The dramatic increase in the number of IoT devices has dramatically increased the attack surface, and thus it needs intrusion detection systems (IDS) with high-performance and the ability to detect potential threats in real-time with limited resources. This research is aimed primarily to offer a new methodology to apply the artifivial intelligence to place a new format upon intelligent intrusion detection. To solve this, the current paper comes up with a new deep intelligent IDS architecture which combines a lightweight Xception neural network with a Modified Pelican Optimization Algorithm (MPOA) in the tuning of hyperparameters and designing of architectures. The MPOA improves convergence, reduces the local optima entrapment and adapts the Xception model to meet the traffic characteristics in the IoT. The proposed framework, evaluated on three benchmark sets (IoT-23, UNSW-NB15, and CICIDS2017) has average F1-scores of 97.8, 96.1 and 97.7, respectively which consistently outperform state of art baselines such as the Random Forest (F1: 89.5%), SVM (F1: 88.2%), LSTM (F1: 93.1%), CNN (F1: 94.5%) and Autoencoder. It is also worth noting that the model has better identification of IoT-specific attacks (e.g., 98.9% F1 with Brute Force on the IoT-23) with low false positive rates (≤ 2.1%) and low computational performance, making it an appropriate model in terms of real time application in IoT contexts that are limited in resources. The study, therefore, provides the effectiveness of synergizing depthwise-separable CNNs with bio-inspired optimization to provide scalable, accurate, and adaptive IoT security.
Pu et al. (Fri,) studied this question.