Abstract

Forensic document authentication is used to identify fakes and verify police evidence and other legal documents, making detection and authentication part of forensic science. Existing approaches face issues about low generalizability and low interpretability of results, and inability to detect subtle amid manipulations or cross-modal inconsistencies. ForenXAI addresses this gap with a smart deep learning (DL) system that differentiates visual and textual forgeries through fused ResNet50-CBAM and LSTM with Attention. It cross-verifies using dual-stream cross-modal methods and performs risk assessment, SHAP-based interpretability, and XAI-derived probabilistic risk scoring. It structures the process within multi-stage pre-processing, image-text alignment, anomaly detection, and decision support to achieve legal accuracy. ForenXAI is optimized for system security and trust through real-time monitoring and evidence security, including auditable logs, encryption, and access control to forensic evidence and other sensitive data. Evaluation over multiple documents and signatures found the framework attains an accuracy of 0.9788 with MCC of 0.9576 and G-mean of 0.9788 at 70% training. This further improved to an accuracy of 0.9894 and MCC of 0.9788 at 80% training. Along with this, the framework attains an optimal F1-score of 0.9894. Comparative analysis across processing time, data precision, time taken to encrypt, delays in real time monitoring, and SHAP feature importance with Cycle-GAN, Ta-RNN, and NSVNN emphasizes ForenXAI’s efficiency and interpretability. These findings affirm that ForenXAI is principled, highly interpretable, secure, and extremely practical for forensic applications, thereby endorsing the use of AI systems in the policing and judiciary frameworks.