Remote identity-verification systems must distinguish bona fide identity cards from print, screen, and composite attacks. The differences between them are often apparent only in subtle texture and color traces whose visibility depends on the device and capture conditions. Because model performance often deteriorates on new documents, we analyze how the choice of foundation model, image preprocessing, and additional binary models affects the overall performance of the system.
The proposed pipeline combines document localization using YOLOv11, a multiclass base classifier, and binary models for detecting individual attack types. The base model distinguishes among bona fide, composite, print, and screen examples and produces an initial document-authenticity score. The included binary models complement this score with targeted assessments of individual attack types, while the final system output is a binary authenticity score. For composite attacks, we compare the use of a full-card crop with the use of a portrait crop. The evaluated base models are CLIP adapted with LoRA, DINOv2, DINOv3, ConvNeXt, EfficientNetV2, CAFormer, and EVA-02.
The approach was evaluated on the PAD-IDCard 2025 and Kid34K datasets using the Equal Error Rate (EER), the Attack Presentation Classification Error Rate (APCER), the Bona Fide Presentation Classification Error Rate (BPCER), and the average score. The most balanced configuration achieved an EER of 1.07% and an average score of 3.78% on the internal test set. In the external evaluation on Kid34K, DINOv2 achieved the lowest EER of 27.24% and the lowest average score of 81.41% among the evaluated base models. The results show that the selective use of binary models and additional crop padding improves system performance. At the same time, generalization to independent data remains challenging.
|
