A Lightweight Model for Malicious Code Classification Based on Structural Reparameterisation and Large Convolutional Kernels

Abstract

Abstract With the advancement of adversarial techniques for malicious code, malevolent attackers have propagated numerous malicious code variants through shell coding and code obfuscation. Addressing the current issues of insufficient accuracy and efficiency in malicious code classification methods based on deep learning, this paper introduces a detection strategy for malicious code, uniting Convolutional Neural Networks (CNNs) and Transformers. This approach utilizes deep neural architecture, incorporating a novel fusion module to reparametrize the structure, which mitigates memory access costs by eliminating residual connections within the network. Simultaneously, overparametrization during linear training time and significant kernel convolution techniques are employed to enhance network precision. In the data preprocessing stage, a pixel-based image size normalization algorithm and data augmentation techniques are utilized to remedy the loss of texture information in the malicious code image scaling process and class imbalance in the dataset, thereby enhancing essential feature expression and alleviating model overfitting. Empirical evidence substantiates this method has improved accuracy and the most recent malicious code detection technologies.