Semi-Supervised Malware Clustering Based on the Weight of Bytecode and API

Abstract

With the rapid advances of anti-virus and anti-tracking technologies, three aspects in malware clustering need to be improved for effective clustering, i.e., the robustness of features, the accuracy of similarity measurements, and the effectiveness of clustering algorithms. In this paper, we propose a novel malware family clustering approach based on dynamic and static features with their weights. In this approach, we employ a new similarity measurement method based on EMD to improve the accuracy of feature similarities. In addition, to reduce convergence time and improve clustering purity, we design a novel semi-supervised clustering algorithm, termed as S-DBSCAN by involving supervision information into the original algorithm known as Density-Based Spatial Clustering of Applications with Noise (DBSCAN). The experimental results demonstrate that the proposed approach can correctly and accurately distinguish the samples among various families and achieve outperformed purity with 98.7%. INDEX TERMS EMD, hybrid features, semi-supervised clustering, weight.

SoK: The Impact of Unlabelled Data in Cyberthreat Detection

“…Representation learning focuses on finding the features that maximize the performance of a ML classifier [78]. It is possible to use unlabelled data to finetune such selection (e.g., [79]- [83]). Such procedures are ancillary to detection tasks, and hence outside our scope.…”