CDNET: CLUSTER DECISION FOR DEEPFAKE DETECTION GENERALIZATION

The fast development of deepfake generation technology has caused serious security threats to human society. Many deepfake detection methods have been proposed recently, but most of them show poor generalizability in detecting unseen deepfakes. In this paper, we propose a lightweight cluster decision network (CDNet) to improve generalizability in deepfake detection. We design a selective attention module that decides the attention areas by manually cropping the facial areas (e.g., eyes, nose, and lips). This design can greatly reduce the model size while ensuring the model can focus on the areas of the deepfake artifacts. We also propose a cluster classifier to equally utilize the feature representation, inspired by the contrastive learning. Compared with existing deepfake detection methods, our model can pay attention to the areas of the general deepfake artifacts and has a simple model structure. Extensive experiments show that our method outperforms existing state-of-the-art methods in deepfake detection generalizability and has the minimum model size.