Simin Li (李思民)

I am a Ph.D. student (2021.09-) at the State Key Laboratory of Software Development Environment (SKLSDE) and School of Computer Science and Engineering (SCSE) , Beihang University, Beijing, China, supervised by Prof. Xianglong Liu. Before that, I obtained my BSc degree in 2020 from Beihang University (Summa Cum Laude).

Email: lisiminsimon@buaa.edu.cn

Google Scholar / CV

I am interested in Trustworthy AI for multi-agent reinforcement learning (MARL). My research goal is to make reinforcement learning safe and robust, including practical adversarial attack for RL/MARL, adversarial defense. Many of my works on robust MARL are tested on real world robot swarms.

• Robust MARL
• Trustworthy RLHF

[2024.02] One co-authored paper on collision avoidance submitted to RAL 2024

[2024.02] One first-authored paper on robust regularization for MARL submitted to ICML 2024

[2024.01] One first-authored paper on defending Byzantine adversary of MARL accepted by ICLR 2024

[2024.01] Two co-authored paper on robotics submitted to IJCAI 2024

[2024.01] One co-authored paper on partial symmetry for MARL accepted by AAAI 2024

[2023.09] One first-authored paper on adversarial attack against MARL submitted to IEEE TCYB

[2022.11] One first-authored paper on naturalness of physical world adversarial attack accepted by CVPR 2023

[2022.12] One co-authored paper of algorithmic testing accepted by Artificial Intelligence Security (In Chinese).

[2022.12] One co-authored survey on RL robustness accepted by Chinese Journal of Computers (In Chinese, top journal in China, CCF-A).

[2022.07] One first-authored paper on privacy protection of fingerprints submitted to IEEE TIP (currently under minor revision).

[2022.04] One co-authored paper on robustness testing of MARL accepted by CVPR 2022 workshop.

Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game
Simin Li, Jun Guo, Jingqiao Xiu, Xin Yu, Jiakai Wang, Aishan Liu, Yaodong Yang, Xianglong Liu.

Accepted by ICLR, 2024
pdf / Project page

We study robustness of MARL against Byzantine action perturbations by formulating it as a Bayesian game. We provide a rigorious formulation of this problem and an algorithm with strong empirical performance.

Mutual Information Regularization is Provably Robust for Multi-Agent Reinforcement Learning
Simin Li, Ruixiao Xu, Jun Guo, Pu Feng, Jiakai Wang, Aishan Liu, Yaodong Yang, Xianglong Liu.

Submitted to AAAI, 2024
pdf

We proof that minimizing mutual information as a regularization term is minimizing a lower bound of robustness in MARL under all potential threat scenarios.

Attacking Cooperative Multi-Agent Reinforcement Learning by Adversarial Minority Influence
Simin Li, Jun Guo, Jingqiao Xiu, Pu Feng, Xin Yu, Jiakai Wang, Aishan Liu, Wenjun Wu, Xianglong Liu.

Submitted to IEEE TCYB (IF=11.8)
pdf / Project page

We propose the first adversarial policy attack for c-MARL, which is strong and practical. Our attack provides the first demonstration that adversarial policy is effective against real world robot swarms.

We take the first step to evaluate the naturalness of physical world adversarial examples by a human oriented approach. We collect the first dataset with human naturalness ratings and human gaze, unveil insights of how contextual and behavioral features will affect attack naturalness, and propose an algorithm to automatically evaluate naturalness by aligning human behavior and algorithm prediction.

Hierarchical Perceptual Noise Injection for Social Media Fingerprint Privacy Protection
Simin Li, Huangxinxin Xu, Jiakai Wang, Aishan Liu, Fazhi He, Xianglong Liu, Dacheng Tao.

Submitted to IEEE TIP(IF=10.86), currently under minor revision
pdf / Project page

While billions of people are sharing their daily life images on social media everyday, hackers can easily steal fingerprint from the shared images. We leverage adversarial attack to protect such privacy leakage, such that hackers cannot extract fingerprints even they get the shared images in social media. Our method, FingerSafe, is strong for protection and natural for daily use.

Towards Comprehensive Testing on the Robustness of Cooperative Multi-agent Reinforcement Learning
Jun Guo, Yonghong Chen, Yihang Hao, Zixin Yin, Yin Yu, Simin Li (corresponding author).
Accepted by CVPR Workshop, 2022
pdf

We propose a testing framework to evaluate the robustness of multi-agent reinforcement learning (MARL) algoritms from the aspect of observation, action and reward. Our work first point out state-of-the-art MARL algorithms, including QMIX and MAPPO, are non-robust in multiple aspects, and point out the urgent need to test and enhance the robustness of MARL algorithms.

Leveraging Partial Symmetry for Multi-Agent Reinforcement Learning
Xin Yu, Rongye Shi, Pu Feng, Yongkai Tian, Simin Li, Shuhao Liao, Wenjun Wu.

Accepted by AAAI, 2024

Symmetry has been used in MARL as a prior to incorporate domain knowledge in the environment, which enhance sample efficiency and performance. In this paper, we extend symmetry to paritial symmetry that considers uncertainties in environment with non-uniform field, including uneven terrain, wind, etc.

Exploiting Spatio-Temporal Symmetry for Multi-Agent Reinforcement Learning
Xin Yu, Rongye Shi, Yongkai Tian, Li Wang, Tianhao Peng, Simin Li, Pu Feng, Wenjun Wu.

Submitted to IJCAI, 2024

Symmetry are everywhere in real world, yet current MARL algorithms are agnostic of such symmetry by design. We extend the idea of symmetry to temporal domain, proposing spatial-temporal symmetry network, which includes adds a stronger induction bias during network training.

Lyapunov-Informed Multi-Agent Reinforcement Learning
Pu Feng, Rongye Shi, Size Wang, Xin Yu, Junkang Liang, Jiakai Wang, Simin Li, Wenjun Wu.

Submitted to IJCAI, 2024

Many MARL tasks specify certain goal states where special rewards are granted. The optimal policy in such task could be characterized by Lyapnov stability, where the policy asymptotically converge to the goal states from any initial, making the goal states stable equilibria. We formulate such process as a Lyapunov Markov game, and proof it facilitates the training process to find a stable suboptimal policy more easily and then converge to an optimal policy more efficiently.

Motivated by soft potential field theory, we propose an algorithm to avoid collision in robot swarms.

We provide a comprehensive survey of attack and defenses for deep reinforcement learning. We first analyze adversarial attacks from the perspectives of state-based, reward-based and action based attacks. Then, we illustrate adversarial defenses from adversarial training, adversarial detection, certified robustness and robust learning. Finally, we investigate interesting topics including adversaries for good and model robustness understanding for DRL, and highlights open issues and future challenges in this field.

We provide an AirSim-based unmanned aerial vehicle (UAV) simulator. Based on this simulator, we identify several critical adversarial attacks in multi-UAV combat.

We advocate the research on behavioral dynamics, which provides both microscopic and macroscopic understanding on adversarial vulnerability. We argue that combining the search of network science and game theory with AI safety could potentially benefit the understanding on micro information transmission and macro agent-wise intereaction.

We propose our recent insight to test the security of an intelligent system from full life cycles, including vulnerabilities in model training, testing and deployment and their testing techniques. We offer insights on safety standards, safety testing platforms and sketch our method on security evaluation of autonomous driving.

[Workshop@CVPR2023]I serve as Program Commitee at workshop The Art of Robustness: Devil and Angel in Adversarial Machine Learning at CVPR 2023.

[Reviewer]I am a reviewer of CVPR, ECCV, AAAI, Pattern Recognition, etc.