File size: 3,426 Bytes
580e48b d509ec2 580e48b d509ec2 580e48b d509ec2 580e48b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 | ---
language:
- en
license: apache-2.0
library_name: pytorch
tags:
- image-classification
- forgery-detection
- document-forgery
- resnet
- redrob
- redrob-verify
pipeline_tag: image-classification
---
# redrob-verify — forgery
Forgery detector weights for **[redrob-verify](https://github.com/redrob-labs/redrob-verify)**,
the Redrob verification stack (document OCR, forgery, face match, identity).
- Hub: [`redrob-labs/redrob-verify-forgery`](https://huggingface.co/redrob-labs/redrob-verify-forgery)
- Code: Apache-2.0 in the GitHub repo (`services/forgery/`)
## Intended use
Score a document scan or photo for **tampering likelihood** in `[0, 1]`
(higher = more likely forged). Tuned for ID / KYC–style pages in the
redrob-verify harness (MIDV authentic + synthetic patches).
Not a face deepfake detector. Not a claim of FMIDV cross-domain pass unless you
run that split yourself.
## Model
| | |
|--|--|
| Architecture | ResNet-50 + FFT + HOG streams + upsample localization head |
| Backbone init | torchvision `ResNet50_Weights.IMAGENET1K_V2` (BSD-3) |
| Input | RGB **320×320** |
| Serving score | Image-level sigmoid (localization used at train time) |
| Code | `services/forgery/model.py` |
Files in this repo:
- `model.safetensors` — weights for Hub / `safetensors` loaders
- `forgerynet_apache.pth` — full training checkpoint (`model_state` + metadata); drop-in for redrob-verify `config.yaml`
- `config.json` — image size, recommended threshold, provenance pointers
## Training data (provenance)
| Source | Role | Terms |
|--------|------|--------|
| torchvision ResNet-50 ImageNet-1K V2 | Backbone init | BSD-3 / torchvision |
| MIDV-2020 authentic pages | Train negatives (JPEG-recompressed, **train-split only**) + eval authentic (**held-out docs**) | Follow MIDV / portal terms |
| `tools/gen_forgery.py` synthetic tampers | Train positives + masks (from train docs only) | Synthetic; generated in-repo |
Weights are **not** derived from TruFor.
## Evaluation (in-domain)
Document-disjoint holdout (`tools/split_forgery_holdout.py`): **400** train /
**100** eval authentic IDs, eval **n=200** (100 auth + 100 forged).
Published checkpoint (seed **7**):
- Joint TC2/TC3 feasible ≈ **[0.69, 0.93]**
- Recommended threshold **0.87** → TPR ≈ **0.92**, F1 ≈ **0.82**
Multi-seed check (seeds 7 / 13 / 42; judge by **minimum**):
| Seed | TPR | F1 | t* |
|------|-----|----|----|
| 7 | 0.92 | 0.821 | 0.87 |
| 13 | 0.88 | 0.811 | 0.98 |
| 42 | 0.93 | 0.798 | 0.96 |
| **min** | **0.88** | **0.798** | — |
Protocol: `./run.sh split-forgery-holdout --regenerate-train --rebuild-eval` then
`./run.sh train-forgery` / `./run.sh eval-forgery`.
## Download & run
```bash
# From the redrob-verify checkout
./tools/fetch_models.sh # pulls face + forgery from Hugging Face
# Or Hub only
huggingface-cli download redrob-labs/redrob-verify-forgery \
--local-dir models/forgery
```
Serve with `forgery.backend: forgery_net`, `image_size: 320`, and
`weights_path: models/forgery/forgerynet_apache.pth` (or load `model.safetensors`
via the same `ForgeryNet` class).
## Limitations
- Domain: MIDV + our synthetic generator; other scanners/tampers may need fine-tuning.
- Optional TruFor backend in the code repo is research-only (nonprofit upstream) and is **not** these weights.
## Citation
Cite redrob-verify and MIDV-2020 per their terms when reporting results.
|