Privacy Prediction of Lightweight Convolutional Neural Network

The growing popularity of cloud-based deep learning raises a problem about accurate prediction and data privacy. Previous studies have implemented privacy prediction for simple neural networks. Since more complex neural networks require more computational overhead, existing privacy prediction schemes are inefficient. To tackles the above problem, this paper introduces a privacy prediction method for lightweight convolutional neural network (CNN) that can be applied to encrypted data. Firstly, the complex CNN is pruned into a lightweight network without compromising the original accuracy, which can realize secure prediction efficiently. Secondly, the FV homomorphic encryption scheme is adopted to encrypt the user’s sensitive data and each layer in CNN is calculated on the ciphertext, so as to protect user’s data privacy. Finally, the security analysis and experiment results demonstrate the privacy-preserving property and practicability of the proposed scheme, where the complex CNN on the MNIST data set can achieve more than 98% accuracy.