Lightweight Intrusion Detection System for IoT Environment through Compression Techniques

Internet of Things has been widely adopted technology and making secure IoT systems is a necessity. Machine-and deep-learning-based intrusion detection systems have been extensively used to detect anomalies in network traffic. However, deploying such models in resource-constrained environments is cumbersome since these models consume huge memory space. Compression Techniques are implemented on such models to make them suitable for deployment in the IoT ecosystem. In this paper, we present a hybrid CNN-BiLSTM based model for classifying network attacks using the CICIoT2023 dataset. For feature selection and reduction, we utilized Permutation Feature extraction and Principle Component Analysis. To make our model lightweight, we implemented three compression techniques- Knowledge Distillation, weight pruning, and post-training quantization and compared the model performance with all three techniques. We achieved the most reduced model size via Knowledge distillation, reducing the original size by $\mathbf{9 4. 7 6 \%}$, followed by quantization, reducing size by $\mathbf{8 9. 5 2 \%}$. Post-training quantization showed a high accuracy of $\mathbf{9 1. 5 0 \%}$. Through this work, we explore the optimal compression technique to be adopted while designing lightweight Deep learning-based IDS solutions for IoT networks.