Federated Learning: Decentralized Machine Learning for Privacy Preservation

Abstract

Federated Learning (FL) is an emerging paradigm in machine learning that enables multiple devices or organizations to collaboratively train a global model without sharing raw data​. This decentralized approach addresses data silo and privacy challenges by ensuring sensitive information remains local to each participant. In this article, we survey recent advances in FL and propose a hypothetical experiment comparing federated and centralized learning on a standard dataset. Our experimental methodology employs the Federated Averaging algorithm​ across simulated clients and measures model accuracy, convergence, and privacy trade-offs. The results indicate that FL can achieve performance nearly comparable to a centralized model while preserving privacy: for example, achieving roughly 98% accuracy versus 98.5% in the centralized case, with only marginal degradation when adding differential privacy. These findings reinforce the practicality of FL in privacy-sensitive domains, supporting trends observed in applications like mobile keyboards and healthcare​. We conclude that FL is a viable strategy for privacy-preserving ML and discuss future challenges in scalability and security.