README.md

---
license: cc-by-nc-sa-4.0
pretty_name: "AIForge-Doc v1 — AI-Forged Document Images"
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - image-classification
  - image-segmentation
tags:
  - scam-ai
  - document-forgery
  - document-forensics
  - ai-generated
  - gemini
  - ideogram
  - tampering-detection
  - document-fraud
gated: true
extra_gated_heading: "Access Scam.AI Research Dataset"
extra_gated_description: |
  This dataset is released by Scam.AI for academic research and non-commercial use.
  Please share a couple of details so we can understand how the community is using our work.
  Access is granted automatically after submission.
extra_gated_button_content: "Agree and access dataset"
extra_gated_fields:
  Full Name: text
  Email: text
  I agree to use this dataset for non-commercial research purposes only and to cite the corresponding paper if I publish results based on it: checkbox
---

# AIForge-Doc: A Benchmark of AI-Forged Document Images

[![License: CC BY 4.0](https://img.shields.io/badge/License-CC%20BY%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by/4.0/)
[![Format: DocTamper-compatible](https://img.shields.io/badge/Format-DocTamper--compatible-blue)]()
[![Images: 4,061](https://img.shields.io/badge/Images-4%2C061-green)]()

**AIForge-Doc** is the first large-scale benchmark of AI-forged document images, targeting
financial and identity document fraud.  Every tampered image was produced by a
diffusion-model inpainting pipeline — a threat model that existing forgery detectors
cannot reliably handle.

---

## At a Glance

| Attribute | Value |
|---|---|
| Total forged images | **4,061** |
| Training split | 3,249 (80 %) |
| Testing split | 812 (20 %) |
| Authentic baseline images | 812 (mirror of test split) |
| AI inpainting tools used | 2 (Gemini 2.5 Flash Image, Ideogram v2 Edit) |
| Source datasets | CORD v2, WildReceipt, SROIE, XFUND |
| Document types | Receipts (89.7 %), Forms (10.3 %) |
| Languages | 9 (EN, ID, DE, IT, ES, FR, PT, ZH, JA) |
| Output format | DocTamper-compatible (binary grayscale masks) |
| Mask convention | `0` = authentic · `255` = tampered pixel |

---

## Directory Layout

```
AIForge-Doc/
├── TrainingSet/
│   ├── images/       # 000000001.png … 000003249.png
│   └── masks/        # same filenames; 0=authentic, 255=tampered
├── TestingSet/
│   ├── images/       # 000000001.png … 000000812.png
│   └── masks/
├── metadata.jsonl    # full provenance for every image (see schema below)
├── README.md         # this file
└── DATASHEET.md      # Datasheets for Datasets (Gebru et al., 2021)
```

File names are **9-digit zero-padded integers** (`000000001.png`), identical to the
DocTamper dataset convention so that existing evaluation pipelines require no modification.

---

## Provenance — metadata.jsonl Schema

Each line is a JSON object with the following fields:

| Field | Type | Description |
|---|---|---|
| `spec_id` | str | Unique forgery spec identifier |
| `image_id` | str | Original image ID from source dataset |
| `source_dataset` | str | `cord` / `wildreceipt` / `sroie` / `xfund` |
| `doc_type` | str | `receipt` or `form` |
| `language` | str | ISO 639-1 language code |
| `field_name` | str | Annotation key of the tampered field |
| `original_value` | str | Ground-truth text before tampering |
| `forged_value` | str | Synthesised replacement text |
| `bbox_xyxy` | list[int] | Tampered region `[x1, y1, x2, y2]` in full-image pixels |
| `assigned_tool` | str | Inpainting model used (`gemini-nano` / `qwen-inpaint`) |
| `split` | str | `training` or `testing` |
| `new_id` | str | 9-digit zero-padded filename stem |
| `final_image_path` | str | Absolute path on generation machine |
| `final_mask_path` | str | Absolute path on generation machine |
| `generated_at` | str | ISO 8601 timestamp |

---

## Dataset Statistics

### By Inpainting Tool

| Tool | API Provider | Images | Share |
|---|---|---|---|
| Gemini 2.5 Flash Image (`gemini-nano`) | Google / OpenRouter | 3,639 | 89.6 % |
| Ideogram v2 Edit (`qwen-inpaint`) | fal.ai | 422 | 10.4 % |

### By Source Dataset

| Source | Document Type | Images | Languages |
|---|---|---|---|
| WildReceipt | Receipt | 1,696 | EN |
| CORD v2 | Receipt | 1,000 | ID |
| SROIE | Receipt | 946 | EN |
| XFUND | Form | 419 | DE, IT, ES, FR, PT, ZH, JA |

### By Language

| Language | Code | Images |
|---|---|---|
| English | `en` | 2,642 |
| Indonesian | `id` | 1,000 |
| Italian | `it` | 81 |
| German | `de` | 78 |
| Spanish | `es` | 67 |
| French | `fr` | 56 |
| Portuguese | `pt` | 62 |
| Chinese | `zh` | 38 |
| Japanese | `ja` | 37 |

### Most Commonly Tampered Field Types

| Field | Count |
|---|---|
| Telephone / store address | 1,388 |
| Free-form text | 880 |
| Store address | 457 |
| Form answer | 419 |
| Menu price | 404 |

---

## Forgery Generation Pipeline

Each forgery is created in four steps to prevent global image drift:

1. **Field selection** — A numeric or key field (price, date, ID, phone) is chosen from
   the source annotation and a plausible replacement value is generated.
2. **Context crop** — The bounding box is expanded 50 % on each side (minimum 100 px) to
   provide font and colour context for the inpainting model.
3. **Masked inpainting** — The context crop is inpainted with the replacement text using
   a diffusion-model API (white mask = tamper region, black = preserve).
4. **Patch-back** — Only the exact field bbox region is pasted back into the full image;
   the ground-truth mask marks those pixels as `255`.

This technique ensures that only the tampered field is replaced while the rest of the
document (background, typography, logos) remains authentic.

---

## Baseline Results

Evaluated on the **TestingSet** (812 forged + 812 authentic paired images).

### Image-Level Detection (AUC-ROC)

| Method | AUC | 95 % CI | AP |
|---|---|---|---|
| TruFor (Guillaro et al., 2023) | **0.751** | [0.726, 0.776] | 0.709 |
| DocTamper (Qu et al., 2023) | 0.563 | [0.535, 0.591] | 0.564 |
| GPT-4o (zero-shot) | 0.509 | [0.481, 0.537] | 0.516 |

### Pixel-Level Localisation (TruFor only)

| Metric | Value |
|---|---|
| IoU | 0.358 |
| F1 | 0.434 |
| Pixel-AUC | 0.916 |

**Key finding:** Even the best-performing detector (TruFor) achieves only 0.751 AUC —
well below the ≥ 0.95 reported on traditional Photoshop-tampered benchmarks.
DocTamper and GPT-4o are near-random, confirming that AI-generated forgeries represent
a qualitatively different and substantially harder threat model.

---

## Licence

The forged images are derived from:

- **CORD v2** — [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
- **WildReceipt** — [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
- **SROIE** — ICDAR 2019 research use
- **XFUND** — [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

The AIForge-Doc dataset itself (forged images + masks + metadata) is released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).  You are free to share and
adapt the material for any purpose provided you give appropriate credit.

---

## Citation

If you use AIForge-Doc in your research, please cite:

```bibtex
@dataset{aiforgedoc2026,
  title   = {{AIForge-Doc}: A Benchmark of AI-Forged Document Images},
  year    = {2026},
  note    = {Dataset paper under submission},
  url     = {https://github.com/YOUR_ORG/aiforge-doc}
}
```

---

## Contact

For questions about the dataset or to report issues, please open a GitHub issue or
contact the authors at [your-email@institution.edu].

---

## Related Research from Scam.AI

This dataset is part of Scam.AI's broader research portfolio on deepfake detection, synthetic media forensics, and adversarial robustness. Other relevant work from our group:

- **DOCFORGE-BENCH: A Comprehensive Benchmark for Document Forgery Detection and Analysis** — Zhao, Xia, Wei et al. (arXiv:2603.01433)
- **When the Forger Is the Judge: GPT-Image-2 Cannot Recognize Its Own Faked Documents** — Wu, Zhou, Ng et al. (arXiv:2604.25213)
- **AIForge-Doc: A Benchmark for Detecting AI-Forged Tampering in Financial and Form Documents** — Wu, Zhou, Xu et al. (arXiv:2602.20569)
- **Can Multi-modal (reasoning) LLMs detect document manipulation?** — Liang, Zewde, Singh et al. (Google Scholar)

Browse our full publications list and dataset catalog at **[scam.ai/research](https://www.scam.ai/en/research)**.

## About Scam.AI

Scam.AI builds detection systems for AI-driven fraud — deepfakes, document forgery, AI-generated synthetic media, and adversarial attacks against identity verification. Learn more at **[scam.ai](https://www.scam.ai)**.