FLiD: Lightweight Field-Localized Forgery Detection for Digital Identity Documents
Researchers have introduced FLiD, a novel lightweight framework designed to detect forgery in digital identity documents used for remote user onboarding. Unlike existing methods tailored for natural images, which struggle with structured documents, FLiD specifically targets critical identity fields such as facial photographs and textual data. The system employs a fine-tuned object detector to localize these fields, followed by a frozen MobileNetV3-Small backbone that extracts compact embeddings. These are classified by a lightweight neural network containing only 191,000 trainable parameters. Performance evaluations demonstrate significant improvements over full-document baselines, with AUC scores reaching 0.880 for face manipulations and 0.954 for text alterations. Furthermore, FLiD outperforms general-purpose manipulation detectors like TruFor and MMFusion while requiring substantially fewer computational resources, specifically 13 times fewer parameters and 21 times fewer FLOPs. This advancement addresses vulnerabilities in digital identity verification systems against localized manipulations, offering a more efficient and accurate solution for security applications.
Wire timeline
FLiD: Lightweight Field-Localized Forgery Detection for Digital Identity Documents
Researchers have introduced FLiD, a novel lightweight framework designed to detect forgery in digital identity documents used for remote user onboarding. Unlike existing methods tailored for natural images, which struggle with structured documents, FLiD specifically targets critical identity fields such as facial photographs and textual data. The system employs a fine-tuned object detector to localize these fields, followed by a frozen MobileNetV3-Small backbone that extracts compact embeddings. These are classified by a lightweight neural network containing only 191,000 trainable parameters. Performance evaluations demonstrate significant improvements over full-document baselines, with AUC scores reaching 0.880 for face manipulations and 0.954 for text alterations. Furthermore, FLiD outperforms general-purpose manipulation detectors like TruFor and MMFusion while requiring substantially fewer computational resources, specifically 13 times fewer parameters and 21 times fewer FLOPs. This advancement addresses vulnerabilities in digital identity verification systems against localized manipulations, offering a more efficient and accurate solution for security applications.
cs.AI updates on arXiv.org