Research on Federated Learning Scheme Based on Secure Aggregation

As an emerging distributed learning framework, federated learning allows multiple clients to jointly train a global model without sharing the original data, which effectively protects data privacy. However, it also makes federated learning vulnerable to malicious clients. This paper aims to study the problem of secure aggregation in federated learning, and proposes a FLSAgg algorithm based on multi-dimensional credibility score. The algorithm combines the evaluation of multiple dimensions such as loss function, cosine similarity, historical performance score and malicious model identification, dynamically adjusts the weight of each evaluation dimension, and selects the client for model aggregation. The experimental results show that the FLSA algorithm performs well in resisting malicious client poisoning attacks, and effectively improves the stability and accuracy of the global model.