Publication

Abstract : The use of deepfake technology has disrupted almost all aspects of digital media, with great implications on privacy, security, and trust. While classic detection methods raised the red flag on manipulated content, they mostly functioned as black boxes without explanations for their conclusions. To counter this, we designed an explainable deepfakeFor detection, whereby both accuracy and interpretability were enhanced in the approach. It combines EfficientNetB0 with an LSTM model to seize spatial and temporal features in videos. Visual features are extracted by EfficientNetB0, whereas sequential dependencies across frames are modeled with an LSTM to identify inconsistencies. For explainability, Grad-CAM generates heatmaps to capture altered areas focusing on key facial features, such as eyes and nose. The disturbed pixels finding, which are accumulated across frames, produced a result dimension of the video-wide, hence simplifying the analysis of edited regions. The system incorporates authentication through password or biometric mechanisms, thus preventing unauthorized access, whereby a web interface permits end-users to upload videos for classifications and to view interactive charts illustrating areas of manipulation. The proposed system attained 99.94% accuracy, thus making an advancement in digital forensics through a reliable and interpretable deepfake detection approach.