Privacy-Preserved Machine Learning for Secure Neural Networks

In this article, we explore the challenges of efficient inference in secure multi-party computation (SMPC) and propose a novel method to reduce the number of ciphertexts required for downsampling convolutional neural networks (ConvNets). Our approach decomposes ConvNets into smaller unit-stride convolution operations, which can be simply added to produce a dense ciphertext without the need for a densification step. This results in a significantly more efficient execution flow compared to previous work, with reduced computational complexity and improved accuracy.
The article begins by discussing the limitations of existing SMPC frameworks, which often result in high computational overhead due to the need for bootstrapping ciphertexts. The authors then introduce their proposed method, which leverages a novel encoding scheme to decompose ConvNets into smaller unit-stride operations. This allows for a more efficient execution flow, with reduced computational complexity and improved accuracy compared to previous work.
To illustrate the effectiveness of their proposal, the authors present experimental results on several benchmark datasets. The results show that their method achieves significant speedups in execution time while maintaining comparable accuracy to state-of-the-art SMPC frameworks. Additionally, the authors provide a detailed analysis of the encoding scheme used in their method, which demonstrates its efficiency and effectiveness in reducing computational complexity.
In summary, this article presents a novel approach to efficient inference in secure multi-party computation by decomposing convolutional neural networks into smaller unit-stride operations. The proposed method reduces computational complexity while maintaining accuracy, making it a valuable contribution to the field of SMPC.