Digital Tampering Detection¶
Definition¶
Specific techniques for detecting digital manipulation of identity document images — photo editing, text replacement, splicing, and AI-generated alterations.
Detection Techniques¶
| Technique | What It Detects | How It Works |
|---|---|---|
| Error Level Analysis (ELA) | Edited regions | Re-compress JPEG, compare error levels — edits show higher residuals |
| Noise inconsistency | Spliced regions from different sources | Different camera sensors produce different noise patterns |
| JPEG ghost detection | Double compression | Edited and re-saved JPEGs show compression ghosts at certain quality levels |
| Copy-move forgery | Cloned regions | SIFT/deep features find duplicate patches within same image |
| Font anomaly detection | Text replacement | Classify font at character level — replaced text uses different font |
| Edge inconsistency | Splicing boundaries | Analyze edge sharpness and artifacts at region boundaries |
| Metadata analysis | Editing software traces | EXIF data reveals if edited with Photoshop, GIMP, etc. |
Deep Learning Approaches¶
| Model | Approach |
|---|---|
| ManTraNet | Pixel-level manipulation detection — no prior knowledge of manipulation type needed |
| MVSS-Net | Multi-view multi-scale — detects both boundary and noise artifacts |
| CAT-Net | Compression artifact tracing — traces JPEG compression history |
| ObjectFormer | Transformer for object-level forgery detection |
Key Takeaways¶
Summary
- Multiple complementary techniques needed — no single method catches all tampering
- ELA + noise analysis are fast baselines; deep learning adds learned pattern detection
- Font analysis is critical for ID documents — most fraud involves text editing
- Cross-validation between OCR text, MRZ, and barcode data catches many edits