Advanced Temporal Analysis for Deepfake Detection using XceptionNet in Media Forensics
Kabilan M, S Logeswaran, S Kubersrinivash, Alfred Daniel J · 2025
Deepfakes videos through advanced AI models, pose a significant threat due to their deceptive realism. This paper presents a deepfake detection framework based on XceptionNet, which provides depthwise separable convolutions for fine-grained image analysis. The proposed system processes both static images and video by detecting and aligning faces, enabling consistent frame-level analysis. Furthermore, for videos, each frame is analyzed using XceptionNet, followed by temporal feature extraction through LSTM and CNN to capture subtle facial expressions and motion variations. The model was evaluated on multiple datasets, achieving strong precision, recall, and F1-scores at both frame and video levels. Results show that the XceptionNet-based approach reliably distinguishes real from manipulated content. However, limitations remain due to dataset diversity and computational complexity, highlighting directions for future research.