Comparative Analysis of Input Image Characteristics in Convolutional Neural Network-based Signature Detection

Abstract

The detection of malware represents a primary concern in contemporary computer security and is therefore imperative for the protection of systems and data integrity. This research presents an innovative approach to comparing diverse input image formats with the objective of identifying the optimal methodology for detecting specific malware-related signatures using convolutional neural networks (CNN), which have been specifically developed by the authors for this purpose. Subsequently, machine code instructions are generated and then converted into four distinct image format options. The four image formats, namely 1xN fixed, 1xN scalable, NxN fixed, and NxN scalable, are subsequently employed for the training of the CNN. The study assesses the formats in question in terms of training time, accuracy, and computational complexity. The results demonstrate that the NxN scalable format exhibits the highest accuracy with accelerated training times in comparison to other formats. Furthermore, the scalable format necessitates only 25% of the original pixel count for a 96% classification success rate. The utilization of the NxN scalable format for machine code instruction representation results in enhanced accuracy, accelerated training, and a considerable reduction in pixel usage, indicating a promising avenue for optimizing the efficiency of malware detection.