ResearchPod Summary
Identity document (ID) authentication is increasingly performed remotely, making systems vulnerable to sophisticated injection attacks. While Generative AI (GenAI) models can now create high-fidelity, localized forgeries, training forensic detection models is difficult due to strict privacy regulations (e.g., GDPR). Most existing databases rely on synthetic templates that lack the complex, high-frequency security patterns—such as holograms and laser engravings—found in real government-issued IDs. This creates a significant domain shift, where models perform well on synthetic data but fail when applied to real-world documents.
To address this, the authors present FakeIDet3-DB, the first comprehensive database of digital manipulations applied to real, government-issued IDs. The dataset includes over 6,400 images featuring eight distinct attack types, ranging from classical methods (copy-move, splicing) to advanced GenAI-driven manipulations (face-swapping, morphing, and text inpainting). A key innovation is the inclusion of a refinement post-processing pipeline that suppresses visual artifacts, resulting in highly realistic forgeries that better simulate real-world threats.
To comply with data protection laws, the authors developed PACE (Pseudo-Anonymized Contextual patch Extraction). Instead of using standard uniform grids, PACE treats patch extraction as a geometrically constrained image processing problem. It uses Integral Image mapping and distance-driven Non-Maximum Suppression (NMS) to extract patches from non-redacted areas. This approach maximizes the semantic density of the data—specifically targeting the "peri-censorship" regions where manipulation artifacts are most likely to reside—while ensuring zero leakage of Personally Identifiable Information (PII).
The authors demonstrate that state-of-the-art forensic models struggle significantly with the high-frequency, complex structures of real IDs, reporting an Equal Error Rate (EER) of 32.45% in detection tasks. By providing a standardized, privacy-compliant benchmark, this work helps bridge the gap between laboratory-based forensic research and the practical requirements of secure, real-world ID verification.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.