A Multiscale Deep Convolutional Malicious Code Classification Method Based on Structural Reparameterization

As malicious code countermeasures evolve, attackers have responded by generating numerous malicious code variants through shelling, code obfuscation, and similar strategies. Addressing the shortcomings of current deep learning-based malicious code classification methods-namely, their limited extraction accuracy and efficiency-this study offers a novel approach that integrates CNNs and Transformer for malicious code detection. This method leverages deep neural networks as its foundation and incorporates a unique fusion module for structural reparameterization. This innovative mechanism reduces memory access costs by eliminating jump connections within the network. Furthermore, the implementation of large kernel convolution techniques enhances network accuracy. Experimental results demonstrate that our proposed method consistently outperforms the latest malicious code detection techniques in terms of both accuracy and operational efficiency.