A Protocol for Privacy Preserving Neural Network Learning on Horizontally Partitioned Data

This article presents a new approach for privacy preserving neural network training. Several studies have been devoted to privacy preserving supervised model learning, but little work has been done to extend neural network learning with a privacy preserving protocol. Neural networks are popular for many applications, among else those calling for a robust learning algorithm. In this study, we elaborate on privacy preserving classification as well as regression with neural networks on horizontally partitioned data. We consider a scenario of more than two parties that are semi-honest but curious. We extend the neural network classification algorithm presented by Rumelhart et al. with protocols for secure sum and secure matrix addition. The extended algorithm does not fully guarantee privacy in the sense defined by Goldreich, but we show that the information revealed is not associated to a specific party and could also be derived by juxtaposing the local data and the final model.