A Training Scheme of Deep Neural Networks on Encrypted Data
Liang Yuan, Gang Shen · 2020
Machine learning based on deep neural network has shown great potential in many application fields. Compared with traditional machine learning algorithms, training a reliable neural network model requires a huge amount of data. However, the data required to train a deep neural network model is often privacy sensitive. In order to preserve privacy data, users need to encrypt the data before uploading them to the cloud server. Since back-propagation algorithm works on plaintext, it is difficult to train a neural network model on ciphertext. To address this issue, we propose a new scheme to support training neural network over encrypted data. We use homomorphic cryptosystem to protect privacy feature data of users. We also improve the back-propagation algorithm in privacy-preserving environment to protect classification label of user. We implement the model based on LeNet-5 and present performance evaluation. The experimental results show that our scheme has similar accuracy to training the MNIST dataset in plaintext.