AI-ENABLED LIGHTWEIGHT INTRUSION DETECTION FOR SMART WIRELESS SENSOR NETWORKS: A RESOURCE-EFFICIENT DEEP LEARNING APPROACH

Abstract

ABSTRACT
Wireless Sensor Networks (WSNs) constitute critical smart infrastructure for sustainable development, enabling applications from precision agriculture to smart city monitoring. However, their deployment faces significant security challenges compounded by severe resource constraints, limited energy, bandwidth, and computational capacity. Existing intrusion detection systems (IDS) prioritize detection accuracy while neglecting resource efficiency, rendering them impractical for battery-powered sensor nodes. Additionally, supervised learning approaches limit detection to known attack signatures, failing against zero-day exploits. This research addresses these critical limitations by developing a Lightweight Intelligent Intrusion Detection System (LIIDS) using unsupervised deep autoencoders optimized for resource-constrained environments. The bottleneck architecture achieves exceptional 86.89% data compression (approximately 7.6× reduction), directly translating to proportional bandwidth and energy savings. Despite aggressive compression, the system maintains superior detection performance: 99.76% accuracy on NSL-KDD and 98.52% on UNSW-NB15 datasets, with false alarm rates reduced by 78% compared to existing methods (1.27% vs. 5.82%). The unsupervised learning paradigm enables detection of both known and zero-day attacks without labeled training data. These results demonstrate that comprehensive WSN security can be achieved without excessive resource consumption, enabling practical deployment in Nigeria's emerging smart infrastructure for sustainable economic growth.