Deep fake technology generates highly realistic virtual images through AI technology, and its misuse poses a serious threat to social stability and national security. In order to cope with the problems of easy generation, difficult recognition and fast propagation, which are unique to deep fake technology, this paper proposes a joint model combining CNN and Transformer for deep fake image detection. The joint model combines Transformer's ability to extract image spatial information on long input sequences and CNN's feature extraction ability to transform the deep fake detection problem into an image classification problem to improve the accuracy of detection and the generalization performance in different detection scenarios. The final classification probability values represent the image authenticity detection results, where 0 < y ≤ 0.5 denotes the Real class and 0.5 < y ≤ 1 the Fake class. In this paper, the experimental results on Face Forensics++ dataset show that the method's deep fake image detection accuracy reaches 96.42%, which has better robustness and generalization performance compared with other methods, and can effectively solve the problem of deep fake image detection. In addition, this paper also develops a practical deep fake detection platform, which realizes the visualization function of image authenticity detection, providing a direct possibility for practical application.