Study of a Hybrid Approach Towards Malware Detection in Executable Files

With the ever-increasing number of Internet users in this digital age, exposure to malicious attacks is increasing. Every day, large volumes of malicious content are generated to exploit 0-day vulnerabilities. There is every possibility of downloading malicious files unintentionally, which could corrupt the system and user data. With the advancements in technology and growing dependence on digital data, malicious software detection has become a crucial task. The existing approaches need modifications to support and detect the latest attacks. Recently, artificial intelligence-based malicious file detection methods have been proposed. In the past, most of the works analyzed the executable file features and visual features from their corresponding images independently. Additionally, image-based analysis has been exploited for categorical classification, i.e., finding the family once it is known to be malware. We propose a CNN-based model that extracts visual features from malware images, which outperforms existing approaches on a benchmark dataset like MalImg. We study the effect of using a hybrid feature set containing these visual features integrated with statically obtained opcode frequencies for the detection of malware. Our experiments on standard datasets demonstrate that there is no significant performance improvement using this hybrid approach. © 2021, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.