Image Based Malware Classification with Multimodal Deep Learning

Today, there are many different methods for analyzing and detecting malware. Some of these methods are basically based on statistical analysis, some on static and dynamic analysis methods, and some on machine learning methods. The studies carried out to classify malware with statistical machine learning-based analysis methods are generally based on complex and challenging feature extraction methods, and manual feature extraction is a very tedious process. However, the capability of deep learning methods to automatically extract complex features in a way simplifies this arduous process. In this study, a novel multimodal convolutional neural network-based deep learning architecture and singular value decomposition-based image feature extraction method are proposed to classify malware files using intermediate-level feature fusion. In addition to this, the performances of classical machine learning algorithms,  neural networks, and the proposed multimodal convolutional neural networks-based deep learning algorithm are compared, and their performance is revealed. The performance of the proposed algorithm was also compared with the results of studies conducted with the same data set in the literature. The experimental results concluded that the proposed method is more successful than other methods or showed the same performance even though it did not use manual feature extraction techniques. It has been observed that with architecture, intermediate fusion approaches have the ability to obtain more specific features more effectively than other methods, thus improving performance values more than other methods.