kriyam-ai/kriyam-tamperflow

Kriyam TamperFlow

The first document tampering detection benchmark built specifically for Indian documents, with a built-in compression stress-test that exposes how quickly forensic models degrade on real-world scanned material.

---

Dataset Summary

State-of-the-art document forgery detectors — CAT-Net, DTD, MVSS-Net, CAFTB, and others — rely on JPEG compression artifacts as their primary forensic signal: inconsistencies in DCT coefficients, block boundaries, and noise patterns left behind when an image is edited and re-saved. This works well on pristine photographic images.

Indian documents, invoices, medical records, and financial forms are routinely scanned, photocopied, emailed, and re-saved. Each JPEG pass erases the very signal these models depend on. A model achieving 90%+ AUC on CASIA2 or Columbia can collapse to near-random performance on a scanned Indian document.

Kriyam TamperFlow v1.0 makes this failure mode measurable:

• 1,050 documents (700 tampered, 350 authentic) drawn from Indian document types: ID cards, invoices, medical records, and financial forms
• Pixel-aligned bounding-box annotations covering five tampering techniques
• Three compression tiers (C0 pristine → C2 double-pass → C4 photocopy simulation) to quantify how model performance degrades as forensic signals are erased
• 3,150 image files total — one triplet per source document

The benchmark evaluates models on two tasks: binary document classification (authentic vs. tampered) and tampered region localisation. Results are reported per compression tier, per tamper type, and with bootstrap confidence intervals so model comparisons are statistically grounded.

---

Supported Tasks

Task 1 — Document Tamper Classification

Goal: Given a document image, classify it as authentic or tampered.

Label	Field in annotation	Description
authentic	"is_authentic": true	Unmodified source document
tampered	"is_authentic": false	One or more regions have been altered

Evaluated with document-level AUC-ROC (using the model's continuous confidence score) and F1 at the optimal decision threshold.

Task 2 — Tampered Region Localisation

Goal: Given an image predicted as tampered, predict the bounding boxes of all altered regions.

Scoring:

1. Build an IoU matrix between ground-truth and predicted bounding boxes
2. Run Hungarian optimal assignment
3. Pairs with IoU ≥ 0.1 → True Positive; unmatched ground-truth boxes → False Negative; unmatched predictions → False Positive
4. Report Region Precision, Region Recall, and Region F1

Models that output heatmaps should convert to bounding boxes via connected-component labelling at a chosen threshold (record the threshold in your model card).

---

Dataset Structure

File naming

images/kriyam_{index:04d}_{tier}.png      # image files
annotations/kriyam_{index:04d}.json       # one annotation per source document

• index: zero-padded 4-digit integer (0001 – 1050)
• tier: C0, C2, or C4
• Authentic and tampered images are mixed in the same folder. The filename reveals nothing about authenticity — that information lives exclusively in the annotation.

Compression tiers

Tier	Code	Processing
Pristine	C0	Lossless PNG. No re-compression. Full forensic signal intact.
Double-pass	C2	Two JPEG saves: Q=85 then Q=80. Simulates a typical scan-share-rescan cycle.
Photocopy-sim	C4	C2 → BMP round-trip → Gaussian blur (σ=0.5) → additive noise → JPEG Q=70. Nearly erases DCT-based forensic signals.

Folder layout

indic-docforgebench-dataset/
├── README.md
├── LICENSE
├── metadata.jsonl
├── images/              ← 3,150 PNG files  (1,050 × 3 tiers)
│   ├── kriyam_0001_C0.png
│   ├── kriyam_0001_C2.png
│   ├── kriyam_0001_C4.png
│   └── …
└── annotations/         ← 1,050 JSON files  (one per source document)
    ├── kriyam_0001.json
    └── …

---

Annotation Format

One JSON file per source document, shared across all three compression tiers. Bounding boxes were annotated using LabelMe and are defined in the C0 (pristine) coordinate space, remaining valid for C2 and C4 images which have identical dimensions.

{
  "id": "kriyam_0001",
  "image_w": 2066,
  "image_h": 2921,
  "is_authentic": false,
  "regions": [
    {
      "region_id": 1,
      "x": 80,
      "y": 862,
      "w": 1751,
      "h": 765,
      "tamper_types": ["splice"]
    }
  ]
}

Field reference

Field	Type	Description
id	string	Unique document identifier, matches the filename stem
image_w, image_h	int	Image dimensions in pixels
is_authentic	bool	true → unmodified document; regions will be []
regions	array	Tampered regions (empty for authentic documents)
region_id	int	1-indexed identifier within this document
x, y	int	Top-left corner of bounding box (pixels)
w, h	int	Width and height of bounding box (pixels)
tamper_types	string[]	One or more of: copy_move, splice, text_replace, inpaint

A single region can carry multiple tamper types simultaneously (e.g. a region that was copy-moved and then inpainted to conceal the boundary).

---

Statistics

Document split

Category	Count
Tampered	700
Authentic	350
Total	1,050

Tamper type distribution

Counts are per document (a document with regions of different types is counted once per type).

Tamper type	Documents
inpaint	525
splice	204
text_replace	89
copy_move	65

Note: totals exceed 700 because a single document can have multiple tamper types.

---

Intended Uses

• Benchmark evaluation: Measure how document forgery detectors perform on Indian documents, across compression conditions. The companion evaluation script produces Region-F1, Doc-AUC, FPR, and the Compression Robustness (CR) score — 1 − (AUC_C0 − AUC_C4) / AUC_C0 — for each model.
• Compression robustness research: Study how JPEG re-compression degrades forensic signals, and develop detectors that are robust to it.
• Indic document understanding: The documents span multiple Indian document types and scripts, providing a realistic test bed for models aiming at real-world deployment in India.

Not intended for: Training production fraud-detection systems, any commercial use, or tasks involving real PII.

---

Ethical Considerations

• Modified documents: The dataset contains manipulated images of Indian origin documents. These are provided strictly for academic research into forgery detection and must not be used to create, distribute, or study methods for committing document fraud.
• No real PII: All personally identifiable information visible in source documents has been removed or synthetically replaced. No real names, ID numbers, or biometric data are present.
• CC BY-NC 4.0: The NonCommercial restriction is intentional. Using this dataset to train or improve commercial fraud-detection products requires a separate agreement with the authors.
• Responsible disclosure: If you discover that a real identity document was inadvertently included, please contact the authors immediately.

---

Licensing

Component	License
Dataset (images + annotations)	CC BY-NC 4.0
Benchmark evaluation code	Apache 2.0

The CC BY-NC 4.0 license permits sharing and adaptation for non-commercial purposes with attribution. Academic research (including at corporate research labs where results are published) is considered non-commercial.

---

Citation

If you use Kriyam TamperFlow in your research, please cite:

@misc{kriyamTamperFlow2026,
  title   = {Kriyam TamperFlow: A Document Tampering Detection
             Benchmark for Indic Documents},
  author  = {Jana, Avishek and others},
  year    = {2026},
  url     = {https://huggingface.co/datasets/kriyam-ai/kriyam-tamperflow}
}

Paper: https://zenodo.org/records/21972485

Repository DOI: https://doi.org/10.5281/zenodo.21469087