Leveraging Spatiotemporal Dependencies for Enhanced Deepfake Video Detection

The rapid advancement of deepfake generation technologies has made synthetic videos increasingly difficult to distinguish from authentic human recordings, posing serious threats to media authenticity and digital security. This study proposes a hybrid deep-learning framework that leverages both spatial and temporal dependencies for enhanced deepfake video detection. The model integrates a ResNeXt-based Convolutional Neural Network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) network for temporal sequence modeling, enabling the detection of subtle artifacts and motion inconsistencies introduced during synthesis. Experiments conducted on three benchmark datasets, FaceForensics++, CelebDF, and the Deepfake Detection Challenge (DFDC), show that the proposed ResNeXt + LSTM architecture achieves accuracies of$\mathbf{9 5. 7 3 \%}, \mathbf{9 7. 2 7 \%}$, and$\mathbf{8 5. 8 9 \%}$, respectively, outperforming standalone CNN and RNN models. These results demonstrate the effectiveness of spatiotemporal feature fusion in improving detection robustness and highlight the potential of hybrid architectures to preserve digital media integrity in real-world scenarios.