A Practical Black-Box Attack on Source Code Authorship Identification Classifiers

Abstract

Existing researches have recently shown that adversarial stylometry of source code can confuse source code authorship identification (SCAI) models, which may threaten the security of related applications such as programmer attribution, software forensics, etc. In this work, we propose source code authorship disguise (SCAD) to automatically hide programmers' identities from authorship identification, which is more practical than the previous work that requires to known the output probabilities or internal details of the target SCAI model. Specifically, SCAD trains a substitute model and develops a set of semantically equivalent transformations, based on which the original code is modified towards a disguised style with small manipulations in lexical features and syntactic features. When evaluated under totally black-box settings, on a real-world dataset consisting of 1,600 programmers, SCAD induces state-of-the-art SCAI models to cause above 30% misclassification rates. The efficiency and utility-preserving properties of SCAD are also demonstrated with multiple metrics. Furthermore, our work can serve as a guideline for developing more robust identification methods in the future.