Usman Opeyemi Lateef, Muniyandi Ravie Chandren, Omar Khairuddin, Mohamad Mazlyfarina, Owoade Ayoade Akeem, Kareem Morufat Adebola
Research Centre for Cyber Security, Faculty of Information Science and Technology, Universiti Kebangsaan Malaysia (UKM), 43600, Bangi, Selangor, Malaysia.
Department of Computer Science, Tai Solarin University of Education, P.M.B. Ijagun, Ijebu-Ode, 2118, Ogun, Nigeria.
Brain Inform. 2025 Apr 30;12(1):11. doi: 10.1186/s40708-025-00256-z.
Recent advancements in cloud-based machine learning (ML) now allow for the rapid and remote identification of neural-biomarkers associated with common neuro-developmental disorders from neuroimaging datasets. Due to the sensitive nature of these datasets, secure deep learning (DL) algorithms are essential. Although, fully homomorphic encryption (FHE)-based methods have been proposed to maintain data confidentiality and privacy, however, existing FHE deep convolutional neural network (CNN) models still face some issues such as low accuracy, high encryption/decryption latency, energy inefficiency, long feature extraction times, and significant cipher-image expansion. To address these issues, this study introduces the HoRNS-CNN model, which integrates the energy-efficient features of the residue number system FHE scheme (RNS-FHE scheme) with the high accuracy of pre-trained deep CNN models in the cloud for efficient, privacy-preserving predictions and provide some proofs of its energy efficiency and homomorphism. The RNS-FHE scheme's FPGA implementation includes embedded RNS pixel-bitstream homomorphic encoder/decoder circuits for encrypting 8-bit grayscale pixels, with cloud CNN models performing remote classification on the encrypted images. In the HoRNS-CNN architecture, the ReLU activation functions of deep CNNs were initially trained for stability and later adapted for homomorphic computations using a Taylor polynomial approximation of degree 3 and batch normalization to achieve high accuracy. The findings show that the HoRNS-CNN model effectively manages cipher-image expansion with an asymptotic complexity of , offering better performance and faster feature extraction compared to its peers. The model can predict 400,000 neural-biomarker features in one hour, providing an effective tool for analyzing neuroimages while ensuring privacy and security.
基于云的机器学习（ML）的最新进展现在允许从神经影像数据集中快速远程识别与常见神经发育障碍相关的神经生物标志物。由于这些数据集的敏感性，安全的深度学习（DL）算法至关重要。虽然已经提出了基于全同态加密（FHE）的方法来维护数据机密性和隐私性，但是，现有的FHE深度卷积神经网络（CNN）模型仍然面临一些问题，例如准确率低、加密/解密延迟高、能源效率低、特征提取时间长以及密文图像显著扩展。为了解决这些问题，本研究引入了HoRNS-CNN模型，该模型将余数系统FHE方案（RNS-FHE方案）的节能特性与预训练的深度CNN模型在云端的高精度相结合，以实现高效、隐私保护的预测，并提供了其能源效率和同态性的一些证明。RNS-FHE方案的FPGA实现包括用于加密8位灰度像素的嵌入式RNS像素位流同态编码器/解码器电路，云端CNN模型对加密图像进行远程分类。在HoRNS-CNN架构中，深度CNN的ReLU激活函数最初经过训练以提高稳定性，随后使用3次泰勒多项式近似和批量归一化进行同态计算调整，以实现高精度。研究结果表明，HoRNS-CNN模型有效地管理了密文图像扩展，其渐近复杂度为 ，与同类模型相比具有更好的性能和更快的特征提取速度。该模型可以在一小时内预测40万个神经生物标志物特征，为分析神经图像提供了一个有效的工具，同时确保了隐私和安全。
