Efficient Privacy-Preserving Machine Learning with Homomorphic Encryption through Pruning

Homomorphic encryption (HE) is an important and promising technology in privacy-preserving machine learning (PPML), which can perform calculations without decrypting data to ensure data privacy. It is widely applied in PPML. However, HE incurs extremely high computational costs, particularly for rotation and multiplication operations, which are far more expensive than plaintext computations. This paper proposes a HE-based weight pruning convolutional neural network (CNN) model — PrHECNN. By pruning convolutional kernels and global weights, PrHECNN effectively reduces the number of rotation and multiplication operations, significantly lowering computational and memory costs, and providing efficient and accurate inference for PPML. Experiments on the MNIST dataset show that PrHECNN improves the encrypted inference speed by 17.33%-74.23% and reduces memory usage by 36.67%-87.72%, showing the best performance among all compared methods. Additionally, the model achieves an accuracy of 98.55%, nearly identical to the plaintext model.