基于空频双域特征融合的高迁移性对抗样本生成方法OA

Despite the remarkable performance of deep neural networks across various fields,the existence of adver-sarial examples reveals significant security vulnerabilities.Existing black-box attack methods typically operate within a sin-gle domain,overlooking the importance of multi-domain feature co-perturbation in enhancing the transferability of adversar-ial examples.Moreover,many methods suffer from a single-purpose loss function,making it difficult to balance target class guidance and gradient stability.To address these issues,this paper proposes a high-transferability adversarial examples gen-eration method based on spatial-frequency dual-domain feature fusion(SFDFF).Specifically,the input examples are first transformed from the spatial domain to the frequency domain using the discrete cosine transform,and region-level feature fusion is performed between the input and clean examples in the frequency domain.Then,the input examples are restored to the spatial domain via the inverse discrete cosine transform,and noise based on the statistical characteristics of the original examples are injected.Next,channel-level fusion of spatial features between the input and clean examples are conducted.Fi-nally,a dual-guidance loss function is designed to simultaneously enhance target class directionality and gradient stability.Extensive experiments on ImageNet-Compatible and CIFAR-10 datasets demonstrate the performance of the proposed method.For instance,the attack success rate of the proposed SFDFF increases by 2.5%compared to the state-of-the-art method when transferred from the adv-RN-50 to LeViT model on ImageNet-Compatible dataset.The code is available at https://github.com/ipkpkpk/SFDFF.