Hybrid Xception-ResNext BiLSTM Model with Attention and Visual Synchronization for Robust Deepfake Detection

Deepfake technology has become a significant problem since it allows for the creationof compelling manipulated videos. This research presents a novel hybrid deepfake detection system that combines the Xception and ResNext networks with a Bidirectional Long Short-Term Memory (BiLSTM) model and Convolutional Block Attention Module (CBAM). Apart from focusing on the temporal and spatial anomalies of deepfakes, our model also includes visual synchronization analysis and physiological signal detection. Extensive experiments are conducted across several datasets, exhibiting cutting-edge performance across various manipulation scenarios. The proposed method achieves good generalization, as they are in real-world circumstances.