Energy-Efficient Privacy-Preserving Neural Networks for Resource-Constrained IoT Devices Using Secure Computation

Abstract

Artificial intelligence at the edge is increasingly embedded in consumer, industrial, medical, and civic infrastructures, yet many internet-of-things devices operate under severe limits in energy, memory, bandwidth, and thermal headroom. Running neural networks locally can reduce latency and communication, but it also exposes sensitive data to device compromise and raises questions about whether privacy can be preserved without sacrificing operational sustainability. This paper examines energy-efficient privacy-preserving neural networks for resource-constrained internet-of-things devices through the lens of secure computation. Rather than treating cryptography, machine learning optimization, and infrastructure governance as separate domains, the analysis frames them as coupled design choices that shape system architecture, cost, robustness, and fairness. The discussion considers secure multi-party computation, homomorphic encryption, secret sharing, differential privacy, and federated learning, with attention to their energy profiles and deployment feasibility in edge and gateway environments. It then analyzes structural trade-offs among privacy guarantees, inference accuracy, latency, battery life, and communication overhead. The paper argues that practical deployments require hybrid protocols, model compression, reduced activation complexity, secure offloading, and policy mechanisms that clarify accountability across device manufacturers, network operators, cloud providers, and public institutions. Case illustrations from smart health, industrial monitoring, and urban sensing show that privacy and energy efficiency are not merely technical constraints but also governance and sustainability challenges. The conclusion outlines a research agenda for auditable, inclusive, and resource-aware secure neural computation at the edge.