RobFL: Robust Federated Learning via Feature Center Separation and Malicious Center Detection

In recent years, the integration of federated learning and deep learning technologies has become increasingly prevalent in privacy-preserved scenarios, such as smart health applications and automatic financial support. However, the inherent robustness issue in deep learning poses potential risks to federated learning systems when subjected to various attack methods. These attacks can inflict damage during the training and testing phases, perturbing models and inputs. To enhance the robustness of existing federated learning systems, we propose a novel framework called RobFL. This framework incorporates a unique feature learning module - feature center separation learning - that is specifically designed to increase the margins between different classes in the feature space, thereby augmenting the difficulty of attacks employing imperceptible perturbations on inputs. Furthermore, we design a malicious center detection method to detect malicious clients and mitigate their adverse impact. Extensive experiments substantiate the robustness of our proposed framework, RobFL, demonstrating its resilience against both evasion attacks and poisoning attacks.