ENNigma: A framework for Private Neural Networks

作     者：Barbosa, Pedro Amorim, Ivone Maia, Eva Praca, Isabel 

作者机构：Polytech Porto ISEP IPP Porto Sch Engn Res Grp Intelligent Engn & Comp Adv Innovat & Dev P-4200072 Porto Portugal Polytech Inst Porto IPP PORTIC Porto Res Technol & Innovat Ctr P-4200374 Porto Portugal 

出 版 物：《FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE》 (Future Gener Comput Syst)

年 卷 期：2025年第166卷

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Norte Portugal Regional Operational Programme (NORTE 2020)   under the PORTUGAL 2020 Partnership Agreement  through the European Regional Development Fund (ERDF) [NORTE-01-0145-FEDER-000044  UIDB/00760/2020] 

主　　题：Homomorphic encryption Private Neural Networks Encrypted neural networks Fast fully homomorphic encryption over the torus Distributed denial of service detection 

摘      要：The widespread use of the Internet and digital services has significantly increased data collection and processing. Critical domains like healthcare rely on this data, but privacy and security concerns limit its usability, constraining the performance of intelligent systems, particularly those leveraging Neural Networks (NNs). NNs require high-quality data for optimal performance, but existing privacy-preserving methods, such as Federated Learning and Differential Privacy, often degrade model accuracy. While Homomorphic Encryption (HE) has emerged as a promising alternative, existing HE-based methods face challenges in computational efficiency and scalability, limiting their real-world application. To address these issues, we introduce ENNigma, a novel framework employing state-of-the-art Fully Homomorphic Encryption techniques. This framework introduces optimizations that significantly improve the speed and accuracy of encrypted NN operations. Experiments conducted using the CIC-DDoS2019 dataset - a benchmark for Distributed Denial of Service attack detection - demonstrate ENNigma s effectiveness. A classification performance with a maximum relative error of 1.01% was achieved compared to non-private models, while reducing multiplication time by up to 59% compared to existing FHE-based approaches. These results highlight ENNigma s potential for practical, privacy-preserving neural network applications.