Revisiting Adversarial Attacks on Deep Neural Networks via Correlating Adaptive Gradient-based Features

Existing works proved that Deep Neural Networks (DNNs) are vulnerable to artificial adversarial features, which pose challenges in the application of DNNs in security-sensitive scenarios. Exploring adversarial strategies benefits developing identical adversarial defenses that robust DNNs in security-sensitive challenges. From an adversary perspective, finding the adversarial properties with considerable adversarial capability is an optimization problem in feature generation. However, existing attacks focused on improving the adversarial performance of crafted features while ignoring the optimization of perturbation strength, which induces noticeable modifications in adversarial features. To generate adversarial features with profound capability in invading deep neural task models while preserving balanced perturbation magnitude, we introduce a simple yet flexible manner: Adaptive Gradient-based Feature Correlated Method (AFCM) that calls a few adaptive operators on original distributions and correlates such crafted gradient-based features in adaptive similarity-based constraints, which benefits from escaping from the local optimum while searching decision boundaries in adversarial processing. Extensive experiments reveal that adversarial features generated by AFCM acquire surpassing effectiveness in fooling deep neural networks while preserving fewer pixel changes on $\ell_{\infty}$ and $\ell_{2}$ norm constraints with perfect generalization and transferability.