基于ViT的高迁移性黑盒对抗样本生成OA

Transfer-based adversarial attacks against convolutional neural networks(CNNs)exploit the observation that models trained on the same task often share similar decision boundaries,crafting examples on a substitute model that transfer to a target.However,due to architectural differences,existing methods struggle to achieve cross-architecture transfer.To address this,we propose the OptiEncode,a Vision Transformer(ViT)‒based method for generating high-transferability black-box adversarial examples.Leverag-ing ViT's strong capability to localize salient features,OptiEncode uses Grad-CAM to identify cross-architecture consensus regions and Sobel to extract architecture-agnostic high-frequency structures,and then injects perturbations only on their intersection to enhance transferability.Unlike SE,PNA,and FPR,which primarily modify the model internals,the OptiEncode explicitly aligns cross-architecture shared features in the input space,thus making it orthogonal and stackable with those approaches.Evaluations on black-box targets across diverse architectures(ViT,CNN,and MLP)show that the OptiEncode improves ViT-to-ViT transfer by about 10%on average and achieves up to 15%gains in cross-architecture settings(e.g.,ViT→CNN).These results indicate that explicitly aligning the"consensus regions ∩ high-frequency structures"in the input space effectively narrows the cross-architecture transfer gap and offers a reusable,complementary path for improving the practicality of black-box transfer attacks.