Data privacy and security are the primary concerns in today’s era, malware attacks are the prime threats. Though the security systems are strengthened with modern and efficient techniques, they can handle attacks from known malware families. However, the development of new malware is progressing in an exponential curve, keeping the demand for a more efficient algorithm to secure the personal, organizational, and other modes of data. Recently, the malware executable files have been visualized as gray-scale images and transformed into an image classification problem. Literature reports that Deep Learning (DL) algorithms, especially the Convolutional Neural Network (CNN) have been achieving significant accuracy in malware image classification. However, a substantial limitation in Deep Learning algorithms is that they expect the input malware image size should be uniform, which is not possible in practice. Hence, the malware images are resized to fit into uniform dimensions and then fed to the algorithms. The dimension of the malware image is closely related to the lines of code and the block, resizing them to a uniform size would affect the meaning of the binary files. A novel malware image representation based on the Gray-Level Co-occurrence Matrix (GLCM) is proposed here to maintain a stable dimension for all the malware images, hence the image resizing step could be ignored. The sparse nature of the GLCM matrix offers the space for constructing a novel dynamic stride-offset matrix which in turn accelerates the training phase of the CNN classifier. The performance of the proposed Sparse CNN classifier having the GLCM-based images (SCNN + GLCM) as input is investigated with the malware images from three benchmark datasets: MalImg, Malevis, and Microsoft Malware Detection (MMD) dataset. With the set of optimized hyperparameters, the proposed SCNN + GLCM malware image classification is analyzed on various metrics to state its significant performance.
原文链接
分享
引用
