USENIX Security '15 - De-anonymizing Programmers via Code Stylometry

De-anonymizing Programmers via Code Stylometry

Aylin Caliskan-Islam, Drexel University; Richard Harang, U.S. Army Research Laboratory; Andrew Liu, University of Maryland; Arvind Narayanan, Princeton University; Clare Voss, U.S. Army Research Laboratory; Fabian Yamaguchi, University of Goettingen; Rachel Greenstadt, Drexel University

Source code authorship attribution is a significant privacy threat to anonymous code contributors. However, it may also enable attribution of successful attacks from code left behind on an infected system, or aid in resolving copyright, copyleft, and plagiarism issues in the programming fields. In this work, we investigate machine learning methods to de-anonymize source code authors of C/C++ using coding style. Our Code Stylometry Feature Set is a novel representation of coding style found in source code that reflects coding style from properties derived from abstract syntax trees.

Our random forest and abstract syntax tree-based approach attributes more authors (1,600 and 250) with significantly higher accuracy (94% and 98%) on a larger data set (Google Code Jam) than has been previously achieved. Furthermore, these novel features are robust, difficult to obfuscate, and can be used in other programming languages, such as Python. We also find that (i) the code resulting from difficult programming tasks is easier to attribute than easier tasks and (ii) skilled programmers (who can complete the more difficult tasks) are easier to attribute than less skilled programmers.

View the full USENIX Security '15 program at https://www.usenix.org/conference/usenixsecurity15/technical-sessions