Privacy-Preserving Convolutional Neural Networks with CKKS-RNS for Encrypted Image Classification

Abstract

The rise of security concerns in cloud-shared infrastructures has introduced significant challenges for maintaining privacy in data processing. Although standard encryption methods provide robust protection for data at rest and during transmission, vulnerabilities arise when data must be decrypted for processing, exposing sensitive raw information to potential privacy risks. This issue is particularly pronounced in sectors governed by stringent regulatory requirements, such as healthcare, genomics, smart government, and finance, among many others, where protecting confidential data is critical. Homomorphic Encryption (HE) cryptosystems are solutions to address privacy concerns by providing encrypted data computations. HE allows a non-trustworthy third-party resource to process encrypted information without disclosure. However, the main challenge toward deploying lattice-based HE schemes in Convolutional Neural Network (CNN) models lies in overcoming the high computational costs associated with these cryptosystems. Efficient cryptographically compatible methods become imperative for designing a privacy-preserving CNN with HE (CNN-HE). This paper proposes a method to improve the performance of CNN-HE using the Residual Number System (RNS)-based Cheon-Kim-Kim Song (CKKS) HE scheme, which enables approximate arithmetic over encrypted real numbers. The CNN-HE with CKKS-RNS enables encrypted inputs to be decomposed into several parts and propagated homomorphically and independently in parallel across the model. The RNS representation enables parallel processing in our models, significantly reducing processing time. Experimental analysis on the MNIST optical character recognition benchmark dataset demonstrates that the proposed CNN-HE-RNS models reduce classification latency concerning state-of-the-art CNN-HE solutions without compromising security and accuracy.