Research On Malicious Code Classification Based On One-stage Target Detection Network

With the development of Internet technology and the popularity of computers,computer technology is applied to various industries of the country,and computers are closely related to people’s lives.The Internet brings great convenience to people’s lives,while the emergence of malicious software also brings information security problems,and even direct economic losses to enterprises and the country.Therefore,finding an effective malicious code detection method is an urgent issue.Since traditional malicious code detection methods rely heavily on manual analysis of static files,which consumes a lot of time and is extremely inefficient.With the development of computer vision and the hot artificial intelligence technology,researchers have been exploring the application of image processing technology and deep learning algorithms to malicious code detection,which is a very important research direction at present.This paper uses computer vision technology to convert malicious code sample files into grayscale images,constructs a grayscale image dataset of malicious code based on VOC format,and proposes an improved migration YOLOv4 network for malicious code detection method,which extracts grayscale image texture features through backbone feature network and feature pyramid network,and combines with prediction network to classify and locate malicious code.Meanwhile,the original CSPDar Net53 network is replaced by the lightweight Mobile Net V3 network as the backbone network,which reduces the overall computation of the model and the overall computation time of the model while maintaining the detection accuracy;migration learning is used to migrate the model weights pre-trained on large datasets to the improved malicious code classification model to speed up the model convergence;the Focal loss function is used to improve the loss function and solve the problem of positive and negative sample imbalance during model training to improve the model detection accuracy.The experimental results show that the improved model proposed in this paper is better than the traditional malicious code detection method in terms of accuracy,false alarm rate and convergence speed,and the accuracy of malicious code classification reaches 94.3%.