EVADING STYLISTIC ANALYSIS OF BINARY PROGRAMS USING ADVERSARIAL LEARNING

Abstract

Recent work suggests that it may be possible to determine the author of a binary program simply by analyzing stylistic features preserved within it. As this poses a threat to the privacy of programmers who wish to distribute their work anonymously, we consider steps that can be taken to mislead such analysis. We begin by exploring the effect of non-standard compiler optimizations on the features used for stylistic analysis. Building on these findings, we propose a black-box attack on a state-of-the-art classifier using compiler optimizations. Finally, we discuss our results, as well as implications for the field of binary stylometry.