Experimental study demonstrates efficient multi-class intrusion detection in IoT networks using a lightweight CNN-BiLSTM architecture, highlighting feasibility for edge deployment.
Key Points
To develop and evaluate a lightweight hybrid deep learning model combining 1D-CNN and BiLSTM architectures that achieves high multi-class intrusion detection accuracy while remaining computationally feasible for IoT edge devices.
Designed a hybrid architecture utilizing a one-dimensional convolutional neural network (1D-CNN) for spatial feature extraction coupled with a bidirectional long short-term memory (BiLSTM) network for temporal pattern recognition.
Preprocessed the CICIoT2023 benchmark dataset—comprising 33 attack variants across seven categories—using class balancing, feature normalization, and feature reduction techniques.
Evaluated detection quality, parameter count, and CPU inference latency against standalone CNN, standalone BiLSTM, and classical machine learning baselines.
The hybrid 1D-CNN–BiLSTM architecture effectively identifies diverse attack categories, including DDoS, Mirai botnet, spoofing, reconnaissance, and web attacks.
Feature reduction and balanced architectural design lower overall parameter counts and CPU inference latency compared to full-scale deep learning models, demonstrating deployability on resource-constrained edge hardware.