|基金项目:本课题得到浙江省自然科学基金(No. LQ20F020008)与中科院信工所开放课题(No. 2021-MS-03)提供资助。
|Empirical Study on the effects of BFV Scheme Configuration on Secure Neural Networks Inference
YANG Tao1, DONG Jianfeng1,2
|(1.School of Computer Science and Technology, Zhejiang Gongshang University, Hangzhou, 310012, China;2.State Key Laboratory of Information Security, Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China)
|Homomorphic encryption has emerged as an effective solution for preserving privacy in neural network applications. However, the parameter configuration of the BFV homomorphic encryption scheme has not yet been thoroughly investigated for various application domains. This has resulted in difficulties in applying homomorphic encrypted neural networks to other scenarios, as a lack of understanding of how parameter configuration affects inference speed and accuracy can result in suboptimal performance. To address this issue, this paper conducted an investigation of parameter configuration not only in the BFV scheme but also in the homomorphic encrypted neural network model itself. The aim was to reveal the relationships between parameter settings and computation efficiency and inference accuracy. The results of the experiments showed that the polynomial modulus was the most important parameter influencing computation efficiency and accuracy. Larger polynomial moduli resulted in greater accuracy, while smaller polynomial moduli resulted in a significant loss of accuracy. Other parameters in the BFV scheme had no effect on inference accuracy and only affected computation efficiency. Additionally, the polynomial modulus was found to be strongly related to both space and time complexities. As the polynomial modulus increased, both time and space complexity increased, with time complexity increasing much faster than the polynomial modulus. The ciphertext was nearly ten times larger than the plaintext, indicating a significant increase in storage requirements when using homomorphic encryption. The experiments also revealed that pooling and convolutional layers were the most time-consuming layers in homomorphic encrypted neural networks due to the large volume of ciphertext computation. To mitigate this issue, the paper advised reducing the use of pooling layers or replacing them with other network structures. Moreover, increasing the size of the convolution kernel was found to reduce the time-complexity of inference computation. In summary, the investigation of parameter configuration in homomorphic encrypted neural networks is critical for achieving optimal performance in various application domains. The findings of this paper provide valuable insights into the impact of different parameters on computation efficiency and inference accuracy, allowing for more informed decisions regarding parameter selection and network design. As homomorphic encryption continues to widely applied in privacy-preserving machine learning, further research is needed to explore additional aspects of parameter configuration and their impact on performance.
|Key words: homomorphic encryption|neural networks|parameter settings