Objective In view of the current lack of a powerful fake face detection model that can expose fake face images in complex scenes a new network called two-stream manipulation trace network TSMTN is proposed for learning subtle manipulation traces on facial regions in fake images. Methods This method is different from the previous methods of directly learning features from images. Instead it first extracts manipulation traces from the image and then uses the manipulation traces to detect whether the face has been manipulated. The network consists of three key modules spatial domain manipulation trace extraction SDMTE frequency domain manipulation trace extraction FDMTE and feature fusion module FFM based on self-attention mechanism. SDMTE uses convolutional neural networks CNNs to learn subtle manipulation traces in the image spatial domain. FDMTE learns the manipulation traces of high-frequency information in the frequency domain of images. FFM fuses manipulation traces in the spatial and frequency domains to generate final features for classification. Results The experimental results show that the model has good performance and has reached an advanced level on commonly used detection datasets. Conclusion This method shows good robustness and generalization ability and has made some progress which is of great significance.