Abstract:

In the era of big data, data privacy protection in the field of machine learning has become increasingly important. In multi-party learning scenarios, attackers can reverse-engineer original data features from information such as gradients and model parameters. Moreover, some participants may collude for personal gain, sharing data that should remain confidential, thereby undermining the fairness and privacy requirements of multi-party learning. To address these issues, this paper proposed a fusion scheme of multi-key homomorphic encryption and differential privacy for distributed learning, namely the PrivMPL scheme, whose core objective was to achieve efficient model training while ensuring data privacy security. In this scheme, local clients used an aggregated public key to encrypt updated model parameters, and the decryption process required collaborative participation from all data users. The server achieved differential privacy by adding Gaussian noise to the aggregated parameters. The scheme effectively prevented privacy leakage caused by information sharing during multi-party training and was robust against collusion between data users and the server. To validate the effectiveness of the PrivMPL scheme, it is compared with a Paillier-based homomorphic encryption multi-party learning approach, using model accuracy as the evaluation metric. Experimental results show that the PrivMPL scheme achieves a significant improvement in model accuracy, further demonstrating its advantages in data privacy protection and model performance.

Key words: machine learning as a service, multi-party learning, differential privacy, multi-key homomorphic encryption