DeepTrust: An Explainable, Real-Time Audio Forensic System Robust to Social Media Compression

Abstract

The rapid rise of advanced generative AI has made it harder to tell the difference between real and fake audio, which creates serious security risks. Current detection systems face two main issues: the "Black-Box" nature of deep learning models, which are hard to interpret, and the "Compression" challenge, where social media algorithms lower audio quality. This research presents DeepTrust, a new forensic system aimed at addressing these challenges.
DeepTrust uses a cutting-edge CNN-Transformer hybrid structure. It takes advantage of the local feature extraction strengths of Convolutional Neural Networks (CNNs) while also utilizing the global context modeling of Transformers. The system trains and validates with the Wavefake benchmark, making it resistant to the latest spoofing methods. A significant contribution of this project is the inclusion of an Explainable AI (XAI) module. With Grad-CAM (Gradient-weighted Class Activation Mapping), DeepTrust creates visual heatmaps that highlight specific forged parts of an audio file. This gives forensic investigators clear evidence instead of just a yes-or-no answer.

Additionally, the system is designed for practical use. It has a real-time web dashboard that keeps high detection accuracy, even under heavy social media compression. Experimental results show that DeepTrust provides a strong, clear, and scalable solution for audio forensics in today's digital world.