FedEqual: Defending Model Poisoning Attacks in Heterogeneous Federated Learning

Abstract

With the upcoming edge AI, federated learning (FL) is a privacy-preserving framework to meet the General Data Protection Regulation (GDPR). Unfortunately, FL is vulnerable to an up-to-date security threat, model poisoning attacks. By successfully replacing the global model with the targeted poisoned model, malicious end devices can trigger backdoor attacks and manipulate the whole learning process. The traditional researches under a homogeneous environment can ideally exclude the outliers with scarce side-effects on model performance. However, in privacy-preserving FL, each end device possibly owns a few data classes and different amounts of data, forming into a substantial heterogeneous environment where outliers could be malicious or benign. To achieve the system performance and robustness of FL's framework, we should not assertively remove any local model from the global model updating procedure. Therefore, in this paper, we propose a defending strategy called FedEqual to mitigate model poisoning attacks while preserving the learning task's performance without excluding any benign models. The results show that FedEqual outperforms other state-of-the-art baselines under different heterogeneous environments based on reproduced up-to-date model poisoning attacks.

Ähnliche Arbeiten