Federated Learning: Strategies for Communication Efficiency – A Comprehensive Review

Abstract

Federated Learning (FL) is a machine learning paradigm that employs distributed devices at disparate locations to perform collaborative model training without having centralized sensitive information. While FL alleviates privacy and regulatory challenges, the massive volume of server-client communication introduces a bottleneck, particularly in low bandwidth. This paper brings together foundational and contemporary research to synthesize methods that enhance the effectiveness of communication in FL. After the initial work of Konečný et al. (2016) and McMahan et al. (2017), discussions on secure aggregation (Bonawitz et al., 2017), Homomorphic Encryption schemes (Gentry, 2009; Acar et al., 2018; Phong et al., 2018), vertical FL platforms (Hardy et al., 2017), domain-specific methods such as SecureBoost (Cheng et al., 2021), and the attack concern from membership inference attacks (Truex et al., 2018) continue. Together, these papers demonstrate the trade-off among security, efficiency, and scalability and map out the road ahead for future research. The survey is completed with a general comparison of the main contributions and indicates the open problems to finish rigorous, communication-efficient Federated Learning.   :