Efficient Adversarial Attacks on Online Multi-agent Reinforcement Learning

Due to the broad range of applications of multi-agent reinforcement learning (MARL), understanding the effects of adversarial attacks against MARL model is essential for the safe applications of this model. Motivated by this, we investigate the impact of adversarial attacks on MARL. In the considered setup, there is an exogenous attacker who is able to modify the rewards before the agents receive them or manipulate the actions before the environment receives them. The attacker aims to guide each agent into a target policy or maximize the cumulative rewards under some specific reward function chosen by the attacker, while minimizing the amount of manipulation on feedback and action. We first show the limitations of the action poisoning only attacks and the reward poisoning only attacks. We then introduce a mixed attack strategy with both the action poisoning and the reward poisoning. We show that the mixed attack strategy can efficiently attack MARL agents even if the attacker has no prior information about the underlying environment and the agents' algorithms.

Paper

References (41)

Scroll for more · 29 remaining

Similar papers

Peer review

Reviewer Mdqm6/10 · confidence 3/52023-06-26

Summary

This paper proposes a series of novel attacks against MARL under different assumptions. From the theoretical perspective, it then analyzes the efficacy of the proposed attacks.

Strengths

+ The paper proposes a series of novel attacks against online MARL and provides a decent theoretical analysis for the efficacy of the proposed attack. This paper lays a theoretical foundation for the adversarial attacks against online MARL. The proposed attacks can be potentially extended by follow-up works to launch practical attacks against MARL systems. The novelty of the problem and the technical depth are enough for a top-tier ML conference.

Weaknesses

I do not idenitfy any critical weakness in the proposed technique and the corresponding theoretical analysis. As a theoretical paper, it is a little bit over to ask questions about the practical perspective. However, as an attack paper, it is somewhat important to discuss the practicability of the proposed attacks. As such, I would suggest the authors add such a discussion related to the proposed technique. For example, the authors could provide some suggestions for practitioners to solve the proposed objective functions, which could be hard to optimize. In addition, what would be the computational cost for the proposed objective functions?

Questions

1. Why the proposed assumptions that the attackers can freely alter the target agents' actions and rewards are practical and realistic? 2. What are the practical suggestions for optimizing the proposed objective functions? 3. What is the computational cost of solving the proposed attack objective functions?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This paper does not include a section discussing the limitation of the proposed attacks. Given that the paper is mainly about theoretical analysis of attack efficacy. It could include a discussion section (maybe in the appendix) to discuss the potential limitations maybe on the practical side, i.e., how realistic are the assumptions that the attackers can freely manipulate the action and reward? What is the computational cost of solving the proposed attack objective functions? In addition, the authors checked the broader impact of the checklist. However, I do not find a (sub)section or paragraph clearly addressing the potential negative social impact.

Reviewer qHBd5/10 · confidence 2/52023-07-05

Summary

This paper presents a novel approach to adversarial attacks in Multi-agent Reinforcement Learning (MARL), offering significant insights into various attack settings. The authors explore different attack strategies, including reward poisoning, action poisoning, and a combination of both, across various settings: white-box, grey-box, and black-box. The paper's key contribution is the efficient adversarial attack strategy that perturbs both the action and reward of the agents.

Strengths

- The paper addresses a novel and significant problem of adversarial attacks in Multi-agent Reinforcement Learning (MARL) for reward/action poisoning. - The paper is moderately easy to follow. - The authors have considered different settings such as white-box, grey-box, and black-box, which adds depth to the study.

Weaknesses

- The related work section of the paper does not seem to cover the literature comprehensively. It would be beneficial for the authors to provide a more thorough review of the existing literature, especially focusing on the works that have addressed similar problems in the past, how they did it (for reward/attack poisoning). - The paper lacks numerical results. While the authors have provided some simulation results in the appendix, it would be beneficial to include more numerical results in the main body of the paper to strengthen their arguments. Furthermore, this is a quite practical problem, and it is beneficial to thoroughly assess the concepts introduced by the authors by means of numerical simulations. - Attack detection is not discussed by the authors. -The authors have analyzed the performance of the proposed attack algorithm on V-learning. It would be interesting to see how the proposed black-box strategy performs on other MARL algorithms. - It's not clear how tight the bounds proposed by the authors are.

Questions

See above

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Authors do not seem to discuss limitations thoroughly.

Reviewer YKVp6/10 · confidence 2/52023-07-06

Summary

This paper investigates the impact of adversarial attacks on Multi-Agent Reinforcement Learning (MARL) models. The authors propose an attacker who can manipulate rewards and actions to guide the agents into a target policy or maximize cumulative rewards under a specific reward function. The paper presents an adversarial attack model where the attacker aims to force the agents to learn a target policy or maximize cumulative rewards. Loss and cost functions are used to evaluate the effectiveness of the attack, with cost representing the cumulative manipulation and loss measuring the deviation from the target policy or regret compared to the attacker's optimal policy. The attack problem is studied in three settings: white-box, gray-box, and black-box. The paper demonstrates the limitations of action poisoning-only attacks and reward poisoning-only attacks. Certain Markov Games (MGs) are identified where these strategies are inefficient. However, sufficient conditions are provided under which these attacks can efficiently target MARL algorithms. Efficient strategies for action poisoning and reward poisoning attacks are introduced, and their costs and losses are analyzed. Then a mixed attack strategy is proposed in the gray-box setting, and an approximate mixed attack strategy is introduced for the black-box setting. These strategies can force sub-linear-regret MARL agents to choose actions according to the attacker's target policy with sub-linear cost and loss. The impact of the approximate mixed attack strategy on V-learning, a decentralized MARL algorithm, is investigated. No experiments have been done.

Strengths

Originality: Firstly, this paper introduces an adversarial attack model specifically tailored for Multi-Agent Reinforcement Learning (MARL) systems. While adversarial attacks in single-agent RL have been studied before, this paper focuses on the challenges and implications of attacks in the context of MARL. The consideration of both action poisoning and reward poisoning, as well as the introduction of a mixed attack strategy, adds originality to the research. Quality: The analysis considers different attack settings, providing a comprehensive understanding of the impact of adversarial attacks on MARL algorithms. The paper also provides conditions under which the attack strategies can be effective, contributing to the quality of the research. Clarity: The paper is easy to follow. The introduction provides an overview of the problem and motivation, while the contributions are explicitly listed, making it easy for readers to understand what the paper aims to achieve. The attack model, attack settings, and attack strategies are explained in a concise and understandable manner. Significance: With the increasing use of MARL in various applications, understanding the vulnerabilities and countermeasures against adversarial attacks is crucial for the safe and reliable deployment of these systems. By investigating the limitations of existing attack strategies and proposing new attack strategies, the paper sheds light on the security aspects of MARL and contributes to the development of more robust and trustworthy MARL algorithms.

Weaknesses

Limited comparison with existing work: The paper lacks a comprehensive comparison with prior research on adversarial attacks in MARL. While it briefly mentions existing works on attacks in single-agent RL and MARL, it does not provide a detailed comparison or highlight the novelty and differentiation of its proposed attack model and strategies. A more extensive discussion and comparison with related works would strengthen the paper's contribution and contextualize its findings within the existing literature. No empirical evaluation: The paper would benefit from an empirical evaluation of the proposed attack strategies. While the analysis provides theoretical insights into the effectiveness of the attack strategies, it lacks practical validation through experiments on real-world or simulated MARL environments. Conducting experiments and presenting empirical results would enhance the credibility and applicability of the proposed attack strategies.

Questions

Some experiments would benefit the paper. It would be great if the authors can resolve my concerns in the weakness.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer SPgM6/10 · confidence 4/52023-07-08

Summary

This paper studies adversarial attacks in online multi-agent RL, focusing on reward and action poisoning attacks. It provides a set of characterization results for three different attack modalities: white-box, gray-box and black-box attacks. The authors first discuss the limitation of action-only or reward-only poisoning attack, showing that there are instances of the problem setting where these are not efficient and successful. However, for white-box versions of these attacks, the authors provide sufficient conditions that enable feasibility. For all three attack modalities, they demonstrate that the combination of action and reward poisoning attacks can always force targeted behavior, and provide upper bounds on the expected cost and loss of proposed attack strategies.

Strengths

- The paper is clearly written and is enjoyable to read. While there are typos, these should be easy to fix. More importantly, the technical content is clearly conveyed, and the paper provides intuitions bending the main technical results. - The paper complements prior work on poisoning attacks in MARL, which considered offline RL setting. Hence, the characterization results are novel and contribute to the line of work on poisoning attacks in RL. That said, the results appear to be similar to the ones presented in [Rangi et al., 2022]. It would be useful to have additional discussions on how these differ in terms of their technical content, proof techniques, etc.. - The paper considers several variations of poisoning attacks under different attack modalities, while showcasing the existence of efficient and successful reward+action poisoning attack strategies. Such a systematic study is likely to be valuable to researchers working on poisoning attacks in RL. - The paper provides a good overview of the related work on adversarial attacks in RL. Specifically for MARL, the most important references seem to be covered. However, there are a couple of recent references that could be included in the list (see below). I encourage the authors to do so. > - Wang et al;, BACKDOORL: Backdoor Attack against Competitive Reinforcement Learning > - Mohammadi et al., Implicit Poisoning Attacks in Two-Agent Reinforcement Learning: Adversarial Policies for Training-Time Attacks > - Liu et al., Controllable Attack and Improved Adversarial Training in Multi-Agent Reinforcement Learning > - Chen et al., Byzantine-Robust Online and Offline Distributed Reinforcement Learning

Weaknesses

- In terms of results, the paper does not discus the optimality of the upper bounds in their formal results. For example, it could be interesting to discuss the tightness of the bounds in Theorem 9, and how they compare to those of the underlying learning algorithms (V-learning). - Following up on my previous remark, additional discussions regarding some of the assumptions in the problems setting would be useful to have. For example, the paper assumes that the target policy receives strictly positive rewards (page 4). Additionally, it would be good to formally specify the learners' model, and reference it in Theorem 1 and 2. - This work is primarily theoretical, but the paper could benefit from having additional experiments (in addition to the numerical simulations reported in the appendix), e.g., those that would demonstrate the efficacy of the black-box approach. Namely, bound in (8) and (9) are inversely proportional to $R_{min}$, so it would be interesting to investigate how this dependency affects the practicality of the proposed approach. - Practical considerations are not discussed, e.g., it would be useful to have some discussion on how to enable scalability and go beyond the tabular setting.

Questions

It would be great if the authors could additionally answer the following clarification questions: - Could you explain how these results compare to the ones from [Rangi et al., 2022] in terms of, e.g., proof techniques? - For Eq. (1), Theorem 1 and 2, what is the assumed model of the MARL agents? I.e., what is needed for these statements to hold? - Can you explain why the analysis does not allow sparse rewards for the target policy (the assumption on page 4)? What happens if the rewards are equal to $0$? - In the section that talks about reward-only white-box attacks, why doesn't it suffice to set the reward to the lowest value whenever the target action is not taken? - The dependency on $m$ seems to be smaller in gray-box than in white-box attacks. Could you comment on the practical benefits of white-box attacks? - Could you comment the optimality of the upper bounds (attack loss/cost)? More specifically, could you comment on the tightness of the bounds in Theorem 9, e.g., in terms of horizon $H$?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

This paper provides a theoretical analysis of poisoning attacks in multi-agent RL, so I don't believe this work will have any negative societal impact. The limitations of this work could have been discussed in greater detail, e.g., in the concluding remarks. For example, instead of summarize the results of the paper, the concluding remarks could focus more on the tightness of the bounds in Theorem 9, or the practical aspects of this work, e.g., how to go beyond the tabular setting.

Reviewer YKVp2023-08-15

Thank you for your response

I would thank the authors for their detailed responses to my questions. I have no concerns now and would like to raise my rating from 5 to 6. Though I got to know the experiments are in the appendix, I still suggest putting some experiment results in the main text when preparing the camera-ready version.

Authorsrebuttal2023-08-17

Thanks!

We thank the reviewer again for the thoughtful comments and valuable recommendations. We will definitely add some experiment results in the main text of the revised version.

Reviewer Mdqm2023-08-15

Thank the authors for the clarification. I do not have further questions. Please add the rebuttal changes to the paper.

Authorsrebuttal2023-08-17

Thanks!

We thank the reviewer again for the valuable feedback and the time invested. We will definitely add the rebuttal changes to the paper.

Reviewer SPgM2023-08-20

Thank you for your response

Thank you for your detailed response and for answering my questions. It would be great if you could update the paper accordingly. I don't have further clarification questions.

Reviewer qHBd2023-08-20

I thank the authors for their comments. However, I still have concerns regarding the problem of detection and tightness of the bounds. Furthermore, I understand that the numerical results are in the appendix, but given the nature of the topic, I believe it is better to move the results in the main body of the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC