Probabilistic-Robustness-Driven Explainable Universal Adversarial Perturbations Attack Deep Reinforcement Learning Intrusion Detection Systems
Probabilistic Robustness-driven Universal Adversarial Perturbations with Explainability against Deep Reinforcement Learning-based Intrusion Detection System
AI summary
A method for generating universal adversarial perturbations (UAPs) targeting probabilistic robustness (PR) is proposed for attacking deep reinforcement learning (DRL) intrusion detection systems (IDS). The work claims to be the first to introduce an explicit PR objective into such UAPs, and proposes PX-UAP, which uses explainable AI to shape perturbations under realistic domain constraints, along with a theoretical analysis. Experiments show that PX-UAP consistently outperforms existing UAP methods in attack effectiveness.