MFSC: A Federated Learning and Semantic Communication Framework for Efficient Digital Twin Creation in Metaverse

In Metaverse, creating a digital twin of real-world objects involves transmitting a digital copy to the Metaverse Service Provider (MSP). Some research works employ semantic communication to create a digital twin of Metaverse instead of raw data, helping to deal with communication overload. However, ensuring the quality of semantic information is critical to building an optimal digital replica in Metaverse. In this paper, we propose a new framework, namely Metaverse-based Federated Learning and Semantic Communication (MFSC), where the MSP multi-servers are associated with a set of semantic extraction learners (IoT devices) and assign them semantic extraction tasks using different compression rates based on their qualities of communication channel, and computational resources. MSP servers send the semantic extraction models to their respective incentivized IoT devices, which jointly train the model locally on their local data, helping to enhance semantic extractions accuracy and the reconstructed data qualities and maximizing the utilities for both MSP and distributed IoT devices. We optimize resource allocation by optimizing the models compression rates and association between IoT devices and MSP servers, while aligning with requirements of servers and device resource limitations. Due to the non-convexity, we propose a partially decomposed approach to breaking up our problem into two sub-problems: Association and Compression Rate Optimization (ACRO), considering the utility of Metaverse and IoT devices and energy consumption. Simulation results demonstrate that the proposed framework achieves near-optimal efficiency in energy consumption and system utility.