A Hybrid Xception-LSTM Model with Channel and Spatial Attention Mechanism for Deepfake Video Detection

The great strides taken in recent times in image and video manipulation have raised serious concerns. Deepfake technology uses deep learning approaches to create highly realistic, astonishing content. Detecting such videos is the only promising defense against such fraudulent data. To counter the malicious intent of the user, a deepfake detection model is proposed that employs channel and spatial attention mechanisms(CBAM) along with Xception and LSTM pretrained models. Xception uses depthwise separable convolution to capture the latent spatial artifacts. LSTM captures the discrepancies among the manipulated sequences; hence, this hybrid ensembling of models allows the learning of powerful features. The evaluation is performed on the recently proposed Div-DF dataset consisting of varied video manipulation like face swap, facial reenactment, and lip-sync. It shows that the model works well (Accuracy~ 93 % & AUC ~ 0.98) on the diversified dataset and easily beats the score of various state-of-the-art deepfake detection and image classification models.