Efficient Integer Vector Homomorphic Encryption Using Deep Learning for Neural Networks

Machine learning techniques based on neural networks have achieved significant applications in a wide variety of areas. There is a great risk on disclosing users’ privacy when we train a high-performance model with a large number of datasets collected from users without any protection. To protect user privacy, we propose an Efficient Integer Vector Homomorphic Encryption (EIVHE) scheme using deep learning for neural networks. We use EIVHE to encrypt users’ datasets, then feed the encrypted datasets into a neural network model, and finally obtain the trained model for neural networks. EIVHE is an innovative bridge between cryptography and deep learning, which aims at protecting users’ privacy. The experiments demonstrate that the deep neural networks can be trained by encrypted datasets without privacy leakage, and achieve an accuracy of 89.05% on MNIST. Moreover, this scheme allows us to conduct computation in an efficient and secure way.