@article{Ullah2026, 
author = {Saeed Ullah and Junsheng Wu and Mian Muhammad Kamal and Mohammed K. Alzaylaee and Heba G. Mohamed},
title = {Quantized Intrusion Detection for Resource-Constrained IoT: A Comparative Evaluation of Efficiency and Adversarial Robustness},
year = {2026},
journal = {Computers, Materials & Continua},
volume = {88},
number = {3},
pages = {62},
keywords = {IoT security, intrusion detection, neural network quantization, adversarial robustness, energy-efficient computing, TensorFlow lite},
url = {https://www.sciopen.com/article/10.32604/cmc.2026.084409},
doi = {10.32604/cmc.2026.084409},
abstract = {The proliferation of Internet of Things (IoT) devices has introduced unprecedented security challenges, necessitating efficient intrusion detection systems (IDS) capable of operating under severe resource constraints. This research presents a hardware-informed empirical study of quantized neural-network-based intrusion detection for resource-constrained IoT platforms, using an ARM Cortex-M4 deployment target as a reference. We evaluate FP32, FP16, and INT8 TensorFlow Lite model variants derived from a lightweight 1D-CNN and assess their trade-offs in clean-data accuracy, model size, estimated inference latency, estimated energy consumption, and adversarial robustness. INT8-quantized model achieves 99.10% accuracy on clean data while maintaining 97.50% adversarial accuracy under Projected Gradient Descent (PGD) attacks with perturbation budget = 0.3. The quantized model achieves 12.0× latency reduction (0.083 vs. 0.995 ms) and 92.7% energy reduction (0.0083 vs. 0.1135 mJ) when compared to FP32. The memory footprint of the model is reduced by 55.7% from 58.02 to 25.72 KB. Our comprehensive analysis includes confusion matrices, ROC curves (AUC = 0.9964 for INT8), adversarial robustness heatmaps, and statistical significance testing via McNemar’s test. The results establish INT8 quantization as a viable solution for deploying robust IDS on resource-constrained IoT devices, achieving practical deployment feasibility without reducing detection performance.}
}