A Hybrid Lightweight Deep Learning based Intrusion Detection Approach in IoT Utilizing Feature Selection & Explainable Artificial Intelligence

Due to the resource constraints of IoT devices, standard cryptographic-based intrusion detection system techniques are ineffective in IoT environments. This paper introduces DL IID, a lightweight deep learning-based IoT intrusion detection model that not only improves security but also addresses computational challenges. The model, based on deep neural networks (DNN) layers along with bidirectional long-short-term memory (BiLSTM) modules, called DLB, enables symmetric and bidirectional feature removal from complex RF signal data, thereby enhancing the accuracy of detection by capturing complex patterns of legitimate and malicious IoT devices. The model employs an evolutionary genetic algorithm (GA) for feature selection to optimize memory and computational efficiency. A key aspect of this work is the use of Explainable AI (XAI) methodology, particularly Local Interpretable Model-Agnostic Explanations (LIME), which ensures transparency in decision-making and increases trust in the model predictions. Furthermore, dynamic quantization is applied after training to reduce model size while preserving detection accuracy, making it a good choice for resource-constrained IoT environments. The model is evaluated on an RF fingerprinting database comprising more than 450 IoT devices with signal changes in frequency, amplitude, and phase to mimic real-world scenarios with non-idealities. The test results demonstrate that DL-IID is superior to traditional machine learning models and modern deep-learning algorithms. It achieves an accuracy of 99.84%, a precision of 100.0%, a recall of 99.69%, and an F1-score of 99.84% while reducing the model in size down to 108.42 KB. The model is further evaluated using three benchmark datasets: CICIDS2017, CICIoMT2024, and UNSW-NB15.