Personalized Wireless Ad Hoc Federated Learning for Label Preference Skew
Ryusei Higuchi, Hiroshi Esaki, Hideya Ochiai · 2023
Wireless ad hoc federated learning (WAFL) has been proposed to allow fully distributed collaborative learning among the physically encountered devices in a peer-to-peer manner without relying on any centralized mechanisms. The previous studies of WAFL mainly focused on the generalization of models from distributed users over label distribution skew cases in not independent and identically distributed (Non-IID) scenarios. In general, the generalized model does not always provide correct answers to each individual because each person also has label preference skews. In this paper, we study the personalization of WAFL, especially focusing on label preference skew problems. We take two parameter de-coupling approaches considering public layers and private layers in machine learning model architecture for such personalization. We have carried out two benchmark-based evaluations using modified MNIST and MobiAct datasets. The results indicate that the structure of the model with public layers near the input and private layers near the output performs better and achieves higher accuracy than the original WAFL.