Low-Resource Malware Family Detection by Cross-Family Knowledge Transfer

Low-resource malware families are highly susceptible to being overlooked when using machine learning models or deep for automated detection because of the small amount data samples. When we target train a classifier low-resource family, training family itself is not sufficient good classifier. In this work, study relationship between different and improve performance model based on method in detection. First, propose an empirical supportive score measure transfer quality find that transferring varies lot families. Second, Sequential Family Selection (SFS) algorithm select multiple as data. With SFS, only knowledge from several We conduct experiments 16 4 models, results show our could outperform best baselines by 2.29% average achieves 14.16% improvement accuracy at highest. Third, transferred capture common characteristics proposing achieve family. Our also be applicable image signal