EM Meets Malicious Data
Yongkang Jiang, Shenghong Li, Tong Li · 2020
This paper examines the problem of inferring underground family truth from inconsistent antivirus vendor labels. Our insight is that vendors are not equally reliable, so we construct a two-dimensional probability matrix for each vendor to model its ability to identify diverse families. Then we formalize the inference task as a maximum likelihood estimation problem with hidden random variables and propose a solution based on the expectation-maximization algorithm.