Towards Safety in Multi-agent Reinforcement Learning through Security and Privacy by Design

Kyle Tilbury, Bailey Kacsmar, Jesse Hoey

Keywords: Multi-agent reinforcement learning, Safety in multi-agent reinforcement learning, Security by design, Privacy by design, Privacy enhancing technologies

Abstract: In multi-agent reinforcement learning (MARL), the integration of security and privacy by design is critical for safe deployment in real-world applications. This position paper explores the unique security and privacy challenges inherent to MARL, identifying potential attack vectors and their implications on system security and user privacy. We emphasize the necessity of embedding security and privacy considerations starting from the initial stages of designing MARL systems, especially in settings involving humans. We highlight theoretical foundations and potential deployment challenges, advocating for a design paradigm that prioritizes security and privacy by design in MARL systems.

Submission Number: 13