NPNNL: A Non-interactive Privacy-preserving Neural Network Learning Scheme

Recently, the neural network has been widely employed to train predictive models in artificial intelligent applications, such as risk assessment, facial recognition, and disease diagnosis. Since the performance of neural network models is directly proportional to the amount of training data, collecting massive data is essential for training neural network models. However, collecting such amounts of data from different sources raises data security issues and thus may pose a major challenge to the development of neural networks. In this paper, by introducing two non-colluding servers, we design a non-interactive privacy-preserving neural network learning protocol, NPNNL. NPNNL enables two cloud servers cooperatively train neural network models and provide prediction services without the help of users. Additionally, NPNNL uses homomorphic encryption and data perturbation mechanism to protect the privacy of data and models, respectively. Extensive security analysis demonstrates that our scheme holds data privacy and model privacy. Finally, we evaluate the performance of our scheme on a real-world dataset. The experimental results show that the scheme is computationally and communicationally efficient.