Toward Federated Learning Models Resistant to Adversarial Attacks
Fei Hu, Wuneng Zhou, Kaili Liao, Hongliang Li, Dongbing Tong
Abstract
With the popularity of the Internet of Things (IoT) and crowdsensing, sample data are more detailed and diverse. Users tend to avoid uploading personal data for privacy protection. Federated learning (FL) provides a new learning paradigm to complete training tasks without compromising user privacy. To deal with the challenge of malicious client attacks in FL systems, we present a robust framework for FL (RFFL) that can iteratively filter out malicious clients before federated aggregation, which results in defense capability against different types and levels of attacks. Then, we provide a convergence analysis of RFFL. Since client devices and edges distribute in different environments, which may cause client data heterogeneity, we offer an extension of RFFL (Ext. RFFL) to mitigate the effects of heterogeneity with no loss of defense capacity. Extensive experiments with real-world data sets demonstrate that our frameworks are competitive with benchmark algorithms in defending against various types and rates of attacks.