关键词: 双分支, 特征融合, 多分支卷积, 注意力机制, 下采样

Abstract: In the homology analysis of malicious code, a large number of malicious code variants are generated due to techniques such as encryption, obfuscation, and packing, which leads to the problem that the deep learning model has insufficient ability to extract the features of malicious code. To solve this problem, a multibranch convolution and transformernet (MCATNet) homology analysis model based on feature fusion was proposed. Firstly, an MCATNet dualbranch network was constructed, one branch was a multibranch convolutional MBC (Multibranch convolution) module, and the MBC module was used to construct the CNN branch, and the CBAM hybrid attention mechanism was introduced to make the network pay more attention to the core features while taking into account the local features. Another branch is the Transformer module with ViT as the backbone, which extracts global feature information of malicious code images and proposes a downsampling module to finely preserve global features while aligning the feature maps of Transformer and CNN at the spatial scale. Secondly, the cascading strategy is used to fuse the local features of the CNN branch and the global features of the Transformer branch to solve the problem that the network only focuses on a single feature. Finally, the Softmax classifier was used to analyze the homology of the malicious code family. Experimental results show that the classification accuracy of the twobranch model based on feature fusion reaches 99.24%, which is 0.11% and 0.65% higher than that of the singlebranch CNN and singlebranch Transformer models, respectively.

Key words: doublebranched, feature fusion, multibranch convolution, attention mechanism, downsampling