Correlating High- and Low-Level Features: Increased Understanding of Malware Classification
Journal article, Peer reviewed
Accepted version
View/ Open
Date
2019Metadata
Show full item recordCollections
Original version
Lecture Notes in Computer Science. 2019, 11689 149-167. https://doi.org/10.1007/978-3-030-26834-3_9Abstract
Malware brings constant threats to the services and facilities used by modern society. In order to perform and improve anti-malware defense, there is a need for methods that are capable of malware categorization. As malware grouped into categories according to its functionality, dynamic malware analysis is a reliable source of features that are useful for malware classification. Different types of dynamic features are described in literature [5, 6, 13]. These features can be divided into two main groups: high-level features (API calls, File activity, Network activity, etc.) and low-level features (memory access patterns, high-performance counters, etc). Low-level features bring special interest for malware analysts: regardless of the anti-detection mechanisms used by malware, it is impossible to avoid execution on hardware. As hardware-based security solutions are constantly developed by hardware manufacturers and prototyped by researchers, research on low-level features used for malware analysis is a promising topic. The biggest problem with low-level features is that they don’t bring much information to a human analyst. In this paper, we analyze potential correlation between the low- and high-level features used for malware classification. In particular, we analyze n-grams of memory access operations found in [6] and try to find their relationship with n-grams of API calls. We also compare performance of API calls and memory access n-grams on the same dataset as used in [6]. In the end, we analyze their combined performance for malware classification and explain findings in the correlation between high- and low-level features.