ASL: Adversarial Attack by Stacking Layer-wise Relevance Propagation

As the adoption of Deep Neural Networks (DNNs) continues to expand in various AI application fields, the concerns about their security and robustness increase. Adversarial samples can test the security and robustness of DNNs. The existing adversarial samples with low transferability and therefore cannot effectively test unknown DNNs. To improve transferability, some adversarial attacks have employed attention mechanisms to obtain and modify the feature information. Though the single attention mechanism has demonstrated good performance, it may fail to obtain complete and accurate feature information. Employing stack attention mechanisms is crucial in obtaining fully complete and accurate feature information, resulting in further enhancement transferability. This paper proposes a novel adversarial attack method, which implements stacked layer-wise relevance propagation technologies, called ASL. Specifically, the ASL implements stacked attention mechanisms at the pixel level to comprehensively explore feature information. The comprehensive feature information is used as the loss function to generate the adversarial perturbation, which is minimized to create adversarial samples. Through extensive experiments on ImageNet, we demonstrate that ASL improves transferability and generates efficiently on victim models owing to its ability to obtain and modify comprehensive feature information.