计算机科学
恶意软件
理论计算机科学
程序设计语言
操作系统
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2023-10-22
卷期号:7: 167-175
被引量:2
标识
DOI:10.1109/qrs-c60940.2023.00072
摘要
Malware family classification is a well-attended research topic in the field of malware analysis. The existing approaches for malware classification typically utilize API call sequences to represnet malware. However, these approaches extract the features of malware from sequential structure, not taking into account that APIs have different behavioral semantics in different control structures. This may degrade the classification performance. In this paper, we propose an approach for malware family classification based on Behavior Tree representations of malware and heterogeneous graph neural networks. The Behavior Trees, which are mined directly from API call sequences, explicitly describe control structures between the API calls for extracting more accurate behavioral features. Compare to other structure models, behavior trees are easier to construct and more compact. Due to the heterogeneity of different node types, the ordered nature of the sibling nodes, and the uncertainty of the distance between API call nodes of the behavior trees, it is difficult for the existing methods to extract their features. To enable feature extraction, we convert the behavior trees into an equivalent binary relation graphs. Then we feed the binary relation graphs into heterogeneous GNNs (simple-HGN and RGCN) to extract the malicious behavioral features and train a malware familial classification model. The experiments were conducted on a public dataset containing API call sequences of 10,620 malwares from 43 families. The experimental results show that our approaches have more than 80% accuracy, outperforming the other compared approaches which are based on API call sequences.
科研通智能强力驱动
Strongly Powered by AbleSci AI