Occlusion Resilient Adversarial Attack For Person Re-Identification

Deep learning-based person re-identification(Re-ID)methods have achieved the significant performance of matching person images across camera views,which plays an important role in the construction of the smart city.Adversarial attacks on deep learning models help identify potential problems with the models.Recent works of adversarial attacks have explored the serious vulnerability of deep learning-based Re-ID systems.However,the existing attacks are performed with idealized conditions,ignoring the changes brought by the real environment,such as the occlusion of the adversarial patch caused by the walking habits of pedestrians.The original attack method is not robust to occlusion.To this end,this thesis proposes a two-stage adversarial patch generation scheme to better evaluate deep learning-based person re-identification systems.The main contributions are as follows:First,we summarize the walking habits of pedestrians from the collected pedestrian walking data and constructs an occlusion template.The application of the occlusion template verifies that the existing attack methods for Re-ID model are not robust to occlusion.Second,we design a partitioning and fine-tuning training strategy to better combine adversarial patches that are easily occluded and not easily occluded.During the training process,this thesis introduces a contextual loss to penalize the semantic distance of the attacked pedestrian images.Extensive experiments on Market1501 dataset and self-collected dataset demonstrate the performance of the proposed scheme.At last,we design an occlusion resilient adversarial patch optimization scheme that adapts to changing environments.In the optimization process,randomly changing environmental factors are added,and the maximum similarity loss is used to generate adversarial patch with environmental adaptability.The complete adversarial patch generated by this scheme can still have high attack performance under changing environmental conditions.