Enhanced intrusion detection in wireless sensor networks using deep reinforcement learning with improved feature extraction and selection

In the Internet of Things, intrusion detection entails keeping an eye on device activity and network traffic in order to spot and address possible security lapses. Early threat detection, sensitive data protection, and cyberattack mitigation are some of its benefits. False positives, resource-intensive monitoring, and the difficulty of staying up to date with changing threats in the ever-changing IoT world are possible downsides. This study suggested a novel Bagging-DRL-based Intrusion Detection model, which comprises of four stages, to address these issues. (i) Gathering and pre-processing data; (ii) extracting features; (iii) selecting features; and (iv) utilizing deep reinforcement learning for intrusion detection. Initially, the CSE-CIC-IDS2018 and NSL-KDD databases are used to get the raw data. Z-Score normalization and data cleaning were used as preprocessing techniques for the gathered data. Features such as correlation, protocol-based features, higher-order statistical features, statistical features, and the newly proposed Enriched Principal Component Optimization with Self-Improved Seagull Algorithm (EPCO-SISA) are extracted from the pre-processed data. A novel Correlation-based Recursive Feature Elimination (C-RFE) method is used to choose the best features from the extracted features. Finally, Deep Reinforcement Learning is used to detect the incursion using the features that have been chosen. In order to improve detection accuracy, the DRL combines Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), and Optimized Recurrent Neural Networks (O-RNN). To do this, the weight function of the RNN is adjusted using the recently developed Self-Improved Seagull Optimization Algorithm (SI-SOA). The result is derived from Deep Reinforcement Learning's bagging value. The suggested model's performance is implemented using the MATLAB platform, and performance metrics including accuracy, precision, recall, and F1-score are used to evaluate the model's performance. The suggested model outperformed previous efforts with the maximum accuracy of 0.9836 and 0.9606 on the NSL-KDD and CSE-CIC-IDS2018 datasets, respectively.