Explainable deep learning framework for advanced deepfake video manipulation detection

Abstract

The growing sophistication of deepfake technologies has emerged as a critical threat to the credibility of digital media by generating highly realistic yet fabricated visual content. This erodes public trust, elevates security vulnerabilities, and challenges information integrity across online platforms. Despite notable advancements, existing research still suffers from limited data diversity, insufficient model explainability, and inadequate model evaluation. To overcome this limitation, a framework for detecting deepfake video manipulation by using a transfer learning approach was introduced. Each extracted frame was processed by a convolutional neural network (CNN)-based model to obtain frame-level predictions, which were subsequently aggregated to produce the final video-level prediction using a predefined threshold. The publicly available, widely adopted FaceForensics++ dataset was used, which contains high-quality videos generated using advanced manipulation techniques. Various CNN architectures, including Xception, Densenet121, InceptionResNetV2, ResNet50, and EfficientNetB3, were explored along with rigorous hyperparameter tuning. Among these, the Xception architecture outperformed others by achieving a test accuracy of 94.5%. Gradient-weighted class activation mapping (Grad-CAM), generalized gradient-based visual explanations (Grad-CAM++), and Shapley additive explanations (SHAP) were employed to enhance model explainability by visualizing the key regions that influence deepfake detection. The research offers an effective approach to address deepfake threats and safeguard information integrity in contemporary industry 4.0.