Deep neural networks (DNNs) are susceptible to adversarial examples generally generated by adding imperceptible perturbations to the clean images, resulting in the degraded performance of DNNs models. To generate adversarial examples, most methods utilize the \(L_p\) norm to limit the perturbations and satisfy such imperceptibility. However, the \(L_p\) norm cannot fully guarantee the semantic authenticity of adversarial examples. Defenses may take advantage of this defect to weaken the attack capability of adversarial examples. Moreover, existing methods with \(L_p\) restriction have poor generalization ability in white-box attacks and have inferior aggressiveness in black-box attacks. To solve the problems mentioned above, we propose a multiple feature interpolation method to generate face adversarial examples. In the proposed method, we perform the multiple feature interpolation to generate face adversarial examples with new semantics in the process of original image reconstruction and conditional attribute-guided image generation based on StarGAN. Experimental results demonstrate that adversarial examples generated by our method possess new attribute-guided semantics and satisfactory attack success rates under both white-box and black-box settings.
