FedGAT: Federated Graph Attention for User Association and Interference Mitigation in C-RAN

This paper presents FedGAT, a federated graph attention network for joint user association and inter-remote radio head (RRH) interference mitigation in cloud radio access networks (C-RAN). Centralized optimization requires global channel state information (CSI) at the baseband unit (BBU) pool, which adds fronthaul overhead and exposes user data. Existing federated learning schemes use flat local models that do not represent the interference graph, which lowers performance in heterogeneous multi-cell settings. In FedGAT, each RRH builds a local interference graph from its own CSI and trains a GATv2 model by local Adam gradient steps. Only the parameter increments are sent to the BBU pool, which combines them by weighted FedAvg without exchanging raw CSI. We formulate the joint user association and power allocation problem as a mixed-integer non-convex program and derive a graph neural network (GNN)-based continuous relaxation suitable for distributed training. A non-asymptotic convergence bound is obtained for non-identically distributed local interference graphs, showing that the optimality gap grows with the local step count and a measure of graph heterogeneity. Under the 3GPP Urban Macrocell channel model, FedGAT increases the steady-state weighted sum-rate by 18.7% over a federated multilayer perceptron baseline and by 49.2% over a centralized Graph Attention Network version 2 (GATv2) model at equal training budgets, while keeping raw CSI local. These margins are averaged over ten Monte Carlo channel realizations and reported with their standard deviations, and the fronthaul benefit of FedGAT refers to privacy and CSI-free inference after training rather than training-phase traffic.