Towards Robust Person Re-Identification Via Efficient and Generalized Adversarial Training

Person re-identification (ReID) is a particular cross-camera person retrieval problem that involves non-overlapping cameras. Adversarial training (AT) is known to dramatically increase the robustness of deep neural network-based ReID. Unfortunately, the high cost of adversarial samples generation process makes the application of AT impractical. Moreover, the model trained by AT has poor generalization. For unfamiliar adversarial samples generated by non-AT trained attack methods, the model may recognize them incorrectly. In this work, we proposed an efficient adversarial training method for ReID tasks to address these issues. Firstly, the factors that lead to the lack of robustness of current ReID models under adversarial attack are explored, and the reasons for this phenomenon are experimentally analyzed and discussed. Based on our explorations, an efficient adversarial training method is proposed by combining the training method of Free-AT with universal adversarial perturbation (UAP), while keeping the cost of adversarial training the same as normal training. In addition, the strategies of random restart for UAP and partially retaining clean samples for training are proposed, which improve model generalization ability by diversifying adversarial samples. Extensive experiments demonstrate the efficiency of the model training process, as well as the robustness and generalization ability of the model.