Summary of Cooperative Backdoor Attack in Decentralized Reinforcement Learning with Theoretical Guarantee, by Mengtong Gao et al.
Cooperative Backdoor Attack in Decentralized Reinforcement Learning with Theoretical Guarantee
by Mengtong Gao, Yifei Zou, Zuyuan Zhang, Xiuzhen Cheng, Dongxiao Yu
First submitted to arxiv on: 24 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
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 The paper explores the safety concerns in decentralized reinforcement learning (RL) by introducing a novel cooperative backdoor attack that compromises the policies of benign agents. Unlike existing methods that conceal the attack within shared policies, this approach breaks down the backdoor behavior into multiple components based on the RL state space. Each malicious agent hides one component in its policy and shares it with others, allowing the backdoor to be assembled when a benign agent learns all poisoned policies. Theoretical proofs demonstrate the effectiveness of this cooperative method, which is more covert than existing attacks since each attacker’s policy contains only a portion of the backdoor. Experimental simulations on Atari environments validate the efficiency and stealthiness of the approach. This research presents the first provable cooperative backdoor attack in decentralized RL. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine a situation where some bad actors try to sneakily change the behavior of good agents in a complex game-like scenario called reinforcement learning. They do this by sharing small parts of their own “strategies” that, when combined, create a hidden problem. The researchers looked at how these bad actors might work together to cause trouble and found that they can be very effective. They also tested this idea using computer simulations and showed that it’s hard to detect. This is the first time someone has demonstrated this kind of attack in a specific type of game-like scenario. |
Keywords
* Artificial intelligence * Reinforcement learning
