Primary Investigator:
Marcus Rogers

Project Members
Yujin Lee, Umit Karabiyik and Marcus Rogers

Artificial intelligence (AI) has been increasingly used in digital forensic investigations. AI-based methods have been applied to tasks such as document fraud detection, image analysis, and large-scale data classification. These methods can support investigators by automating repetitive procedures and identifying patterns that may be difficult to detect through manual inspection alone, thereby helping investigators work more efficiently. However, some AI models operate as black boxes, which can present challenges in legal and forensic settings where model outputs may need to be interpreted and justified. This limitation has motivated the use of explainable artificial intelligence (XAI), which aims to clarify the factors that influence predictions. In this study, we used the IDNet Dataset, a synthetic dataset of identity documents developed for research purposes. For document authenticity classification, we employed a ResNet50-based classifier. To interpret the model outputs, we applied SHAP (SHapley Additive exPlanations) to estimate feature attributions for the input images and aggregated the resulting attributions into a patch-based representation. This patch-based aggregation provides a summary of influential regions and distinguishes between positive and negative contributions within the document image. The resulting visualizations help illustrate which regions contributed to the model prediction and how those contributions were distributed across the input. These explanations support investigators and analysts in interpreting model behavior and in communicating findings from analysis conducted with AI support within the digital forensic process to experts, non-expert audiences, and legal professionals.