An Adversarial Example Generation Method Based on Backward Dynamic Slope Gating

ABSTRACT Network intrusion detection is an important technical means to ensure cyberspace security. Many intrusion detection systems adopt artificial intelligence methods, and the Rectified Linear Unit (ReLU) activation function is widely used in deep learning‐based intrusion detection models. However, in the process of generating adversarial examples, these models suffer from problems such as gradient sparsity and unstable optimization direction caused by the dead zone of ReLU and gradient degradation. To address these issues, this paper proposes an adversarial example generation scheme based on backward dynamic slope gating, named DyReLU‐Atk. DyReLU‐Atk assigns adaptive nonzero slopes to candidate positions in the negative activation region of standard ReLU, where the corresponding gradients are originally zero, through a continuous soft gating mechanism. This allows partially suppressed gradients that are beneficial to the targeted attack objective to contribute to backward optimization, thereby enabling attack‐oriented gradient reshaping. The resulting dynamic gradient is then integrated with the original gradient through a lightweight stabilization strategy to generate feasible adversarial examples under feature constraints. The proposed mechanism is embedded into seven representative attack methods, including FGSM, BIM, PGD, CW‐, FAB‐, APGD, and CAPGD, and evaluated on binary intrusion detection tasks using the NSL‐KDD and UNSW‐NB15 datasets. Experimental results show that DyReLU‐Atk generally achieves higher targeted attack success rate (ASR) and enhances attack effectiveness and stability under limited perturbation budgets in the feature‐constrained white‐box attack scenarios investigated in this work.