复旦大学数字信号处理与传输实验室的郝子若同学的论文Invariant Federated Learning for Edge Intelligence: Mitigating Heterogeneity and Asynchrony via Exit Strategy and Invariant Penalty，被IEEE Transactions on Mobile Computing 录用

论文摘要：This paper presents an invariant federated learning framework for resource-constrained edge intelligence. The framework first identifies and excludes low-value or aberrant clients, and then enforces model-level invariance to preserve generalization. Motivated by pretraining-based client valuation, we show that some clients may have negligible or even detrimental contributions to global aggregation under severe heterogeneity or asynchrony. Accordingly, we propose a practical exit strategy based on client utility and delay, which removes such clients from the main training loop to protect in-distribution (ID) performance. To maintain robust performance for excluded and resource-constrained clients, we introduce FedIPG (Federated Learning with Invariant Penalty for Generalization), which promotes approximate orthogonality between invariant and heterogeneous parameter components through a gradient-based penalty. We provide theoretical analysis showing that low-value client exit can improve aggregation and that FedIPG tightens an upper bound on out-of-distribution (OOD) risk without increasing communication overhead during formal training. Extensive experiments on four domain-generalization benchmarks and multiple client scales demonstrate that the combined exit + FedIPG framework improves ID accuracy, enhances OOD generalization over recent baselines, and maintains stable convergence. Visualization results further indicate that FedIPG reduces reliance on confounding features and encourages more task-relevant feature usage.

论文作者：Ziruo Hao, Zhenhua Cui, Tao Yang, Yunjie Fang, Xiaofeng Wu, Hui Feng