Enhancing Federated Learning Robustness and Fairness in Non-IID Scenarios

Abstract

Federated Learning is a distributed machine learning paradigm that allows multiple clients to collaboratively train a joint model without sharing the raw data. Despite its advantages, FL faces the security issues inherent to its decentralized nature, and FL clients often encounter unfair treatment from the design that prioritizes server interests. Today, many studies have been proposed to mitigate the research gap; nevertheless, in the absence of a non-IID setting, ensuring robustness and fairness in FL remains an open problem. Therefore, in this thesis, we study several topics on the robustness and fairness of FL in non-IID scenarios, including attack surface reduction, poisoning attack defense, and implicit class-level fair enhancement. We start by investigating FL's non-IID resource and propose the Mini FL framework. Based on a predefined grouping principle, Mini FL assigns similar clients to different groups and aggregates them respectively to achieve attack surface reduction. Then, we focus on defending against FL poisoning attacks. For the Label Flipping Attack, we introduce the HSCS FL method. It evaluates the accuracy of each class in both global and local models in each iteration. These accuracies are then translated into a score, and only clients with top scores are included in the current aggregation. For the Class Imbalance Attack, we introduce the Class-Balanced FL framework. This approach dynamically determines the aggregation weight for each client, considering their potential contribution to the current global model, thereby preventing the joint model biases toward specific data distributions. Lastly, we propose the ICB FL method to enhance FL fairness. This framework enables the server to identify implicit classes and dynamically distribute weights, ensuring a similar learning performance across these implicit classes. We provide mathematical proofs for each scheme and framework we proposed and conduct experiments to show their effectiveness.