Multi-Environment Training Against Reward Poisoning Attacks on Deep Reinforcement Learning

Myria Bouhaddi, Kamel Adi

2023

Abstract

Our research tackles the critical challenge of defending against poisoning attacks in deep reinforcement learning, which have significant cybersecurity implications. These attacks involve subtle manipulation of rewards, leading the attacker’s policy to appear optimal under the poisoned rewards, thus compromising the integrity and reliability of such systems. Our goal is to develop robust agents resistant to manipulations. We propose an optimization framework with a multi-environment setting, which enhances resilience and generalization. By exposing agents to diverse environments, we mitigate the impact of poisoning attacks. Additionally, we employ a variance-based method to detect reward manipulation effectively. Leveraging this information, our optimization framework derives a defense policy that fortifies agents against attacks, bolstering their resistance to reward manipulation.

Paper Citation

in Harvard Style

Bouhaddi M. and Adi K. (2023). Multi-Environment Training Against Reward Poisoning Attacks on Deep Reinforcement Learning. In Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT; ISBN 978-989-758-666-8, SciTePress, pages 870-875. DOI: 10.5220/0012139900003555

in Bibtex Style

@conference{secrypt23,
author={Myria Bouhaddi and Kamel Adi},
title={Multi-Environment Training Against Reward Poisoning Attacks on Deep Reinforcement Learning},
booktitle={Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT},
year={2023},
pages={870-875},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012139900003555},
isbn={978-989-758-666-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 20th International Conference on Security and Cryptography - Volume 1: SECRYPT
TI - Multi-Environment Training Against Reward Poisoning Attacks on Deep Reinforcement Learning
SN - 978-989-758-666-8
AU - Bouhaddi M.
AU - Adi K.
PY - 2023
SP - 870
EP - 875
DO - 10.5220/0012139900003555
PB - SciTePress