LAHENet : A Lightweight Additive Homomorphic Edge Neural Network Framework for Industrial IoT

Gong, Mowei and Li, Zhe and Lu, Xuepeng and Gong, Bei and Zhu, Haotian and Meng, Weizhi (2025) LAHENet : A Lightweight Additive Homomorphic Edge Neural Network Framework for Industrial IoT. IEEE Transactions on Dependable and Secure Computing. ISSN 1545-5971

Text (TDSC-1274764960)
TDSC-1274764960.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (5MB)

Abstract

Edge nodes in the Industrial Internet of Things (IIoT) often face a fundamental trade-off between limited computational resources and stringent real-time inference requirements. Moreover, sensitive data they generated are exposed to significant privacy and security threats during transmission and computation. To address these challenges, this paper proposes a lightweight additive homomorphic edge neural network framework called LAHENet. This framework achieves millisecond-level inference latency in real-world industrial environments through a combination of a dual-metric feature selection strategy, an efficient additive homomorphic signcryption protocol, and a lightweight linear computation layer with adaptive layer collapsing. It ensures end-to-end confidentiality, unforgeability, forward security, and verifiable computation correctness. Experimental results show that LAHENet maintains a constant communication overhead at a few kilobytes per inference while preserving high model accuracy. It significantly enhances inference efficiency and reduces bandwidth consumption in edge environments, offering a practical private inference solution for large-scale IIoT deployments.