Research map / Image generation / Autoregressive and latent generation
Beyond Bit Matching: Orthogonal Watermarks for Collusion-Resistant Image Fingerprinting
Zhakshylyk Nurlanov, Tobias Weißberg, Florian Bernard
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OrthoMark, an image fingerprinting method that maps user keys to near-orthogonal watermark patterns to resist averaging collusion attacks.
Problem
Image fingerprinting for source tracing is vulnerable to averaging collusion, where several recipients average differently marked copies to suppress individual watermarks; discrete bit-string watermarks are especially fragile because averaging pushes bit evidence toward ambiguous decisions.
Approach
Maps each user key to a near-orthogonal watermark pattern instead of a discrete bit string, so that averaged copies retain identifiable evidence for colluders rather than becoming ambiguous.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use continuous near-orthogonal vector keys with cosine retrieval instead of binary payload bit matching | VectorMark vs TrustMark · MS-COCO, K=5 averaging | strict all-colluder TPR: 0.000 -> 1.00 | Improved |
| train a neural encoder/extractor instead of spread-spectrum embedding | VectorMark vs SS-DFT · MS-COCO, K=10 averaging | strict all-colluder TPR: .171 -> .996 | Improved |
| train with collusion objective (K_train=20) (ablation) | VectorMark · short runs, K=20 averaging | strict all-colluder TPR: .234 -> .650 | Improved |
| add collusion objective at K=2..10 (ablation) | VectorMark · short runs, K<=10 | strict all-colluder TPR: 1.00 -> 1.00 | No clear effect |
| add collusion training (ablation) | VectorMark · short runs, geometric single-key TPR | TPR: .926 -> .901 | Hurt |
| increase key dimension from 100 to 1024 (ablation) | VectorMark · short runs, K=10 averaging | strict all-colluder TPR: .000 -> 1.00 | Improved |
| increase key dimension from 100 to 576 (ablation) | VectorMark · short runs, single-key geometric | TPR: .969 -> .995 | Improved |
| use soft cosine scores instead of hard codeword decisions (ablation) | VectorMark · MS-COCO registry N=1e6, K=10 | exact-set recovery: .624 -> .974 | Improved |
| retain logits for cosine matching instead of hard agreement (ablation) | ChunkySeal · registry N=1e6, K=10 | exact-set recovery: .002 -> .441 | Improved |
| two-pass decoding with horizontally flipped image (ablation) | VectorMark payload decoding · MS-COCO geometric payload | bit accuracy: .827 -> .975 | Improved |
| scale registry from 1e3 to 1e6 keys | VectorMark · K=10 exact-set recovery | exact-set recovery: .989 -> .980 | No clear effect |
| increase coalition size from 10 to 20 at N=1e6 | VectorMark · exact-set recovery | exact-set recovery: .980 -> .007 | Hurt |
Closest papers by meaning
- Revisiting Coding-Based Approaches to Overcome the Curse of Dimensionality in Learning-Based Watermarking (ICML 2026 · similarity 0.86)
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Can Simple Averaging Defeat Modern Watermarks? (NeurIPS 2024 · similarity 0.84)
- Mask Image Watermarking (NeurIPS 2025 · similarity 0.84)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.83)
- On the Coexistence and Ensembling of Watermarks (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
