DAE-IHOG: An Improved Method for Classification Malware
Zhen Wang, Hongjiao Li · 2023
Among grayscale image-based malware classification methods, the histogram of gradients (HOG) is a mainstream image feature descriptor. HOG considers the gradient magnitude and gradient direction of individual pixels in feature processing, which makes it possible to solve the problem of malware variants. However, HOG ignores the difference in the degree of contribution of different parts of the malware image to the classification, resulting in low classification accuracy. Meanwhile, the high dimensionality of HOG features can make the computational overhead of the classification model high. Thus, this paper proposes an improved malware classification method, named deep autoencoder and improved HOG (DAE-IHOG), which obtains improved HOG features (IHOG) by using statistical feature information of blocks of different sizes and introduces a deep autoencoder to downscale the high-dimensional IHOG features. Finally, a soft voting strategy is used to further aggregate the output of the ensemble model to obtain the final classification accuracy. Experimental results show that compared with the HOG-based method, the classification accuracy of the proposed method is improved by 5.12% and the computation time is reduced by 93.8%.