Multi-Constraint Evaluation of Edge-Deployable Intrusion Detection Architectures for Smart City IoT

Abstract

The integration of IoT devices into smart cities introduces critical vulnerabilities, requiring Intrusion Detection Systems (IDSs) deployable on resource-constrained edge devices. This paper evaluates LightGBM, a Standard Deep Neural Network (SDNN), and a Multi-Task Supervised Autoencoder (MSAE) using a 175,341-sample Smart City dataset to explicitly determine their quality-efficiency-security trade-offs. Pareto Multi-Criteria Decision Making (MCDM) was applied across five constraints: AUC-ROC, False Positive Rate (FPR), False Negative Rate (FNR), Inference Latency, and Memory Consumption. Results show the MSAE achieved the highest sensitivity (98.57% Recall, 1.43% FNR) but suffered high memory overhead (6.34 KB) and FPR (16.02%). LightGBM achieved the highest static baseline score (6.63) via memory compression (2.50 KB) but lacked real-time inference speed (0.0064 ms). The SDNN emerged as the most dynamically robust architecture, dominating a sensitivity analysis of 120 constraint-weighting scenarios by winning 83 Rank-1 placements compared to LightGBM’s 37 and MSAE’s 0. We conclude the SDNN is the statistically optimal architectural choice for balancing edge constraints, while the MSAE suits critical infrastructure where threat sensitivity outweighs alert fatigue.