Privacy-Preserving Multi-party Neural Network Learning Over Incomplete Data
Fengxiao Liu, Gang Shen, Mingwu Zhang · 2022
Neural networks have been widely used in var-ious data-driven businesses. In many scenarios, it is neces-sary for multiple data owners to train neural networks in collaboration to improve the accuracy of learning results, in which privacy becomes a considerable challenge. Many existing cryptographic schemes have guaranteed the privacy of neural network learning when the training datasets are horizontally or vertically partitioned. By contrast, researches into arbitrarily partitioned data are less common. A non-negligible situation is that data from multiple parties usually contain a certain amount of missing data. We propose a privacy-preserving neural network encryption learning scheme, which allows more than two participants to collaboratively train neural network model, particularly for incomplete datasets with arbitrary partition. This scheme preserves the data privacy by using homomorphic encryption, then all parties upload data to a cloud server, and continuously collaborate with all others to train the neural network model. In this process, we design a share splitting scheme with FHE to protect the security of the intermediate results of the computation. We evaluated our scheme on real datasets and compared it with related schemes. The results showed that our scheme has high classification accuracy and low cost. We also compared the incomplete datasets (simulated by our scheme) with the imputed datasets, proving that our data imputation benefits neural network learning.