CascadeNet: A Two-Stage Hybrid Learning Framework for Explainable Deepfake Forensics

This paper presents a novel two-stage deepfake detection framework that helpsin finding the exact locations of the region of the image where a deepfake hashappened. In the first stage, a binary classification approach is used to find outwhether an image or video frame has undergone any manipulation or not. Thisclassifier achieves a high accuracy of 99.9% in classifying whether the imageis deepfake or not while maintaining computational efficiency with the help oftransfer learning and data augmentation methods. After the image is successfullyclassified as a deepfake, the second stage uses a U-Net segmentation model with aResNet encoder to find out the exact regions within each frame where a deepfakehas happened, providing pixel-level manipulation boundaries. Extensive experi-ments on the challenging FF++ dataset demonstrate 99.9% accuracy in detectionand 99.8% Receiver Operating Characteristic - Area Under Curve (ROC-AUC)Score in localization tasks, with minimal false positives. The proposed architec-ture significantly decreases computational costs and time by specifically choosingonly those frames that are deepfake and using a pre-trained model to helpreduce time. The model’s efficacy is also evaluated using various post-processingtechniques.