A Practical Approach to Federated Learning

Abstract

Machine learning models benefit from large and diverse training datasets. However, it is difficult for an individual organization to collect sufficiently diverse data. Additionally, the sensitivity of the data and government regulations such as GDPR, HIPPA, and CCPA restrict how organizations can share data with other entities. This forces organizations with sensitive datasets to develop models that are only locally optimal. Federated learning (FL) facilitates robust machine learning by enabling the development of global models without sharing sensitive data. However, there are two broad challenges associated with deploying FL systems: privacy challenges and training/performance-related challenges. Privacy challenges pertain to attacks that reveal sensitive information of local client data. Training/Performance-related challenges include high communication costs, data heterogeneity across clients, and lack of personalization techniques. All these concerns have to be addressed to make FL practical, scalable, and useful. In this thesis, I discuss techniques I've designed for addressing these challenges and describe two systems that I've developed to mitigate them - PrivacyFL, a privacy-preserving simulator for FL, and DynamoFL, an easy-to-use production-level system for FL.