FedTGR: Topology-Guided Byzantine-Resilient Federated Learning for Massive AI Agents

Abstract

5G-Advanced (5G-A) networks are evolving into intelligent coordination platforms for massive AI agents, including large-model-enabled devices, edge services, robots, vehicles, and network-side autonomous functions. These agents need to collaboratively adapt models and exchange knowledge without centralizing raw data, making federated learning (FL) a natural mechanism for distributed model evolution. However, compromised or malfunctioning agents may launch Byzantine attacks by uploading manipulated updates, thereby corrupting the shared model and weakening trustworthy coordination. Existing Byzantine-resilient FL defenses mainly rely on robust aggregation rules under the canonical star architecture, leaving the role of network-agent topology largely underexplored. This paper proposes FedTGR, a topology-guided grouping framework for Byzantine-resilient federated learning among massive AI agent clients. FedTGR overlays a virtual group topology on participating agent clients, filters suspicious updates through plug-in robust aggregation rules at a coordinator server, and enforces intra-group and inter-group model alignment through an ADMM-based consensus mechanism. By adjusting the number of groups, FedTGR reshapes the coordination topology and suppresses residual Byzantine influence while preserving the standard client-server FL workflow. Experiments on MNIST, CIFAR-10, and FEMNIST show that FedTGR improves test accuracy under diverse poisoning attacks and aggregation rules; a three-run ablation further shows reduced final-accuracy variability for the complete framework. These results support topology-guided coordination as a practical learning-layer mechanism for trustworthy federated model evolution in 5G-A autonomous agent systems.