Visual Feature Mapping and Feature Selection for Robust IDS Based on CSE-CIC-IDS2018 Dataset

Abstract

The rapid growth of Internet of Things (IoT) ecosystems posed a great risk to ubiquitous access to advanced cyber threats and thus the design of effective Intrusion Detection Systems (IDS) has become critical in ensuring the safeguard of secure network structures. Nonetheless, most classic deep learning methods, especially Recurrent Neural Networks (RNNs) tend to exhibit low generalization and low interpretability to tabular network traffic data and may fail to discriminate between different and intricate attacks. To overcome this issue, this paper presents a visual feature-based IDS model, which converts network traffic characteristics in the form of tabular features into RGB matrix images to harness the power of effective computer vision algorithms to detect intrusions. Recursive Feature Elimination (RFE) is applied to determine the best informative feature before generating an image, enhance features and decrease redundancy. The acquired visual representations are subsequently categorized by the EfficientNetB0 design by way of transfer learning. Experimental testing shows that the suggested method has a 98 percent accuracy, precision, recall, and F1-score, Matthews Correlation Coefficient (MCC) of 0.97, and AUC-ROC of 0.99. Comparatively, the RNN original does not reach the same level as it only gives 89% on performance metrics with MCC of 0.88 and AUC-ROC of 0.88. The above findings indicate the ability of feature selection combined with visual deep learning to facilitate scalable, precise, and real-time intrusion detection of modern-day IoT environments.