Adv-Makeup++: an adversarial makeup generation method for simultaneous optimization of disturbance distribution and improvement of attack success rate

To address the issues of insufficient attack capability against black-box face recognition models and poor visual quality of face adversarial samples, we propose the Adv-Makeup++ model. First, we introduce the CBAM attention mechanism module in the encoder. Second, we propose a discriminator momentum enhanced adversarial training strategy, which strengthens the discriminator's robustness to perturbations, thus forcing the generator to produce more powerful adversarial samples. Finally, to overcome the problem of unbalanced adversarial efficiency between the generator and discriminator during GAN training, we introduce a multi-scale feature matching technique to ensure the generator can accurately simulate the feature distribution of real face samples at different scales, thereby achieving a more stable GAN training process. Experiments show that compared to Adv-Makeup, our method significantly improves the attack success rate on the Makeup dataset, while also achieving better visual quality metrics.