- ROC-AUC on the clean split
- 0.947
- ROC-AUC after JPEG q60 laundering
- 0.936
- forensic engines, zero servers
- 5
Error Level Analysis, copy-move detection, noise consistency, EXIF forensics and JPEG quantisation analysis run in the browser and fuse into a verdict with a downloadable evidence report; a PyTorch research layer trains on CASIA v2.
Problem
Forgery detection literature is rarely usable by a person with a suspicious receipt. DocuForge had to make peer-reviewed techniques run in a browser tab with measurable accuracy, not just pretty heatmaps.
Approach
- 01
Each engine is one vanilla ES2017 module producing a heatmap and a score; a weighted fusion favours spatial evidence.
- 02
The exact production engines are benchmarked end-to-end in Chromium with Playwright on a reproducible 64-document dataset, including recompression and noise attacks.
- 03
training/ ships an EfficientNet-B0 over ELA maps of CASIA v2 with a full eval harness and model card.
Results
- ROC-AUC on the clean split
- 0.947
- ROC-AUC after JPEG q60 laundering
- 0.936
- forensic engines, zero servers
- 5
- Created
- 26 Jul 2026
- Commits
- 9
- Default branch
- main
- License
- MIT
Stack
- Vanilla JS
- Canvas
- Playwright benchmarks
- PyTorch
- EfficientNet-B0