Abstract:Deep neural networks are vulnerable to the attack of adversarial examples, which apply perturbations that are hard to notice to mislead models into wrong judgments. While adversarial examples achieve exceptionally high attack rates in white-box settings, their transferability drops significantly in black-box environments. To address this, this study proposes a novel feature-frequency domain dual-discrimination method for enhancing the transferability of adversarial examples. Specifically, the feature discriminator separates semantics between original and adversarial examples in a high-dimensional feature space. The frequency domain discriminator thoroughly analyzes adversarial changes through high-frequency domain features, enabling it to distinguish between original images and adversarial examples. During training, both discriminators guide the generator to optimize adversarial quality in a coordinating manner, while the generator applies random pixel discard to perturbations, thereby preventing perturbation over-concentration in a specific image region. Experiments show that, in the black-box model, the proposed method outperforms VMI-FGSM, TTP, and MultiANDA in transferability, demonstrating effectiveness.