Department of Computer, Control and Management Engineering

The skyrocketing grow rate of new malware brings novel challenges to protect computers and networks. Discerning truly novel malware from variants of known samples is a way to keep pace with this trend. This can be done by grouping known malware in families by similarity and classifying new samples into those families. As malware and their families evolve over time, approaches based on classifiers trained on a fixed ground truth are not suitable. Other techniques use clustering to identify families but they need to periodically re-cluster the whole set of samples, which does not scale well. A promising approach is based on incremental clustering, where periodically only yet unknown samples are clustered to identify new families, and classifiers are re-trained accordingly. However, the latter solutions usually are not able to immediately react and identify new malware families. In this paper we propose MalFamAware, a novel approach to malware family identification based on an online clustering algorithm, namely BIRCH, which efficiently updates clusters as new samples are fed without requiring to re-scan the entire dataset. MalFamAware is able to both classify new malware in existing families and identify new families at runtime. We present experimental evaluations where MalFamAware outperforms both total re-clustering and incremental clustering solutions in terms of accuracy and time. We also compare our solution with classifiers re-trained over time, obtaining better accuracy, in particular when samples belong to yet unknown families.