Enhanced Channel-Wise Homomorphic Encryption for Image Inference based on Pairwise Activation Functions
Tanuja Tanuja, Rakesh Kumar · 2023
To preserve the data privacy of deep learning models, Homomorphic Encryption(HE) is used as a promising method which has gained attention in privacy-preserving deep learning(PPDL). While PPDL algorithms based on HE has been introduced and tested, but there is a need to improve the accuracy and latency of PPDL models for practical applications. To increase accuracy, the existing method uses batch-normalization utilizing the Cofficients Merging(CM) tool, Channel-wise homomorphic encryption(CHE), and square Activation Function(AF). To improve accuracy, it did not emphasize on more AFs. Literature has explored other AFs for better performance called Pairwise AFs (Combinations of some standard AFs). Proposed work, seeks to enhance the performance of CHE-based image inference by using pairwise AFs in Convolution-Activation-BN(CAB) and Convolution-BN-Activation(CBA) schemes. Mathematically derived equations of CAB and CBA schemes shows that proposed method could significantly improve the accuracy.