A privacy protection strategy based on homomorphic encryption and neural networks

In recent years, incidents of data abuse and privacy leakage remind people that improper use of big data technology can bring disastrous consequences. To realize the true value of data, data sharing is needed. This paper adopts a training method for a privacy-protecting neural network model based on homomorphic encryption in a cloud computing environment. It designs a piecewise linear activation function considering the role of the activation function for homomorphic computing. A ciphertext update strategy is proposed to control weight ciphertext noise growth. Combined with cloud computing, complex neural network computing tasks are entrusted to the cloud. Applied to data privacy protection, it is observed that AlexNet's PPNN takes 293.5 seconds for inference on CIFAR-10 or MNIST. The lightweight architecture SqueezeNet generated by EVA amplifies AlexNet's layers and adds rescaling. Although homomorphic encryption in SqueezeNet is slower, it reduces the delay to 72.7 seconds and ensures data security.