In view of the significant intra-class differences between different cameras, and the change of the clothes of some pedestrians, which leads to the degradation of recognition, an end-to-end combination of generative adversarial network (GAN) data generation and person re-identification learning is proposed, which makes up the federated learning framework. Firstly, on the basis of the DGNet network, the residual block (ResBlock) in the generator and discriminator is replaced with DenseBlock, which strengthens feature propagation and avoids the problem of gradient disappearance. Then, the generative module generates high-quality synthetic images by switching appearance and structure encodings. Finally, a new normalization-based attention module (NAM) is added to the discriminative module to suppress less salient weights and pay more attention to the desired target regions. The images generated by the generation module are fed back online to the identification module to identify the true and fake images, and the results are fed back to the person re-identification module for classification and identification (the identification module and the pedestrian re-identification module are shared). The rank-1/mAP on Market-1501 and DukeMTMC-reID datasets reach 95.7%/88.6% and 87.1%/75.7%, respectively.