Communicative Power Regularization: Making Multi-Agent Teams Robust to Misleading Communication

Overview

Multi-agent systems that rely on communication are vulnerable to sabotage from adversarial or misaligned agents sending deceptive messages. We introduce Communicative Power Regularization (CPR), a new training method that makes agents more robust by penalizing them for having too much influence over their teammates’ decisions via communication. Our experiments show that CPR drastically improves a team’s performance and resilience in the face of adversarial attacks, without sacrificing the benefits of communication in normal cooperative settings.

Link to the Paper: Arxiv, CoCoMARL, ECAI

Abstract

Effective communication in Multi-Agent Reinforcement Learning (MARL) can significantly enhance coordination and collaborative performance in complex and partially observable environments. However, reliance on communication can also introduce vulnerabilities when agents are misaligned, potentially leading to adversarial interactions that exploit implicit assumptions of cooperative intent. Prior work has addressed adversarial behavior through power regularization by controlling the influence one agent exerts over another, but has largely overlooked the role of communication in these dynamics. This paper introduces Communicative Power Regularization (CPR), extending power regularization specifically to communication channels. By explicitly quantifying and constraining agents’ communicative influence during training, CPR actively mitigates vulnerabilities arising from misaligned or adversarial communications. Evaluations across benchmark environments Red-Door-Blue-Door, Predator-Prey, and Grid Coverage demonstrate that our approach significantly enhances robustness to adversarial communication while preserving cooperative performance, offering a practical framework for secure and resilient cooperative MARL systems.

The Problem: Communication is a Double-Edged Sword

In multi-agent reinforcement learning (MARL), communication allows agents to coordinate effectively and achieve complex team goals, especially when they can only see a part of the environment. Most models are trained with the assumption that all agents are cooperative and messages are trustworthy.

However, this trust is a critical vulnerability. In real-world scenarios, some agents might be adversarial, hacked, or simply faulty, leading them to send misleading information. An agent that naively trusts these messages can be manipulated, causing the entire team’s performance to collapse. This paper addresses the challenge of building MARL systems that are resilient to such “misaligned communication.”

Our Solution: Communicative Power Regularization (CPR)

Our solution is Communicative Power Regularization (CPR), a method that teaches agents to be robustly skeptical. CPR works by modifying the training objective. In addition to their primary goal (e.g., covering a grid, capturing prey), agents are penalized based on how much “power” other agents’ messages exert over their actions.

“Power” here is the ability to change a teammate’s behavior and expected reward. By penalizing this, CPR discourages agents from becoming overly dependent on communication. Instead, they learn more autonomous policies, treating messages as helpful advice rather than unbreakable commands. This allows them to maintain high performance when communication is reliable, while also being able to ignore it and act on their own when it seems misleading.

Key Results

We evaluated CPR across three distinct and challenging environments:

Grid Coverage: In this task, CPR-trained agents proved far more effective at coordinating to cover a grid while under communication attack. In a scenario with 3 cooperative and 3 adversarial agents, CPR improved the team’s score by a remarkable 463% compared to the baseline.

Predator-Prey: We tested a baseline model and a CPR model in a competitive setting. The cooperatively trained baseline, when faced with misleading communication, completely failed, achieving a 0% success rate. In contrast, the CPR-trained agents achieved a 100% success rate, demonstrating their ability to handle adversarial communication effectively.

Red-Door-Blue-Door: This environment tests sequential coordination. The standard agent’s performance collapsed under adversarial messaging, with its reward dropping from 1.0 to 0.368. The CPR agent, however, learned a more cautious strategy that was highly resilient, achieving a reward of 0.497—nearly identical to the ideal, independent strategy.

Across all experiments, CPR enables agents to learn robust, resilient strategies that can withstand communication-based attacks, providing a clear path toward more secure real-world multi-agent systems.

DECEPTION IN MULTI-AGENT SYSTEMS

Citation: Please use the citation below to cite our work.

@misc{piazza2025robustcoordinationmisalignedcommunication,
      title={Robust Coordination under Misaligned Communication via Power Regularization}, 
      author={Nancirose Piazza and Amirhossein Karimia and Behnia Soleymanib and Vahid Behzadan and Stefan Sarkadi},
      year={2025},
      eprint={2404.06387},
      archivePrefix={arXiv},
      primaryClass={cs.MA},
      url={https://arxiv.org/abs/2404.06387}, 
}