Green AI-Based Cybersecurity Model for IoT System

In this manuscript, the Green AI paradigm is employed to develop an Intrusion Detection System (IDS) using Deep Learning. The aim is to overcome the complexity and generalization issues in IoT cyberattack detection solutions by automatically generating Convolution Neural Network (CNN) models and optimizing their settings to ensure their effectiveness on different datasets. The best model is then converted to a sparse model, achieving the Green AI. To generate the CNN models, five activation functions, one normalization, and four regularization parameters are optimized using a three-level tree. The tree of models starts with the root representing the activation function, followed by the normalization level, and then the regularization level. Each time the tree produces a new model structure by changing a node value in activation, normalization, and regularization levels with different combinations, resulting in several branches in the tree. For each tree model, the kernel and filter values are optimized using an exhaustive search to select the best combination among three kernels and three filter values. The findings demonstrate that the proposed method is efficient in generating a model that achieves an accuracy of 95.71 for the MQTT-IoT-IDS2020 dataset and generalizes well on UNSW-NB-2015 and KDD IoT-2017 datasets with 95.87 and 95.91 respectively. Therefore, automatically generating models is a preferred technique for researchers to ensure generalization on different datasets.