Similarity-guided Model-selection of Adversarial attack Methods Based on Transferability

Black-box attacks based on adversarial instance transferability have become a research hotspot. This method uses known white box models (alternative models) to create adversarial examples that pose a threat to other unknown black box models. However, current black box attack methods based on transferability face many challenges, such as high query counts and difficulty balancing transferability and attack success rates. To address these challenges, this article introduces a similarity guided model selection adversarial attack method (SMAT) based on transferability. It combines transfer-based techniques with score-based attacks to analyze the similarity between the target model and candidate alternative models, and selects the model that is closer to the target model on the decision boundary from the candidate alternative model set as the alternative model.. Experiments on MNIST, CIFAR-10, and ImageNet have shown that LBAT reduces queries while maintaining transferability and success. This study enhances the theoretical framework of black box attacks, which contributes to the robustness evaluation of DNNs and advances adversarial machine learning.