Abstract
Deep Reinforcement Learning models inherit not only generalization abilities but also vulnerabilities under adversarial attacks from Deep Neural Networks. The recent external model based defense method for Reinforcement Learning (RL) detects and corrects the action relying on only the observation prediction method. The observation prediction method may not perform well in complicated applications because of the knowledge of environment, which will downgrade the defense efficacy. This study proposes a multiple-model based defense method for RL which considers detection and correction tasks separately. Since the problem is broken down into two tasks, their complexity and difficulty is also lower, i.e., a better performance is expected. We propose a Correlation Feature Map to extract the observation consistency in the temporal sequence which is destroyed by adversarial noise to separate clean and attacked states. Our correction only deal with the states classified as contaminated and maps them to proper actions. The performance of our proposed method is evaluated and compared to the state of the art method experimentally in various settings. The results confirm the superiority of our methods in terms of robustness and time.
Supported by the Natural Science Foundation of Guangdong Province, China (No. 2018A030313203).
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
References
Behzadan, V., Munir, A.: Mitigation of policy manipulation attacks on deep q-networks with parameter-space noise. In: Computer Safety, Reliability, and Security - SAFECOMP 2018 Workshops, vol. 11094, pp. 406–417 (2018)
Carlini, N., Wagner, D.A.: Towards evaluating the robustness of neural networks (2017)
Chan, P.P., Wang, Y., Yeung, D.S.: Adversarial attack against deep reinforcement learning with static reward impact map. In: Proceedings of the 15th ACM Asia Conference on Computer and Communications Security, pp. 334–343 (2020)
Gallego, V., Naveiro, R., Insua, D.R.: Reinforcement learning under threats. In: Proceedings of the 33rd AAAI Conference on Artificial Intelligence, pp. 9939–9940 (2019)
Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Proceedings of the 3rd International Conference on Learning Representations (2015)
Havens, A.J., Jiang, Z., Sarkar, S.: Online robust policy learning in the presence of unknown adversaries. In: Proceedings of the 2018 Annual Conference on Neural Information Processing Systems, pp. 9938–9948 (2018)
Huang, S., Papernot, N., Goodfellow, I., Duan, Y., Abbeel, P.: Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284 (2017)
Kos, J., Song, D.: Delving into adversarial attacks on deep policies. In: Proceedings of the 5th International Conference on Learning Representations (2017)
Lin, Y.C., Hong, Z.W., Liao, Y.H., Shih, M.L., Liu, M.Y., Sun, M.: Tactics of adversarial attack on deep reinforcement learning agents. arXiv preprint arXiv:1703.06748 (2017)
Lin, Y.C., Liu, M.Y., Sun, M., Huang, J.B.: Detecting adversarial attacks on neural network policies with visual foresight. arXiv preprint arXiv:1710.00814 (2017)
Mnih, V., et al.: Asynchronous methods for deep reinforcement learning. In: International Conference on Machine Learning, pp. 1928–1937 (2016)
Mnih, V., et al.: Human-level control through deep reinforcement learning. Nature 518(7540), 529–533 (2015)
Moosavi-Dezfooli, S., Fawzi, A., Frossard, P.: DeepFool: a simple and accurate method to fool deep neural networks. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2574–2582 (2016)
Pattanaik, A., Tang, Z., Liu, S., Bommannan, G., Chowdhary, G.: Robust deep reinforcement learning with adversarial attacks. In: Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 2040–2042 (2018)
Silver, D., et al.: Mastering the game of go with deep neural networks and tree search. Nature 529(7587), 484–489 (2016)
Szegedy, C., et al.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2021 Springer Nature Switzerland AG
About this paper
Cite this paper
Chan, P.P.K., Wang, Y., Kees, N., Yeung, D.S. (2021). Multiple-Model Based Defense for Deep Reinforcement Learning Against Adversarial Attack. In: Farkaš, I., Masulli, P., Otte, S., Wermter, S. (eds) Artificial Neural Networks and Machine Learning – ICANN 2021. ICANN 2021. Lecture Notes in Computer Science(), vol 12891. Springer, Cham. https://doi.org/10.1007/978-3-030-86362-3_4
Download citation
DOI: https://doi.org/10.1007/978-3-030-86362-3_4
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-86361-6
Online ISBN: 978-3-030-86362-3
eBook Packages: Computer ScienceComputer Science (R0)