LMSA: A Lightweight Multi-Key Secure Aggregation Framework for Privacy-Preserving Healthcare AIoT

Hyunwoo Park^1,2, Jaedong Lee^1,3,*
1 Healthcare AI Team, National Cancer Center, Goyang, 10408, Republic of Korea
2 Cancer Big Data and AI Branch, National Cancer Center, Goyang, 10408, Republic of Korea
3 Department of Cancer AI & Digital Health, Graduate School of Cancer Science and Policy, National Cancer Center, Goyang, 10408, Republic of Korea
* Corresponding Author: Jaedong Lee. Email: email
(This article belongs to the Special Issue: Machine learning and Blockchain for AIoT: Robustness, Privacy, Trust and Security)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2025.061178

Received 18 November 2024; Accepted 06 February 2025; Published online 17 March 2025

Download PDF

Abstract

Integrating Artificial Intelligence of Things (AIoT) in healthcare offers transformative potential for real-time diagnostics and collaborative learning but presents critical challenges, including privacy preservation, computational efficiency, and regulatory compliance. Traditional approaches, such as differential privacy, homomorphic encryption, and secure multi-party computation, often fail to balance performance and privacy, rendering them unsuitable for resource-constrained healthcare AIoT environments. This paper introduces LMSA (Lightweight Multi-Key Secure Aggregation), a novel framework designed to address these challenges and enable efficient, secure federated learning across distributed healthcare institutions. LMSA incorporates three key innovations: (1) a lightweight multi-key management system leveraging Diffie-Hellman key exchange and SHA3-256 hashing, achieving O(n) complexity with AES (Advanced Encryption Standard)-256-level security; (2) a privacy-preserving aggregation protocol employing hardware-accelerated AES-CTR (CounTeR) encryption and modular arithmetic for secure model weight combination; and (3) a resource-optimized implementation utilizing AES-NI (New Instructions) instructions and efficient memory management for real-time operations on constrained devices. Experimental evaluations using the National Institutes of Health (NIH) Chest X-ray dataset demonstrate LMSA’s ability to train multi-label thoracic disease prediction models with Vision Transformer (ViT), ResNet-50, and MobileNet architectures across distributed healthcare institutions. Memory usage analysis confirmed minimal overhead, with ViT (327.30 MB), ResNet-50 (89.87 MB), and MobileNet (8.63 MB) maintaining stable encryption times across communication rounds. LMSA ensures robust security through hardware acceleration, enabling real-time diagnostics without compromising patient confidentiality or regulatory compliance. Future research aims to optimize LMSA for ultra-low-power devices and validate its scalability in heterogeneous, real-world environments. LMSA represents a foundational advancement for privacy-conscious healthcare AI applications, bridging the gap between privacy and performance.