Explainable AI for Identity-Document Fraud Detection in Digital Forensics Using Grad-CAM++

Abstract

Artificial intelligence (AI) systems can assist digitalforensic investigators in reviewing large volumes of digital evidence and identifying patterns for further examination. However, some AI models remain difficult to use in forensic and legal settings because their internal decision-making processes are not readily open to examination. Explainable AI (XAI) addresses this limitation by providing methods for examining which features or regions contributed to a model output. This study evaluated Grad-CAM++ as an explainability method for AIassisted identity-document fraud detection. A multitask ResNet-50 model was trained and evaluated on IDNet, a research dataset containing synthetic U.S. driver's-license images. The model used multiple task heads to capture overall fraud presence, broad faceand text-related evidence, and subtype-specific fraud patterns. Grad-CAM++ was applied to selected task heads, and the analysis combined heatmap visualization, salient-region extraction, annotation-based localization measures, and blur-ablation response analysis. The overall fraud-detection head showed high recall and strong ranking performance as a screening output. Across five fraud types, localization and perturbation behavior varied: face-related manipulations and text-field replacement showed the clearest region alignment, whereas inpainting-based rewriting and crop-and-replace produced more diffuse localization. A representative face-replacement case further showed that the Grad-CAM++ explanation was concentrated mainly on the portrait region and that perturbing this region reduced the corresponding model output. These findings indicate that Grad-CAM++ can provide a reviewable visual and quantitative explanation layer for AI-assisted identity-document fraud detection.