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

Abstract

Due to the resource constraints of IoT devices, standard cryptographic-based intrusion detection techniques are ineffective in real-world IoT environments. This study presents DL-IID, a lightweight deep learning-based framework for IoT intrusion detection. The model integrates deep neural networks (DNN) with bidirectional long short-term memory (Bi-LSTM) to detect complicated temporal and nonlinear attack patterns. A wrapper-based genetic algorithm (GA) is employed for feature selection, eliminating redundant features and further reducing memory consumption. In addition, dynamic quantization after training reduces the model size to 108.42 KB while retaining high accuracy, which is within the reach of limited IoT nodes. A key aspect of this study is the application of XAI methodology, specifically Local Interpretable Model-Agnostic Explanations (LIME), which ensures transparency in decision-making and enhances trust in the model predictions. The proposed method is compared with four standard datasets: RF fingerprinting (450 IoT devices), CICIDS2017, CICIoMT2024, and UNSW-NB15. Experimental results demonstrate improved performance with 99.84% accuracy, 100% precision, 99.69% recall, and 99.84% F1-score, while significantly reducing model size and computational overhead. Experimental results confirm that DL-IID is a viable and practical solution for intrusion detection in next-generation IoT environments.