A Multi-Modal Rgb Encoding Framework for Vision Driven Network Intrusion Detection Via Optimized Variable-To-Pixel Mapping and Hybrid Dimensionality Reduction

Abstract

Recent advances in computer vision–driven deep learning have enabled highly accurate classification of two dimensional (2D) images, surpassing both traditional machine learning approaches and even human-level performance in certain tasks. These advanced vision-based models have increasingly been explored in the field of network intrusion detection, where experimental studies have demonstrated both their potential and inherent challenges. To address these challenges, this study introduces an enhanced approach for analyzing NIDS datasets using vision-based deep learning techniques.The approach converts network data into two-dimensional (2D) visual formats using different image transformation methods, which are then combined to form three channel RGB images. This representation significantly improves the effectiveness of deep learning-based intrusion detection models.This work adopts the variants of CNN for intrusion detection to enhance the model quality. The proposed vision-based intrusion detection framework demonstrates significant performance improvements across benchmark datasets. On the ISCXIDS2012 dataset, the model achieves an overall accuracy of 99.12%, precision of 98.87%, recall of 98.95%, and F1-score of 98.91%, outperforming the conventional non-image approach by approximately 4.8% in accuracy and 6.2% in F1-score. Similarly, on the CSE-CIC IDS2018 dataset, the proposed method attains an accuracy of 98.76%, recall of 98.41%, precision of 98.53%, and F1-score of 98.47%, showing an improvement of 5.1% over single-channel image representations. The multi-channel RGB encoding contributes to an average performance gain of 3.5–6.0% across all evaluation metrics compared to grayscale-based models. Additionally, the false positive rate is reduced to below 1.8%, ensuring reliable classification of benign traffic. The ROC analysis further confirms robustness, with an average AUC score exceeding 0.995 across all classes.