RL-Based Method for Benchmarking the Adversarial Resilience and Robustness of Deep Reinforcement Learning Policies

This paper investigates the resilience and robustness of Deep Reinforcement Learning (DRL) policies to adversarial perturbations in the state space. We first present an approach for the disentanglement of vulnerabilities caused by representation learning of DRL agents from those that stem from the sensitivity of the DRL policies to distributional shifts in state transitions. Building on this approach, we propose two RL-based techniques for quantitative benchmarking of adversarial resilience and robustness in DRL policies against perturbations of state transitions. We demonstrate the feasibility of our proposals through experimental evaluation of resilience and robustness in DQN, A2C, and PPO2 policies trained in the Cartpole environment.

cs.LG cs.AI cs.CR eess.SY stat.ML Computer Science Engineering Mathematics

Source

Paper

References (13)

01The Faults in Our Pi Stars: Security Issues and Open Challenges in Deep Reinforcement LearningVahid Behzadan, Arslan Munir2018 · arXiv.org · 27 citations In Library

02Ensemble Adversarial Training: Attacks and DefensesFlorian Tramèr, Alexey Kurakin, Nicolas Papernot et al.2017 · International Conference on Learning Representations · 3.1k citations In Library

03Imitation LearningA. Hussein, M. Gaber, E. Elyan et al.2017 · ACM Computing Surveys · 559 citations

04Tactics of Adversarial Attack on Deep Reinforcement Learning AgentsYen-Chen Lin, Zhang-Wei Hong, Yuan-Hong Liao et al.2017 · International Joint Conference on Artificial Intelligence · 475 citations In Library

05Adversarial Attacks on Neural Network PoliciesSandy H. Huang, Nicolas Papernot, I. Goodfellow et al.2017 · International Conference on Learning Representations · 957 citations In Library

06Vulnerability of Deep Reinforcement Learning to Policy Induction AttacksVahid Behzadan, Arslan Munir2017 · IAPR International Conference on Machine Learning and Data Mining in Pattern Recognition · 311 citations In Library

07OpenAI GymGreg Brockman, Vicki Cheung, Ludwig Pettersson et al.2016 · arXiv.org · 5.6k citations In Library

08Explaining and Harnessing Adversarial ExamplesI. Goodfellow, Jonathon Shlens, Christian Szegedy2014 · International Conference on Learning Representations · 23k citations In Library

09We propose two RL-based techniques and corresponding metrics for the measurement and benchmarking of resilience and robustness of DRL policies to perturbations of state transitions

10Train the adversarial agent against the target policy π ∗ in its training environment, report the maximum adversarial regret R ∗ adv ( T ) for time T achieved atadversarial optimality

11Apply the adversarial policy against the target in N episodes, record total cost C adv for each episode

12Train the adversarial agent against the target following π in itstraining environment, report the optimal adversarial return R ∗ perturbed and the maximum adversarial regret R ∗ adv ( T )

Scroll for more · 1 remaining