Encrypted Data Learning and Prediction Using a BFV- based Cryptographic Convolutional Neural Network

Machine learning services are widely used for big data, cloud computing, and distributed artificial intelligence applications. Multiple parties participating in the provision of these services may access the users' sensitive data because most machine learning models use and share plaintext directly. Therefore, it is necessary to utilize cryptographic mechanisms for protecting user privacy. Homomorphic encryption provides an important information security guarantee for machine learning models. However, the complexity of fully homomorphic encryption increases with the depth of neural networks. Especially with the increase in the number of ciphertext multiplications, the time and space costs will also raise exponentially. Using homomorphic encryption in order to protect the model and data security while ensuring the computational efficiency of the employed model over encrypted data is a challenging problem. This paper proposes a BFV-based cryptographic low-latency convolutional neural network (CLOL-CNN) for solving this problem. This new network model performs deep learning and prediction over encrypted data instead of sharing plaintext data. A series of optimization operations are elaborately presented and implemented, such as cryptographic batch normalization, polynomial approximation, cryptographic convolution, and full cryptographic connection. The performance of the proposed model is evaluated with regard to its accuracy and computational overhead obtained by employing deep learning for homomorphically encrypted data. The experiments were conducted on a MNIST image dataset. The obtained results demonstrated that the proposed model has a higher accuracy and a lower time cost than other models and that it is an effective privacy-preserving deep neural network.