A Review of Federated Learning Approaches for PrivacyPreserving Artificial Intelligence

Abstract

To tackle the core limitations of privacy in collaborative AI model development, federated learning (FL) has now become a cornerstone approach for training machine learning models across multiple data sources while maintaining the privacy of local data. We review the algorithms and techniques, privacy safeguards, communication efficiency approaches, and practical applications of federated learning for privacy-preserving AI. We systematically review the landscape of FL algorithms, ranging from the pioneering FedAvg to recent personalized, strong, and DP versions, as well as other supporting building blocks for privacy, such as differential privacy, secure multiparty computation, and homomorphic encryption. Application areas studied are in the fields of healthcare, mobile computing, finance, autonomous driving and pharmaceuticals. Here, we discuss the most important open challenges: data heterogeneity, communication limitations, Byzantine robustness, privacy-utility, and regulatory compliance. The review offers researchers and practitioners a structured ground for being able to navigate the terrain of federated learning, as it outlines the key characteristics of ten representative algorithms and six deployment case studies. There are suggestions on future directions in the areas of personalization, efficient communication, and theoretical privacy guarantees in realistic threat models.