Enhancing Federated Learning Robustness Using Data-Agnostic Model Pruning

Abstract

Federated learning enables multiple data owners with a common objective to participate in a machine learning task without sharing their raw data. At each round, clients train local models with their own data and then upload the model parameters to update the global model. This multi-agent form of machine learning has been shown prone to adversarial manipulation by recent studies. Byzantine attackers impersonated as benign clients can stealthily interrupt or destroy the learning process. In this paper, we propose FLAP, a post-aggregation model pruning technique to enhance the Byzantine robustness of federated learning by effectively disabling the malicious and dormant components in the learned neural network models. Our technique is data-agnostic, without requiring clients to submit their dataset or training output, well aligned with the data locality of federated learning. FLAP is performed by the server right after the aggregation, which renders it compatible with an arbitrary aggregation algorithm and existing defensive techniques. Our empirical study demonstrates the effectiveness of FLAP under various settings. It reduces the error rate by up to 10.2% against the state-of-the-art adversarial models. Moreover, FLAP also manages to increase the average accuracy by up to 22.1% against different adversarial settings, mitigating the adversarial impacts while preserving learning fidelity.