Vereye/benchmark-synthid-image-protocol
Viewer • Updated • 84k
Vereye
company
AI & ML interests
Content provenance, invisible watermarking, owner identification, tamper localization, AI Act Article 50 compliance
Recent Activity
Organization Card
Vereye
Vereye is an invisible image watermark that names the registered owner of any copy of an image, read blind from the copy alone, with a computed bound on naming a wrong owner. Vereye is an early-stage startup founded by Adam Weisman; Gil Weisman is co-founder, advisor and engineering reviewer.
On Meta's own Watermark Anything benchmark (200 COCO photos, 35 attacks, identical attacked pixels for every method), Vereye recovered the exact owner key from 98.74 % of attacked copies, against 70.27 % for Meta's Watermark Anything and 66.67 % for Adobe's TrustMark, with no false alarm on 7,000 attacked unmarked copies (Meta: 162, Adobe: 178). On Google DeepMind's SynthID-Image protocol at 512 px, on 300 photos never used in development, it named the exact owner of 99.70 % of attacked copies at random strength and 95.80 % at the most severe, with its stronger mark for small images; Google reports higher values for its closed SynthID on its own images.
What it does
When an image is sealed, Vereye writes a random 64-bit owner key with a 24-bit check into its pixels and records the owner in a signed registry. From any copy, cropped, compressed, rotated, recoloured, screenshotted or pasted into another picture, it:
- names the registered owner, or says that no seal could be read;
- says whether the copy is pixel-identical to the sealed original;
- maps the areas that changed, without the original.
The decision is an exact key that passes its check and is found in the registry, not a score above a threshold. Sealing and reading are deterministic and training-free: no machine-learning model, and the geometry of a copy is solved in closed form. The registry stores no image and nothing from which one could be rebuilt, and its records are signed with post-quantum signatures. The mark lives in the pixels, so it still reads after a screenshot strips C2PA Content Credentials or EXIF metadata: it complements metadata provenance rather than replacing it.
Results
| test | Vereye | Meta WAM | Adobe TrustMark |
|---|---|---|---|
| Meta's WAM protocol, 200 COCO photos: exact key or message recovered | 98.74 % | 70.27 % | 66.67 % |
| Same: false alarms on 7,000 attacked unmarked copies | 0 | 162 | 178 |
| SynthID-Image protocol, 512 px, current system, 300 photos never used, stronger mark for small images (PSNR 35.58 dB): exact owner, random / most severe | 99.70 % / 95.80 % | ||
| Same protocol, earlier reader, 100 held-out photos, pre-registered, standard mark (PSNR 41.6 dB) | 99.03 % / 86.38 % | ||
| Same protocol, earlier reader, the 50 of those photos all three ran on: exact message, random / most severe | 99.17 % / 86.21 % | 80.41 % / 68.14 % | 75.59 % / 62.97 % |
| Tamper set, 40 photographs: edits found from the copy alone / untouched copies flagged | 197 of 200 / 0 of 80 | 163 of 200 / 2 of 80 | |
| Attacked unmarked copies given an owner, all benchmarks | 0 of 46,467 |
Read before comparing
- SynthID is closed, so it was not run. Google reports 99.98 % / 99.72 % detection and 99.96 % / 98.84 % message recovery for SynthID-O on its own 10,000 ImageNet images, ahead of Vereye at both strengths.
- Every SynthID-protocol verdict is public, with the script that scores it: anyone can re-check those numbers.
- Wrong owner: an analytic bound of 5.7 × 10⁻¹³ per check with 10¹¹ registered images, under the assumption stated in the paper; measured, 0 of 46,467.
Links
- Every verdict (SynthID-Image protocol): Vereye/benchmark-synthid-image-protocol
- Paper, Identification, Not Detection (method, results, limits): vereye.ai/paper
- Live demo: vereye.ai/demo · Press kit: vereye.ai/press
- Contact: adam@vereye.ai
Patents pending.
models 0
None public yet