Title:MrTF: Model Refinery for Transductive Federated Learning

Abstract:We consider a real-world scenario in which a newly-established pilot project needs to make inferences for newly-collected data with the help of other parties under privacy protection policies. Current federated learning (FL) paradigms are devoted to solving the data heterogeneity problem without considering the to-be-inferred data. We propose a novel learning paradigm named transductive federated learning (TFL) to simultaneously consider the structural information of the to-be-inferred data. On the one hand, the server could use the pre-available test samples to refine the aggregated models for robust model fusion, which tackles the data heterogeneity problem in FL. On the other hand, the refinery process incorporates test samples into training and could generate better predictions in a transductive manner. We propose several techniques including stabilized teachers, rectified distillation, and clustered label refinery to facilitate the model refinery process. Abundant experimental studies verify the superiorities of the proposed \underline{M}odel \underline{r}efinery framework for \underline{T}ransductive \underline{F}ederated learning (MrTF). The source code is available at \url{this https URL}.