A Fourier transform-based homomorphic friendly convolutional neural network efficient inference method

Abstract

This invention relates to the field of homomorphic encryption and privacy-preserving deep learning technology, specifically a high-efficiency inference method for homomorphic-friendly convolutional neural networks based on Fourier transform. Addressing the issues of slow online generation of evaluation representations for plaintext weights and high memory consumption for key switching in homomorphic-friendly convolutional inference, this invention demonstrates that the evaluation representation generated after two Fourier transforms of periodic plaintext weights retains its periodicity. Therefore, only one period's plaintext weight evaluation representation needs to be stored, saving memory usage for plaintext weights. Ciphertext messages use the same arrangement and the number of basic rotation keys is fixed, reducing the memory consumption required for rotation keys. Furthermore, the online convolution stage only requires copying and decompressing small-volume compressed weights, saving time on online generation of plaintext weights. This method ensures data privacy and, compared to existing homomorphic-friendly convolutional neural network inference methods, has lower memory consumption and faster inference speed.

Images