Mengjia Lu, Ping Zhang, An Bao, Keyi Xu, Jiechang Wang · Research Square 2026 · 2026
DOI: 10.21203/rs.3.rs-11417646/v1
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Personalized federated learning (PFL) aims to balance shared knowledge with client-specific adaptation under heterogeneous data. However, data that are not independently and identically distributed (non-IID), together with partial client participation, can cause shared representations to vary across communication rounds, hindering knowledge transfer. Existing methods mainly address heterogeneity through model personalization or historical knowledge transfer, while sample-level temporal representation changes remain underexplored. We propose FedEAGD, a trajectory-guided PFL framework that uses a client-specific exponential moving average (EMA) teacher to summarize shared-model history. The discrepancy between each sample's current and historical shared representations quantifies temporal drift and adaptively determines the strength of historical distillation. Progressive gated residual fusion (PGRF) introduces personalized information according to a stage-wise schedule. Representation anchoring and residual decoupling regularize this process, while post-aggregation calibration and late-stage server smoothing coordinate local adaptation with global model updates. Only the shared feature extractor is communicated, while personalized components remain local. Experiments on MNIST, CIFAR-10, CIFAR-100, and DomainNet evaluate FedEAGD under varying data heterogeneity, client participation rates, and federation sizes. On CIFAR-100 with a Dirichlet concentration parameter of $\alpha=0.1$, FedEAGD achieves 63.05\% accuracy, outperforming the strongest evaluated baseline by 5.96 percentage points.
No comments yet — start the discussion below.