provenance metadata can be stripped, and ledger anchoring plus multi-field consistency checks are used to mitigate tampering, watermark degradation/removal, watermark copy/transfer, adversarial embedding manipulation, and semantic hash collision attempts.

Methods

  • Design: a broker-assisted trust chain workflow with registration (watermark embedding, semantic fingerprint extraction, C2PA-compatible manifest generation, IPFS storage, ledger commitment) and verification (candidate retrieval, ledger record validation, signature checks, optional watermark detection/sealed-media retrieval, signed report issuance).
  • Data and retrieval evaluation: COCO 2017 validation subset, N = 200 randomly sampled images with a fixed random seed; transformed queries used JPEG compression at Q = 50 and Q = 80, resizing to 0.5× with bicubic interpolation, and center cropping with 10% area reduction. Semantic fingerprinting used a pretrained CLIP image encoder (ViT-B/32) to extract L2-normalized embeddings; FAISS was used for nearest-neighbor retrieval; performance was reported as Recall@k.
  • Ledger evaluation: an EVM-based proof-of-concept using a local Anvil testbed and a minimal EvidenceRegistry smart contract; each registration anchored fixed-size evidence fields including a CID hash (bytes32), semantic fingerprint evidence hash (bytes32), provenance manifest hash (bytes32), and timestamp (uint64), with per-registration overhead measured as gasUsed.

Limitations

  • The paper states that the retrieval evaluation is conducted on a moderate-scale subset and under moderate transformations; stronger distortions such as rotation, heavy noise, adversarial perturbations, and more severe cropping may reduce similarity and should be evaluated further.
  • The paper states that broker compromise or collusion is not fully prevented by design; the architecture aims to make broker-issued claims auditable through signatures and immutable record inclusion. It also notes that similarity-index configurations introduce accuracy-latency tradeoffs that should be characterized for large-scale deployments.
  • Evident from the text: the evaluation does not benchmark specific watermarking algorithms or watermark robustness, and it measures retrieval performance plus per-transaction gas rather than full end-to-end overhead across sealing, storage, retrieval, verification, and IPFS workflows. The retrieval experiment uses COCO 2017 images, not generative media specifically, although the paper targets generative media provenance.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/app16073391

Cites

23 of 30 references verified0 citation errors · 5 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Overall architecture of a broker-assisted trust chain for sealing and verifying generative media.
Figure 1 Overall architecture of a broker-assisted trust chain for sealing and verifying generative media.CC-BY · From the PDF
Figure 2 Internal architecture of the trust chain broker service.
Figure 2 Internal architecture of the trust chain broker service.CC-BY · From the PDF
Figure 3 Registration sequence of the broker service (content sealing).
Figure 3 Registration sequence of the broker service (content sealing).CC-BY · From the PDF
Image 4
CC-BY · From the PDF
Figure 5 The fi gure shows the commands and receipt output used for the measurement reported in Table 3.
Figure 5 The fi gure shows the commands and receipt output used for the measurement reported in Table 3.CC-BY · From the PDF

Verification

Claims the paper makes

5 Valid · 1 Not valid · 5 Not checked · 58 Own results and interpretations

22 more claims

Create a free account to see all 69 claims and to verify the 3 references still open.

See all claims — free account
  • Probably not a substantial claim. 27
  • It may state two separate claims as one. 27

Create a free account to see these claims.

Citations

Cites

Metadata

Authors
Yang, C., & Kim, M.
Affiliations
Mokpo National University (KR)
Corresponding author
Kim, Minchul
Venue
Applied Sciences
Publisher
MDPI AG
Published
2026-03-31
Volume
16
Issue
7
Pages
3391
ISSN
2076-3417 (ISSN-L)
OpenAlex
W7147460381
Cited by
0
References
28
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Applied Sciences (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Ministry of Education (2026-RISE-14-001)
Topics
  • Physical Sciences › Computer Science › Information Systems › Blockchain Technology Applications and Security
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
Keywords
blockchain anchoring, C2PA, provenance metadata, content-addressed storage, FAISS similarity search, candidate retrieval

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (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

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.