Deepfake face video detection and interpretation using an explainable XceptionNet and Long Short-Term Memory framework

Abstract Deepfakes are a serious threat to digital trust and information integrity because AI-generated media is becoming more sophisticated. Since human perception cannot distinguish between real and fake visual content, automated and reliable detection systems are desperately needed. In order to simultaneously model subtle temporal inconsistencies and spatial artifacts in deepfake videos, this paper introduces DeepTemporalFace, a dependable and understandable deep learning framework. A refined XceptionNet is incorporated into the system to successfully capture minute manipulation traces and extract discriminative spatial features from facial regions. To identify abnormal transitions that are frequently overlooked by traditional frame-based classifiers, a Long Short-Term Memory (LSTM) network is integrated to learn sequential temporal dependencies and motion-based anomalies. Importantly, DeepTemporalFace has an extensive Explainable AI (XAI) suite, which includes Temporal Attention Analysis to identify key frames in the sequence and LIME for local pixel-level justification. Grad-CAM improves forensic analysis and user trust by displaying the prominent facial regions that impact the model’s decisions. Extensive experiments show that the model achieves high accuracy, strong resilience under compression, and improved generalization when compared to conventional CNN-only or RNN-only baselines. DeepTemporalFace was trained and evaluated on the FaceForensics++ dataset. The results demonstrate a balance between accuracy, robustness, and transparency, which offer a reliable and effective solution for secure multimedia authentication and deepfake forensics.