Secure federated learning based on multi-round critical parameters

Federated learning enables global model training in a distributed manner without requiring clients to upload private data. However, the presence of malicious clients uploading incorrect model updates can severely degrade the performance of the global model. Existing methods often rely on limiting the number of malicious clients or require access to additional clean datasets. To address these limitations, we propose a secure aggregation algorithm named FedMCP to detect and remove malicious clients. FedMCP first constructs judgment vectors based on parameter importance. It then distinguishes benign from malicious clients based on the similarity between their judgment vectors. Moreover, the discrimination between benign and malicious clients is achieved by incorporating the historical distribution characteristics of the judgment vectors from known benign clients. Finally, the Isolation Forest algorithm is employed to remove malicious models those mimic the judgment vectors. Even in individual rounds where a large number of malicious clients participate in training, FedMCP can still accurately distinguish between benign and malicious clients. FedMCP maintains high identification accuracy for both benign and malicious clients, even under heavy adversarial participation. Extensive experiments across multiple datasets and models confirm that FedMCP effectively identifies and excludes malicious clients, achieving superior robustness and performance. • A novel, importance-based aggregation method for Byzantine-robust FL. • Demonstrates versatility across various models and datasets. • No prior knowledge of malicious clients number and additional clean dataset. • Effective even when malicious clients constitute the majority.