Decentralized security for multi-user VR applications using multi-agent reinforcement learning
Virtual Reality (VR) is being increasingly adopted in collaborative settings, where maintaining secure visual outputs from VR IoT devices is critical to prevent malicious content and interactions. Previous approaches have relied on centralized security mechanisms at the server level. Although simple to deploy, they often suffer from single points of failure, performance bottlenecks, and scalability issues. In this paper, we introduce a novel and highly efficient framework capable of decentralizing the security mechanism by shifting protection to the end-user level. Our framework also accounts for the diverse nature of end users, which results from different levels of privileges and permissions. We formulate the problem within a Multi-Agent Reinforcement Learning framework. Empirical experiments demonstrate the effectiveness of our approach in safeguarding VR environments, achieving 100% precision, 97% recall, and 97% F1-score. Moreover, we conducted extensive ablation studies to evaluate various design choices, including reward schemes, network architectures, training protocols, and aggregation strategies. Finally, we open-source our code to support reproducibility and enable further benchmarking.