Summary of A Spatiotemporal Stealthy Backdoor Attack Against Cooperative Multi-agent Deep Reinforcement Learning, by Yinbo Yu et al.
A Spatiotemporal Stealthy Backdoor Attack against Cooperative Multi-Agent Deep Reinforcement Learning
by Yinbo Yu, Saihao Yan, Jiajia Liu
First submitted to arxiv on: 12 Sep 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Cryptography and Security (cs.CR)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Recent studies have highlighted the vulnerability of cooperative multi-agent deep reinforcement learning (c-MADRL) systems to backdoor attacks. Existing proposed backdoors suffer from limitations such as fixed visual trigger patterns, additional network requirements, or all-agents being compromised. In response, this paper introduces a novel backdoor attack that targets an entire c-MADRL team by embedding the backdoor in a single agent. The proposed method employs adversary spatiotemporal behavior patterns as triggers and controls attack duration for stealthiness and practicality. Additionally, the reward function of the backdoored agent is modified via reward reverse and unilateral guidance during training to ensure its adverse influence on the entire team. The proposed attacks are evaluated on two classic c-MADRL algorithms (VDN and QMIX) in the popular SMAC environment. Experimental results demonstrate a high attack success rate (91.6%) with low clean performance variance (3.7%). This study contributes to the development of effective backdoor attacks against cooperative multi-agent deep reinforcement learning systems. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Researchers have found that certain types of team-based artificial intelligence can be tricked into doing bad things. This is called a “backdoor” attack. The problem is that existing ways to create these backdoors are not very good, such as using fixed patterns or requiring extra networks. To address this issue, scientists have come up with a new way to make backdoors that can affect an entire team of AI agents by hiding the bad instructions in just one agent. This new method uses unusual behavior patterns and allows the attack to last for a certain amount of time. The study also changes how the attacked agent gets its rewards to ensure it does what’s being told. Researchers tested this new backdoor attack on two types of team-based AI systems and found that it was very effective, with 91.6% success rate. |
Keywords
» Artificial intelligence » Embedding » Reinforcement learning » Spatiotemporal
