Certifying Safety in Reinforcement Learning under Adversarial Perturbation Attacks

• Published in: Proceedings (IEEE Security and Privacy Workshops. Online) pp. 57 - 67
• Main Authors: Wu, Junlin, Sibai, Hussein, Vorobeychik, Yevgeniy
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.05.2024
• Subjects: Adversarial Perturbation; Markov decision processes; Perturbation methods; Prediction algorithms; Privacy; Reinforcement learning; Robustness; Safe Reinforcement Learning; Safety; Verified Safety

Function approximation has enabled remarkable advances in applying reinforcement learning (RL) techniques in environments with high-dimensional inputs, such as images, in an end-to-end fashion, mapping such inputs directly to low-level control. Nevertheless, these have proved vulnerable to small adv...