Efficient Byzantine-Robust Federated Learning Based on the Multimessage Shuffle Protocol for Consumer Internet of Things

The accelerated advancement of Consumer Internet of Things (CIoT) devices has resulted in their pervasive integration into the fabric of contemporary lifestyles. As these devices are capable of collecting vast quantities of data pertaining to users, they present a significant risk to the privacy and security of individuals. Federated learning (FL) is an advanced machine learning paradigm that prioritises the preservation of user data privacy and enhances security by eliminating the need for centralised data collection. As a result, it is a promising approach to address user data privacy and security concerns. However, FL still faces challenges such as privacy attacks, Byzantine attacks and high communication overheads. To address these challenges, we propose an efficient Byzantine-robust FL method based on multimessage shuffle protocol (MSP). To achieve privacy protection and Byzantine robustness, we introduce an innovative aggregation strategy inspired by the MSP for differential privacy (DP). To improve communication efficiency, we develop an adaptive gradient compression scheme. Theoretical analysis indicates that our MSP is capable of satisfying the requisite DP conditions. Compared with the state-of-the-art methods, our proposed scheme demonstrates superior convergence capabilities on the real-world dataset when approximately 40% of participants are Byzantine, achieving accuracies of 96.78%, which illustrates its strong robustness. Upon reduction of the transmitted model parameters to 1% of the original size, the proposed scheme exhibits a maximum accuracy degradation of 0.35%. which significantly reduces the communication overhead while achieving high accuracy. In addition, the scheme also achieves an accuracy of 88.73% on the Non-iid MotionSense dataset, which is superior to other schemes.