多尺度梯度对抗样本生成网络

Traditional person re-identification (ReID) adversarial attack methods hold some limitations, such as the dependence on the registry(Gallery) to generate adversarial examples and too single examples generation method . To address these problems, an efficient ReID adversarial attack model, multi-scale gradient adversarial examples generation network (MSG-AEGN), is put forward. MSG-AEGN is based on the multi-scale gradient adversarial networks. A multi-scale network structure is adopted to obtain different semantic levels of the input images and the intermediate features of the generator. An attention module is adopted to convert the intermediate features of the generator into multi-scale weights, thereby modulating the image pixels. Finally, the network outputs high-quality adversarial examples to confuse the ReID models. On this basis, an improved adversarial loss function based on the average distance of image features and the triplet loss is proposed to constrain and guide the training of MSG-AEGN. Experiments on three pedestrian ReID datasets, namely Market1501, CUHK03 and DukeMTMC-ReID, show that the proposed method produces promising attack effects on both the mainstream Re-ID models based on deep convolutional neural networks and the transformer networks. Moreover, MSG-AEGN possesses the advantages of low required attack energy and high structural similarity between adversarial samples and original images.