Scalable Active User Detection Based on Federated Learning in Cell-Free Massive MIMO System

In the context of 6G, one of the emerging and dominant services is massive ultra-reliable and low-latency communication (mURLLC). To reduce signaling overhead and access delay, mURLLC widely adopts grant-free random access (GFRA). The first step of GFRA is active user detection (AUD), which aims to accurately and promptly identify the set of active users and determine the active channels. Traditional AUD schemes are based on iterative computations for a large number of users, leading to excessive computational load and processing delay, which severely limits the system scalability in mURLLC. Considering the high reliability and near real-time requirements of mURLLC, based on the spatial extension units (SEUs)-assisted cell-free massive multiple-input multiple-output (CF-mMIMO) system, we exploit channel sparsity to propose a scalable AUD method based on federated learning (FL). Our scheme integrate the model parameters trained at various SEUs at the centralized computing unit (CPU) without data sharing, thereby enhancing the channel detection performance within each AP cluster. Simulation results demonstrate that our scheme significantly improves detection performance compared to traditional AUD methods.