DaDDNet: A Dynamic-aware Dual-discriminators Network for Generating Adversarial Patches

Generating adversarial patches aims to deceive the classifier with generated samples in the real world. Although many current methods have made advances, the realistic and attack capability are still challenging issues. We propose a dynamic-aware dual-discriminators Network (DaDDNet) for generating adversarial patches, which is composed of a generator and dual-discriminators
$D_{\alpha}$
and
$D_{\gamma}$
. The dual-discriminators improve the global authenticity and local aggressiveness of the adversarial patches. To alleviate the phenomenon of overfitting, a dynamic-aware strategy is assigned during the training process. Considering the DaDDNet's effectiveness, 10-fold cross-validation experiments are carried out on seven classifiers in four datasets. The results show that our DaDDNet on 23/28 groups of experiments leads to a decreased recall rate, which is better than the conventional GANs. Furthermore, our method accelerates the convergence rate and the training method is more stable by observing the training loss function.