FedCLLM: Federated client selection assisted large language model utilizing domain description

Abstract

Federated Learning has become an emerging topic since the rise of privacy regulation regarding personal data protection and sensitivity. It provides a decentralized training approach to train a global model between a server and multiple clients while ensuring client data confidentiality. However, in practical scenarios, there are malicious clients within a large pool of client candidates, and selecting trustworthy honest clients becomes a crucial problem. Some previous works tried to solve the problem by exchanging client data for comparison and using a labelled dataset for evaluating client models to select honest clients. However, they pose a limitation when the server possesses unlabelled data from other sources and has limited or unavailable resources to label the data. To address the issue, this work proposes FedCLLM, a novel approach using Large Language Models (LLM) proficiency in semantic tasks on text-based data to compare client and server domain descriptions summary in a text format and assess their similarity. Experiments on popular benchmark datasets show that FedCLLM effectively distinguishes honest clients from potentially malicious ones and outperforms other previous works in terms of performance.