Design of data security check mechanism and research on prevention and control of privacy leakage based on federated learning

Aiming at the security challenges faced by federated learning in practical application, such as data reconstruction attack, gradient leakage and malicious node poisoning, this paper proposes a joint prevention and control system combining hierarchical dynamic security check with multi-level privacy enhancement. The system introduces a lightweight gradient anomaly detection mechanism in the local layer, and judges and cuts the abnormal gradient through dynamic threshold; Design a two-way challenge-response verification protocol in the aggregation layer to improve the consistency check ability of model update; Build a reputation system based on game theory at the global level to realize the dynamic identification and isolation of malicious nodes. Meanwhile, this paper designs multi-level privacy protection strategies, including local gradient sparsification, layered differential privacy (LDP), and trusted execution environment (TEE), balancing model performance and privacy security. Experimental results demonstrate that the proposed method significantly outperforms existing approaches across multiple datasets and attack scenarios, with the attack success rate reduced to a minimum of 3.2% while maintaining a model accuracy of over 90%. The method exhibits strong generalization ability and practical value.