An Efficient SVM-Based Method for Client Access Permission Distribution in Federated Learning

Federated learning has become a promising distributed machine learning paradigm capable of breaking down the information silos while achieving privacy protection. However, ensuring the reliability of individual participant remains a major challenge, that is, to ensure that each participant contributes positively to the federation. In this work, we focus on the candidate stage and assume all candidate clients are kind. Given a candidate client who wishes to join a federation that has already been trained, instead of calling for a re-training globally, we proposed an efficient SVM-based method which can quickly evaluate his contribution to the federation by classifying him based on the performance of his local model and original global model on a shared test set. Theoretically, we proved the convergence of this method by using SMO algorithm. Empirically, we also show that this method has stable performance on different datasets while significantly reducing the communication cost incurred by global training.