CDNet: Cluster Decision for Deepfake Detection Generalization

The fast development of deepfake generation technology has caused serious security threats to human society. Many deep-fake detection methods have been proposed recently, but most of them can only show high detection performance for the deepfakes generated by the similar techniques with the training dataset. To improve the ability of detecting unseen types of deepfakes, some deepfake detection methods have constructed self-generated datasets to train their models. However, the artifacts on these self-generated datasets are usually caused by some specific face-blending algorithms and lack of generality. In this paper, we propose cluster decision network (CDNet) to improve the deepfake detection generalizability. We design a selective attention module that decides the attention areas by manually cropping the facial areas (e.g., eyes, nose, and lips), which greatly reduce the model size and ensure a small model size. Inspired by the contrastive learning, we also propose a cluster classifier to equally utilize the feature representation. Extensive experiments show that our method outperforms existing state-of-the-art methods in deepfake detection generalizability and has the minimum model size.