Abstract: Aiming at the problem that  deepfake  detection algorithms  had low detection effect on neural texture (NT) forgery methods in FaceForensics++(FF++) dataset,  we  proposed a reverse knowledge distillation network (RKD-Net) by improving  fine-grained feature extraction of face images. Firstly, the RKD-Net used  reverse knowledge distillation (RKD) as the main framework, 
preserving  the rich fine-grained information of the input face images. Secondly, spatial and channel reconstruction convolution (SCConv) was inserted between the encoder and the decoder to enhance the representation of fine-grained information from both spatial and channel dimensions. Finally, a residual coordinate attention (RCA) classifier was used to enhance the real and detailed features  output by  the reverse knowledge distillation network, and classify the face images input to the network according to these different features. The experimental results  show that RKD-Net achieves  the best detection effect of NT forgery methods  
while guaranteeing the  detection effect of  other  forgery methods.

Key words: deepfake detection, reverse knowledge distillation, spatial and channel reconstruction convolution, coordinate attention