Privacy-preserving convolutional neural network inference scheme based on homomorphic ciphertext transformation

To solve the problems of frequent interaction and low prediction accuracy of existing privacy-protected convolutional neural networks, a homomorphic friendly non-interactive privacy-protected convolutional neural network inference scheme was proposed via homomorphic ciphertext transformation. Utilizing the Pegasus framework, CKKS (Cheon-Kim-Kim-Song) ciphertext was used to parallelize convolution operations in convolution layer. In the activation layer and pooling layer, LWE ciphertext and LUT (look-up table) technology were used to calculate the activation function, maximum pooling and global pooling. Using the ciphertext conversion technology provided by the Pegasus framework, the conversion between different forms of homomorphic ciphertext is realized. Theoretical analysis and experimental results show that the proposed scheme can ensure data security, and has higher inference accuracy and lower calculation and communication overhead.