Privacy-preserving Deep Learning for Grey Scale Image Classification with Pixel-based Encryption

Privacy-preserving deep learning methods have become a hot topic these years because of the potential information security problem in the process of training and testing data in machine learning (ML). Deep neural networks (DNNs) are widely used in image classification, hence images could be the input of the ML algorithm, and a certain amount of vital information can be involved in these images. This paper focuses on the encryption of the grey scale images using DNNs. Two prevalent image encryption techniques and the improved homomorphic encryption (HE) are integrated to protect the information in the images without losing much accuracy of the model. The effectiveness and security of the proposed method are evaluated on the benchmark datasets and the results of the comparison experiment verify the superiority of the proposed method in terms of effectiveness and security.