Improve Robustness of Safe Reinforcement Learning Against Adversarial Attacks
Authors:
- 27 September 2024
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 723-727
Abstract
Models trained with adversarial attack can be significantly improved stability and performance when faced with new uncertain environment. In this paper, we propose the robust training framework based on Wasserstein SA-MDP for safe RL, which combines ideas from reinforcement learning, adversarial examples generation, Wasserstein metric and robust training. This framework can improve the robustness of safe RL models against adversarial attacks by considering the dynamics of the adversarial interaction as a sequential decision-making process. We made the experiment in safety gym based on the proposed framework and report the experimental results.Keywords
Funding Information
- Natural Science Foundation of Jilin Province (20200201290JC)
This publication has 8 references indexed in Scilit:
- Robust Safe Reinforcement Learning under Adversarial Disturbances2023
- Model-Free Safe Reinforcement Learning Through Neural Barrier CertificateIEEE Robotics and Automation Letters, 2023
- Reinforcement Learning based Recommender Systems: A SurveyACM Computing Surveys, 2022
- Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement LearningAnnual Review of Control, Robotics, and Autonomous Systems, 2022
- Deep Reinforcement Learning for Autonomous Driving: A SurveyIEEE Transactions on Intelligent Transportation Systems, 2021
- Robustifying Reinforcement Learning Agents via Action Space Adversarial Training2020
- Towards Deep Learning Models Resistant to Adversarial AttacksarXiv, 2019
- Human-level control through deep reinforcement learningNature, 2015