Deepfake detection via spatial–temporal deep networks: leveraging CNNs and LSTMs for enhanced accuracy

The rapid evolution of deepfake generation methods has raised unprecedented challenges in ascertaining the authenticity and credibility of digital content. This paper offers a thorough benchmark comparative evaluation of deepfake detection models combining spatial and temporal feature descriptions under controlled conditions. The work systematically compares the performance of pre-trained convolutional neural network (CNN) structures—ResNeXt50, XceptionNet, MobileNet, and VGG16—coupled with Long Short-Term Memory (LSTM) networks for detecting forged videos. All CNNs are used as spatial feature extractors, whereas the LSTM module captures temporal relationships between video frames. The experiments are performed on the FaceForensics++ database using k-fold cross-validation to achieve robustness and generalization. Assessment is made with respect to several performance measures such as accuracy, precision, recall, F1-score, and ROC-AUC in order to give a fair and overall comparison among architectures.