Linkage on security, privacy and fairness in federated learning: New balances and new perspectives.

Abstract

Federated learning (FL) has gained significant attention for its ability to facilitate decentralized model training while preserving data privacy. However, privacy leaks, security threats, and fairness concerns remain critical challenges in FL. Existing studies often investigate these aspects in isolation, yet recent research suggests intricate interdependencies among them. For instance, enhancing fairness may compromise privacy, while improving security mechanisms can impact fairness. This survey provides a comprehensive analysis of the interplay between privacy, security, and fairness in FL, identifying key trade-offs and interconnections. Specifically, we argue that fairness can serve as a bridge between privacy and security, influencing model optimization and protection strategies. Additionally, we explore how gradient sharing-the core mechanism in FL-contributes to privacy and security risks. Building upon our observations, we categorize existing research, analyze potential threats and defense mechanisms, and highlight open challenges. We conclude by outlining promising directions for future research to establish a more balanced and robust FL framework.