Federated Learning for Local and Global Data Distribution

Gaurav Goswami, Akshay Agarwal, Nalini K. Ratha, Richa Singh, Mayank Vatsa

Keywords: federated learning, generalization, personalization, bias, data privacy

TL;DR: We propose a novel approach for FL that preserves generalization during personalization and shows promising results in combined performance.

Abstract: Existing research in Federated Learning focuses on synthetic or small-scale datasets, with in-house distribution posing challenges for long-term real-world use cases. We propose a novel approach that maximizes in-house (local) distribution gains while focusing on generalization. Experimental results on several datasets demonstrate the efficacy of the proposed approach.