Abstract
In the rapidly advancing fields of drone swarms and cooperative robotics, AI agents embedded in individua drones and robots must collaborate seamlessly-such as drones flying in formation to encircle an enemy or multiple robots working together in smart factories. However, these multi-agent systems are vulnerable to disruptions caused by adverse conditions or malicious attacks, which can compromise their cooperation and operational integrity.
Addressing this challenge, a research team led by Professor Seungyul Han from the Artificial Intelligence Graduate School of UNIST has developed a groundbreaking adversarial attack framework inspired by wolf pack hunting strategies, alongside a corresponding defense training method. These innovations aim to evaluate and strengthen the robustness of multi-agent reinforcement learning (MARL) systems against coordinated disruptions.
Reinforcement learning enables AI agents to learn optimal behaviors through trial and error across diverse scenarios. In multi-agent settings, collaboration among agents typically ensures system resilience; if one agent encounters issues, others compensate to maintain overall performance. However, existing attack strategies that target individual agents often fall short in exposing vulnerabilities within these cooperative structures, especially under realistic conditions such as sensor failures, weather disturbances, or cyber-attacks.
The proposed Wolfpack Attack simulates a strategic assault where an initial agent is deliberately compromised, triggering a cascading failure among assisting agents-mirroring a wolf pack isolating and overpowering prey. This attack leverages advanced predictive models to determine the optimal moment to initiate disruption and to sequentially compromise agents sensitive to cooperative cues.
Complementing this, the researchers developed the 'WALL (Wolfpack-Adversarial Learning)' framework, which incorporates these adversarial scenarios into the training process. By exposing AI systems to simulated wolf pack attacks, WALL enhances their ability to withstand real-world disruptions, ensuring more stable and reliable cooperative behavior.
Experimental results demonstrate that AI agents trained with WALL exhibit remarkable resilience, maintaining coordination and task performance even under challenging conditions such as communication delays and sensor inaccuracies. This advancement not only provides a powerful tool for evaluating the robustness of multi-agent systems but also paves the way for deploying more resilient autonomous drones, robotic swarms, and industrial automation solutions.
Professor Han commented, "Our approach offers a new perspective on assessing and fortifying the cooperative capabilities of AI agents. By simulating sophisticated adversarial scenarios, we can better prepare systems for unpredictable real-world challenges, contributing to safer and more reliable autonomous technologies."
This research has been accepted for presentation at the prestigious International Conference on Machine Learning (ICML), held from July 13 to 19 in Vancouver, Canada. Out of over 12,000 submissions, only 3,260 papers were selected, underscoring the significance of this work. The study has been carried out with the support from the Ministry of Science and ICT (MSIT), the Institute of Information & Communications Technology Planning & Evaluation (IITP), and the UNIST Artificial Intelligence Graduate School.
Journal Reference
Sunwoo Lee, Jaebak Hwang, Yonghyeon Jo, and Seungyul Han, "Wolfpack Adversarial Attack for Robust Multi-Agent Reinforcement Learning," the 42nd International Conference on Machine Learning (ICML) 2025, Vancouver, Canada, (2025).
