Artificial Intelligence Service Provision with Secure Federated Learning in Metaverse

The Metaverse represents a paradigm-shifting digital ecosystem that seamlessly integrates artificial intelligence (AI) with immersive augmented/virtual reality (AR/VR) systems, delivering intelligent agent-driven services through multi-sensory interaction modalities. However, training AI models in the Metaverse faces significant challenges, including data privacy concerns, communication overhead, and malicious attacks that lead to the uploading of invalid local models, which degrade global model performance and potentially cause system failures. To tackle these challenges, we propose an incentive-driven federated learning (FL) scheme for the Metaverse, designed to counter malicious attacks and encourage honest user participation. Specifically, We first construct a Bayesian game to model the interactions between Metaverse users and AI agents, capturing the strategic decision-making of both AI agent and dishonest users. We analyze the pure-strategy Bayesian Nash equilibrium (BNE) and derive the condition for its existence. When this condition is not met, we further examine a mixed-strategy BNE for more practical scenarios, determining the optimal strategies for all parties. The effectiveness of the proposed scheme is validated through extensive simulations. Experimental results show that the proposed scheme significantly improves AI model accuracy compared to benchmark approaches.