Authors:
Patrick Kochberger
1
;
2
;
Maximilian Gramberger
1
;
Sebastian Schrittwieser
2
;
Caroline Lawitschka
2
and
Edgar Weippl
3
Affiliations:
1
Institute of IT Security Research, St. Pölten University of Applied Sciences, Austria
;
2
Research Group Security and Privacy, University of Vienna, Austria
;
3
SBA Research, Vienna, Austria
Keyword(s):
Software Protections, Code Obfuscation, Large Language Model, GPT.
Abstract:
This study explores the efficacy of large language models, specifically GPT-3.5, in obfuscating C source code for software protection. We utilized eight distinct obfuscation techniques in tandem with seven representative C code samples to conduct a comprehensive analysis. The evaluation was performed using a Python-based tool we developed, which interfaces with the OpenAI API to access GPT-3.5. Our metrics of evaluation included the correctness and diversity of the obfuscated code, along with the robustness of the resultant protection. While the diversity of the resulting code was found to be commendable, our findings indicate a prevalent issue with the correctness of the obfuscated code and the overall level of protection provided. Consequently, we assert that while promising, the feasibility of deploying large language models for automatic code obfuscation is not yet sufficiently established. This study signifies an important step towards understanding the limitations and potential
of AI-based code obfuscation, thereby informing future research in this area.
(More)