Pseudo-Image Representation for CNN-Based Intrusion Detection

Abstract

— Convolutional Neural Networks (CNNs) have demonstrated strong potential for Intrusion Detection Systems (IDSs); however, their performance is highly dependent on input data representation. This paper presents a lightweight pseudo-image transformation framework that converts network traffic into optimized two-dimensional layouts for CNN processing, focusing on data representation rather than architectural modifications. Evaluated on NSL-KDD and CIC-IDS2017 datasets, square and near-square formats are shown to enhance spatial feature extraction. In particular, an 11×11 base representation, upscaled to 64×64 using bicubic interpolation, achieves a precision of 99.2%, accelerates convergence, and remains robust under limited training data conditions. By avoiding complex hybrid architectures, the proposed approach reduces computational overhead while maintaining competitive performance, making it suitable for real-time, large-scale deployment. These results suggest that strategic data representation, rather than increased model complexity, can significantly improve CNN-based IDS performance.