FakeIDet3-DB: Digital Attack Retouching and Image Patch Extraction for Identity Document Security Evaluation

Researchers proposed FakeIDet3-DB, which the authors describe as the first comprehensive database of digital tampering based on real government-issued identity documents, covering classic tampering such as copy-move and generative AI tampering such as face swapping and image inpainting, with retouching applied to suppress visual artifacts. To comply with data protection regulations such as GDPR, the authors proposed the PACE pseudo-anonymized context patch extraction algorithm, which uses integral image mapping and distance-based non-maximum suppression (NMS) to delineate anonymization masks, maximizing semantic density around masked regions while preventing leakage of personally identifiable information (PII), extracting nearly 5.2M image patches from over 6.4K real and forged document images. Evaluations with state-of-the-art models show that these models struggle to detect and localize attacks from generative and classic techniques, with a detection EER of 32.45% and a localization AUC-ROC of 83.48%.