Poisoning attacks resilient privacy-preserving federated learning scheme based on lightweight homomorphic encryption

Federated learning ensures that multiple participants train the same model without leaking the local raw data. Each participant uploads the local gradient model instead of the original data, however, the uploaded local gradient model may contain certain sensitive information, which can be exploited by an adversary to break privacy protection. Meanwhile, some adversaries can make the model training results contrary to the expected results by tampering with the uploaded local gradient model or mixing malicious data into the local dataset, thereby inducing the model to produce wrong results for specific data. To this end, we devise a privacy-preserving federated learning scheme based on lightweight homomorphic encryption, which simultaneously reduces the weight of malicious data in gradient aggregation and supports anomaly detection of data, achieves the effect of resistance to poisoning attacks. Through theoretical analysis and experimental simulation, the proposed scheme has lightweight computation advantages compared with existing federated learning schemes.