IntegriFy: An Explainable AI-Based Framework for Deepfake Audio and Video Detection

Abstract

The rapid spread of manipulated media, particularly deepfake audio and video, poses a significant threat to information integrity on digital platforms. To address this challenge, a web-based detection system called IntegriFy is developed using artificial intelligence techniques. The system integrates a Random Forest classifier trained on Mel-Frequency Cepstral Coefficient (MFCC) features for audio detection with a MobileNetV2 Convolutional Neural Network (CNN) for video detection. To enhance interpretability, Grad-CAM heatmaps are incorporated for visual explanation of video predictions, while textual explanations are provided for audio classification results. The backend infrastructure, hosted on Firebase and DigitalOcean, securely manages file uploads, user authentication, and historical tracking, while the frontend interface is designed for usability and accessibility. Experimental evaluation demonstrates 96% accuracy for audio detection and 78% accuracy for video detection, indicating reliable performance in identifying manipulated media. Future enhancements include improving video detection accuracy, implementing multi-factor authentication, enabling real-time detection, and providing shareable result links. Overall, IntegriFy offers an accessible and explainable solution to mitigate digital misinformation and support users including social media moderators, journalists, and the general public in assessing media authenticity.