With the growing use of Vietnamese citizen identification cards in digital identity verification, reliable forgery detection has become increasingly important, yet research in this area remains limited, especially against ever more realistic manipulations enabled by generative AI. We propose a deep learning approach for detecting and localizing forgeries in Vietnamese citizen identification cards, formulated as pixel-level binary segmentation rather than image-level classification. Using a dedicated dataset of paired genuine and manipulated card images with tampering masks, we train a U-Net with a ResNet-34 encoder to segment forged regions, and convert the predicted masks into bounding boxes for practical use. Grad-CAM is further applied to visualize the regions driving the model's predictions, providing qualitative interpretability alongside the segmentation results. Experiments show strong performance in both pixel-level localization, with a mean IoU of 0.8119, and image-level classification, with an accuracy of 98.21% in distinguishing genuine from forged images. A human evaluation with 45 participants further shows that the visual explanations raise average classification accuracy from 0.59 to 0.88. Overall, the proposed framework offers an effective and practically deployable means of verifying Vietnamese citizen identification cards, a topic that has received limited attention in the literature.

Vietnamese citizen identity card,document forgery detection,U-Net,forgery localization,ResNet,Grad-CAM