Beyond Bit Matching: Orthogonal Watermarks for Collusion-Resistant Image Fingerprinting
Organizations: University of Bonn Lamarr Institute
Abstract
Image fingerprinting assigns each distributed copy a user-specific watermark for source tracing. Averaging collusion is a central threat: several recipients can average their differently marked copies to suppress each individual mark. Discrete bit-string watermarks are especially vulnerable because averaging drives bit evidence toward ambiguous decisions. We introduce VectorMark, which maps user keys to near-orthogonal high-dimensional unit vectors and identifies colluders by cosine similarity to the extracted watermark direction. Under ideal averaging of vector keys, the normalized average preserves a cosine signal of order for each colluder, while unrelated keys remain concentrated near zero. The spherical geometry of random directions gives a shifted-Beta null distribution for cosine similarities, enabling analytic per-key false-positive-rate control after validation on unwatermarked images. VectorMark uses a neural watermark encoder and centered extractor trained with JND masking and a progressive curriculum for photometric, geometric, and collusion robustness. Experiments on MS-COCO, OpenImages, and AI-generated images show that VectorMark maintains robust watermark attribution under photometric and geometric distortions, and achieves the best strict all-colluder detection among evaluated methods, while preserving high visual quality. Our code is publicly available: https://github.com/nurlanov-zh/VectorMark
Figures & tables
| Quality | Robustness | All-colluder detect. | ||||||||
| Method | Dim | PSNR | LPIPS | Clean | Photo. | Geo. | ||||
| Cont. keys | VectorMark | 1024 | 40.29 | .041 | 1.00 | 1.00 | .985 | 1.00 | 1.00 | .996 |
| SS-DCT Cox et al. (1997) | 1024 | 34.25 | .198 | 1.00 | 1.00 | .000 | 1.00 | .829 | .092 | |
| SS-DFT Cox et al. (1997) | 1024 | 33.85 | .207 | 1.00 | .997 | .000 | .999 | .945 | .171 | |
| SS-FM Lin et al. (2001) | 74 | 32.84 | .103 | .813 | .680 | .041 | .060 | .000 | .000 | |
| Binary keys | VINE Lu et al. (2025) | 100 | 36.98 | .017 | 1.00 | 1.00 | .000 | .307 | .000 | .000 |
| Dataset | Method | PSNR | Clean | Photo. | Geo. | |||
| OpenImages | VectorMark | 40.36 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| OpenImages | SS-DFT | 35.22 | .998 | .997 | .249 | .994 | .934 | .184 |
| OpenImages | TrustMark | 40.81 | 1.00 | 1.00 | .750 | .369 | .000 | .000 |
| AI-generated | VectorMark | 41.53 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| AI-generated | SS-DFT | 36.68 | 1.00 | 1.00 | .250 | 1.00 | .938 | .194 |
| AI-generated | TrustMark | 42.21 | 1.00 | 1.00 | .750 | .388 | .000 | .000 |
| VectorMark | Baselines, sorted by all-transform mean | |||||||||||
| Setting | Ours raw | Ours ECC | Ours 2-pass ECC | Riva GAN | MaskWM | Trust Mark | HiDDeN | Stega Stamp | VINE | MBRS | DWT- DCT SVD | DWT-DCT |
| bits predicted | ||||||||||||
| Clean | 1.00 | 1.00 | 1.00 | .997 | 1.00 | 1.00 | .884 | .999 | 1.00 | 1.00 | .998 | .855 |
| Photometric | .945 | .956 | .957 | .987 | 1.00 | .981 | .866 | .991 | .994 | .983 | .770 | .730 |
| Geometric | .838 | .827 | .975 | .787 | .725 | .668 | .686 | .588 | .500 | .499 | .535 | .519 |
| All trans. avg. | .874 | .870 | .969 | .854 | .816 | .772 | .746 | .722 | .665 | .660 | .613 | .589 |
| Recall | Precision | All-colluder | No-FA | Exact set | FWER | |
| 1.000 | .999 | .999 | .990 | .989 | .010 | |
| 1.000 | .999 | .998 | .990 | .988 | .010 | |
| .999 | .999 | .996 | .988 | .984 | .012 | |
| .998 | .999 | .990 | .989 | .980 | .011 |
| Dim. | PSNR | Clean | Photo. | Geo. | ||||||
| 1024 | 20 | 32.97 | 1.00 | 1.00 | .901 | 1.00 | 1.00 | 1.00 | .977 | .650 |
| 1024 | – | 37.94 | 1.00 | 1.00 | .926 | 1.00 | 1.00 | 1.00 | .927 | .234 |
| 100 | 20 | 32.38 | 1.00 | .991 | .969 | 1.00 | .008 | .000 | .000 | .000 |
| 256 | 20 | 31.89 | 1.00 | 1.00 | .989 | 1.00 | .985 | .007 | .000 | .000 |
| 576 | 20 | 32.43 | 1.00 | 1.00 | .995 | 1.00 | 1.00 | .988 | .321 | .001 |
| 1024 | 20 | 32.97 | 1.00 | 1.00 | .901 | 1.00 | 1.00 | 1.00 | .977 | .650 |
| Method | Score | PSNR | |||
| VectorMark | 1024 | Soft cosine | .992 | .974 | 40.28 |
| VectorMark | 1024 | Hard agreement | .991 | .624 | 40.28 |
| ChunkySeal | 1024 | Logit cosine | .949 | .441 | 44.70 |
| ChunkySeal | 1024 | Hard agreement | .923 | .002 | 44.70 |
| SS-DFT | 1024 | Soft cosine | .862 | .008 | 33.85 |
| DWT-DCT | 1024 | Hard agreement | .490 | .036 | 36.79 |
| Attack | Exact set | Recall | Precision | PSNR | |
| Mean averaging | 10 | .979 | .998 | .999 | 41.03 |
| Subspace removal, | 10 | .704 | .945 | .999 | 37.73 |
| Subspace removal, | 10 | .083 | .448 | .878 | 34.03 |
| Subspace removal, | 10 | .000 | .012 | .058 | 28.99 |
| WmForger | 5 | .725 | .838 | .904 | 32.68 |
| Regen-VAE | 5 | .980 | .998 | .998 | 36.81 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Distortion | Ours | SS-FM ‡ | SS-DCT | SS-DFT | Mask- WM | Stega- Stamp | VINE | Trust- Mark | MBRS | Riva- GAN | DWT DCT SVD | DWT DCT | HiDDeN |
| Clean | 1.000 | 0.813 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.996 | 0.996 | 0.716 | 0.293 |
| Valuemetric (aggregated) | 1.000 | 0.680 | 1.000 | 0.997 | 1.000 | 1.000 | 1.000 | 0.978 | 0.967 | 0.971 | 0.561 | 0.467 | 0.425 |
| JPEG | 0.999 | 0.774 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.992 | 0.997 | 0.746 | 0.459 |
| JPEG | 0.994 | 0.758 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.992 | 0.998 | 0.736 | 0.415 |
| Gauss. Blur | 1.000 | 0.768 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.995 | 0.997 | 0.768 | 0.492 |
| Gauss. Blur | 1.000 | 0.724 | 1.000 | 0.999 | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.995 | 0.997 | 0.669 | 0.462 |
| Distortion | Ours raw | Ours key | Ours 2-pass raw | Ours 2-pass key | Mask- WM | Stega- Stamp | VINE | Trust- Mark | MBRS | Riva- GAN | DWT DCT SVD | DWT DCT | HiDDeN |
| score | acc | acc | acc | acc | acc | acc | acc | acc | acc | acc | acc | acc | acc |
| dim/bits | |||||||||||||
| Clean | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | 0.999 | 1.000 | 1.000 | 1.000 | 0.997 | 0.998 | 0.855 | 0.884 |
| Valuemetric (aggregated) | 0.945 | 0.956 | 0.944 | 0.957 | 1.000 | 0.991 | 0.994 | 0.981 | 0.983 | 0.987 | 0.770 | 0.730 | 0.866 |
| JPEG | 0.789 | 0.832 | 0.788 | 0.834 | 1.000 | 0.998 | 0.999 | 0.998 | 1.000 | 0.996 | 0.996 | 0.857 | 0.872 |
| JPEG | 0.720 | 0.655 | 0.716 | 0.658 | 1.000 | 0.998 | 0.999 | 0.998 | 1.000 | 0.996 | 0.995 | 0.854 | 0.868 |
| Field | Evidence |
| Training dataset | MS-COCO 2014 train, 82,783 images |
| Evaluation dataset | MS-COCO 2014 val subset, 1,000 images |
| Checkpoint training step | 280,000 |
| Encoder parameters | 2,980,006 |
| Extractor parameters | 3,000,049 |
| Stored key-generator tensors | 1,150,976 |
| Method | Native Res. | Encode (img/s) | Extract (img/s) |
| Traditional (non-learned) | |||
| DWT-DCT-SVD † | 54 | 80 | |
| DWT-DCT † | 138 | 222 | |
| SS-DCT | 11,937 | 54,578 | |
| SS-FM | 35,329 | 81,974 | |
| SS-DFT | 38,103 | 111,782 | |
| Noise Category | Deterministic Configurations |
| Valuemetric Operations | |
| JPEG Compression | Quality: , |
| Gaussian Blur | Kernel with |
| Gaussian Noise | (Fixed seed: 42) |
| Median Filter | Kernel size: |
| Salt & Pepper | Probability (Fixed seed: 42) |
| Setting | Trials | Successes | TPR | Wilson 95% CI |
| Clean | 1,000 | 1,000 | 1.0000 | [0.9962, 1.0000] |
| Photometric | 15,000 | 14,993 | 0.9995 | [0.9990, 0.9998] |
| Geometric | 30,000 | 29,554 | 0.9851 | [0.9837, 0.9864] |
| Averaging collusion, | 1,000 | 1,000 | 1.0000 | [0.9962, 1.0000] |
| Averaging collusion, | 1,000 | 1,000 | 1.0000 | [0.9962, 1.0000] |
| Averaging collusion, | 1,000 | 996 | 0.9960 | [0.9898, 0.9984] |
| Recall | Precision | All-colluder | No-FA | Exact set | FWER | ||
| 5 | 1.000 | .998 | 1.000 | .990 | .990 | .010 | |
| 10 | 1.000 | .999 | .999 | .990 | .989 | .010 | |
| 15 | .995 | .999 | .950 | .989 | .940 | .011 | |
| 20 | .951 | 1.000 | .538 | .991 | .533 | .009 | |
| 5 | 1.000 | .999 | 1.000 | .991 | .991 | .009 | |
| 10 | 1.000 | .999 | .998 | .990 | .988 | .010 |
| Method | Score | PSNR (dB) | |||||
| VectorMark | 1024 | Soft cosine | 99.2 100.0 / 99.9 | 97.4 99.8 / 99.9 | 91.4 98.9 / 99.9 | 52.8 93.5 / 99.9 | 40.28 |
| VectorMark | 1024 | Hard agreement | 99.1 100.0 / 99.8 | 62.4 94.4 / 99.9 | 13.4 82.3 / 99.6 | 0.0 56.5 / 99.0 | 40.28 |
| DWT-DCT | 256 | Hard agreement | 0.3 11.7 / 27.4 | 0.0 0.6 / 5.0 | 0.0 0.2 / 2.3 | 0.0 0.1 / 1.0 | 36.79 |
| DWT-DCT | 512 | Hard agreement | 9.4 18.8 / 32.3 | 0.0 3.5 / 12.6 | 0.0 1.8 / 10.1 | 0.0 0.6 / 7.2 | 36.79 |
| DWT-DCT | 1024 | Hard agreement | 49.0 71.8 / 88.9 | 3.6 22.3 / 58.4 | 1.0 14.5 / 48.4 | 0.0 7.2 / 31.7 | 36.79 |
| SS-DFT | 1024 | Cosine | 86.2 94.2 / 98.4 | 0.8 40.1 / 85.5 | 0.0 19.4 / 72.2 | 0.0 5.7 / 46.4 | 33.85 |
| Attack | VectorMark | SS-DFT | TrustMark | ChunkySeal | PixelSeal | VideoSeal | |
| Mean | 10 | 97.9 99.9 99.8 41.03 | 1.3 85.4 40.5 43.54 | 0.0 0.2 0.0 42.39 | 0.6 68.7 32.9 51.32 | 0.0 10.8 1.1 50.49 | 0.0 3.3 0.4 50.83 |
| Subsp. ( ) | 10 | 70.4 99.9 94.5 37.73 | 0.1 73.9 28.0 36.18 | 0.0 0.1 0.0 35.63 | 0.0 32.1 9.5 40.08 | 0.0 3.2 0.3 36.55 | 0.0 2.1 0.2 36.23 |
| Subsp. ( ) | 10 | 8.3 87.8 44.8 34.03 | 0.0 28.5 5.1 30.71 | 0.0 0.0 0.0 30.55 | 0.0 8.3 2.0 34.39 | 0.0 0.7 0.1 30.72 | 0.0 0.1 0.0 30.34 |
| Subsp. ( ) | 10 | 0.0 5.8 1.2 28.99 | 0.0 0.0 0.0 24.82 | 0.0 0.0 0.0 24.98 | 0.0 0.9 0.2 28.51 | 0.0 0.0 0.0 24.78 | 0.0 0.0 0.0 24.37 |
| WmForger | 5 | 72.5 90.4 83.8 32.68 | 61.0 89.5 74.3 32.98 | 0.0 1.0 0.2 32.67 | 10.5 25.0 18.7 33.26 | 0.0 7.5 1.9 33.24 | 0.0 4.0 0.9 33.29 |
| Regen-VAE | 5 | 98.0 99.8 99.8 36.81 | 80.5 97.0 91.2 36.30 | 0.0 3.0 0.6 36.73 | 3.5 10.5 6.8 37.98 | 0.0 75.5 22.0 37.93 | 0.0 48.8 12.0 37.94 |
| Recall | Precision | All-label | No-FA | Exact set | FWER | |
| 1 | 1.000 | .994 | 1.000 | .988 | .988 | .012 |
| 5 | 1.000 | .998 | 1.000 | .988 | .988 | .012 |
| 10 | 1.000 | .999 | 1.000 | .994 | .994 | .006 |
| 15 | .995 | .999 | .926 | .989 | .915 | .011 |
| 20 | .924 | .999 | .184 | .989 | .183 | .011 |
| 25 | .750 | 1.000 | .002 | .992 | .002 | .008 |
| WAM | MaskWM | VectorMark | |||
| Blind | GT mask | Blind | GT mask | Global | |
| 1 | 98.5 99.1 99.6 37.01 | 98.5 99.1 99.6 37.01 | 99.2 99.6 100.0 35.76 | 99.2 99.6 100.0 35.76 | 98.8 99.4 100.0 40.16 |
| 5 | 88.3 99.9 96.9 37.58 | 87.8 99.9 96.8 37.58 | 94.3 99.9 98.1 30.97 | 99.4 99.9 100.0 30.97 | 98.8 99.8 100.0 40.16 |
| 10 | 30.2 99.7 82.9 37.96 | 29.4 99.7 82.4 37.96 | 79.7 99.9 95.1 30.41 | 99.0 99.9 100.0 30.41 | 99.4 99.9 100.0 40.16 |
| 12 | 14.1 99.4 74.4 38.10 | 13.1 99.4 74.0 38.10 | 77.5 100.0 94.8 30.35 | 99.8 100.0 100.0 30.35 | 98.6 99.9 100.0 40.16 |
| 15 | 1.9 99.2 63.6 38.28 | 1.7 99.2 63.1 38.28 | 67.9 99.9 93.1 30.32 | 99.3 100.0 100.0 30.32 | 91.5 99.9 99.5 40.16 |
Robust Watermarking
Robust watermarking embeds or verifies ownership and provenance signals that survive editing, compression, model changes, and deliberate removal.
See all 228 papers