Adversarial Code Obfuscation Against LLM-Based Analysis
Acoda offers an effective defense mechanism against LLM-based code analysis.
The rise of Large Language Models (LLMs) in software engineering has introduced significant security and privacy risks. To mitigate these risks, a genetic algorithm-based framework named Acoda is proposed, which employs eight semantics-preserving obfuscation methods to defend against LLM analysis. Experimental results indicate that Acoda can effectively induce LLMs to misinterpret or refuse code analysis, achieving an attack success rate of up to 70% across various models.