Personalized Federated Learning Framework for Heterogeneous Multi-Task Semantic Communication

A recent study published on arXiv (2608.15256) presents a tailored decentralized federated learning (DFL) framework aimed at enhancing distributed semantic communication (DSC) networks. The researchers pinpoint a limitation in current DFL methods: the topology-agnostic aggregation in diverse, multi-task settings leads to negative transfer and overconsensus bias (OCB). They propose a multi-path routing mechanism driven by node-level policies that distinguishes between task-specific features and shared representations. Additionally, a 'communication-while-aggregation' protocol fine-tunes a consensus matrix based on task affinities. The paper includes a Lyapunov drift analysis that uncovers a U-shaped trade-off in convergence at optimal depth, enabling each node to uphold a personalized model while reaping the benefits of collaborative training, particularly for IoT networks and edge AI challenges.