Privacy-Preserving Classification Method for Neural-Biomarkers using Homomorphic Residue Number System CNN: HoRNS-CNN

The increasing generation of MRI dataset and the recent cloud deployment of deep learning (DL) algorithms have enabled timely remote classifications of discrepancies in neural-biomarkers of critical health conditions such as dyslexia. Using these untrusted platforms to implement a secure DL algorithm will identify and resolve potential security attacks or patient data theft, hence, the need for a privacy-preserving method. However, existing homomorphic (FHE) privacy-preserving methods are still inefficient in terms of accuracy, latency, and feature extraction time with significantly large cipher-image expansion problem. This study proposes homomorphic residue number system-CNN (HoRNS-CNN) model for the privacy-preserving classification of dyslexia neural-biomarkers. The HoRNS-CNN architecture is composed of the RNS-FHE scheme and pre-trained CNN models. The RNS-FHE scheme was used to design an encryption module for each pixel in the dataset, while pre-trained CNNs were applied directly to the encrypted data in the cloud after adapting their activation layers to homomorphic computations. Results from proposed HoRNS-CNN model demonstrated an improved performance.