Federated Learning with Provable Security Against Malicious Clients in IoT Networks
Pubudu L. Indrasiri, Dinh C. Nguyen, Bipasha Kashyap, Pubudu Nishantha Pathirana · 2023
Federated learning (FL) is a promising approach for allowing many Internet of Things (IoT) devices to perform large-scale machine learning while preserving data privacy. However, traditional federated learning is vulnerable to attacks from loT devices (clients) that submit incorrect updates, which can harm the global model. In order to tackle this problem, previous approaches have concentrated on developing secure aggregation rules, however, these methods still have their limitations. Thus, we propose a new framework called FedXPro that combines the PCIBC-DIM neural network and the Geometric Median (GM) to detect and filter out Byzantine clients in FL. The PCIBC-DIM is a neural network that performs Bayesian inference by combining priors and likelihoods to determine posteriors. To identify the priors, the proposed framework uses GM to specify valid clients as priors for the PCIBC-DIM network. During the training phase, the proposed framework estimates the distribution of valid clients chosen from the GM. In the testing phase, the framework tries to reconstruct the same distribution from updates received from clients with respect to the distribution from valid clients. The reconstruction power is then used to filter out invalid clients and prevent their updates from being included in the final model. Our simulations demonstrate that FedXPro outperforms other state-of-the-art methods in terms of accuracy, assured convergence rate, and attack detection in various network settings.