TLDR: A team led by Professor Han Seung-yeol at UNIST has developed ‘Wolfpack Attack,’ an artificial malfunction strategy, and ‘WALL,’ a defense framework, to bolster the resilience of cooperative AI systems in drones and robots. This innovative technology simulates intense crisis scenarios to rigorously test and improve multi-agent cooperative structures, drawing inspiration from wolf hunting tactics that isolate weak individuals. The ultimate goal is to foster robust cooperative AI models for applications in autonomous drones, smart factories, and swarm robotics, ensuring their functionality even under challenging conditions.
Researchers at the Ulsan National Institute of Science and Technology (UNIST) have unveiled a groundbreaking approach to fortify cooperative artificial intelligence (AI) systems, particularly those employed in autonomous drones and robots. On July 30, Professor Han Seung-yeol’s team from the Graduate School of Artificial Intelligence announced the development of ‘Wolfpack Attack,’ an artificial malfunction attack strategy designed to systematically dismantle multi-agent cooperative structures, and ‘WALL’ (Wolfpack-Adversarial Learning for MARL), a corresponding defense framework that utilizes this attack for enhanced learning.
Cooperative technologies, such as drones flying in formation for surveillance or multiple robots collaborating in smart factories, rely heavily on the seamless interaction of AI agents. However, these systems are vulnerable to collapse under adverse conditions like severe weather or sensor malfunctions. The newly developed technology aims to create and ‘preview’ intense crisis situations, thereby enabling the maintenance of robust cooperation even in real-world scenarios.
The ‘Wolfpack Attack’ strategy is particularly innovative. Unlike traditional methods that randomly disrupt single agents, this approach first induces a malfunction in one agent and then sequentially creates problems for other agents attempting to assist, ultimately leading to the collapse of the entire cooperative structure. This strategy mimics the hunting behavior of wolves, which isolate weak individuals and then sequentially subdue those coming to their aid. The attack model incorporates a transformer-based prediction model that automatically determines the initial attack timing by calculating future losses, with subsequent targets identified by analyzing behavioral changes in agents sensitive to cooperation.
Lee Seon-woo, the study’s first author, explained the significance of this new method: “Previously, we only checked how well AI functioned in predetermined situations, but this attack strategy creates crisis situations that constantly change and are difficult to predict, like real-life scenarios, allowing us to evaluate how well AI responds within them.”
Complementing the attack strategy, WALL is a defense learning structure that integrates these disruption tactics into AI training environments. Experimental results have demonstrated that AI models trained using WALL exhibit high adaptability and stable cooperative performance. For instance, they were able to reach target points without collisions or maintain formation while moving objects, even when encountering position errors or communication delays.
Professor Han emphasized the broader implications of their research, stating, “The technology developed this time can be used for accurate performance evaluation of cooperative AI models and creating cooperative AI models that are strong in crisis situations.” He further added that it could significantly contribute to the advancement of autonomous drone, smart factory, and swarm robot industries, particularly in military and disaster response applications.
Also Read:
- Australian Report Highlights Novel Risks in Multi-Agent AI Deployments
- KISA Spearheads Asia-Pacific Ransomware Defense Against AI-Powered Threats
This research received support from the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning and Evaluation (IITP) through several key projects, including the “Regional Intelligence Innovation Talent Training Project,” “Development of Human-Centered Artificial Intelligence Core Source Technology,” and “Support for Artificial Intelligence Graduate School (UNIST).” The findings have been recognized and accepted at the International Conference on Machine Learning (ICML) 2025, held in Vancouver, Canada, from July 13 to 19. Out of approximately 12,107 papers submitted globally, only 3,260 were accepted, underscoring the prestige of this achievement.
