Adversarial Code Obfuscation for Defending Against LLM-Based Analysis
7 hours ago
- Large Language Models (LLMs) pose security and privacy risks by analyzing, reconstructing, or reverse-engineering source code logic, potentially leaking intellectual property.
- Acoda is a genetic algorithm-based adversarial code obfuscation framework that defends against LLM-based code analysis.
- It uses 8 semantics-preserving obfuscation methods leveraging LLMs' safety alignment and token-based information processing.
- Acoda iteratively optimizes obfuscation strategies via a genetic algorithm to generate adversarial samples that maximize defensive effectiveness.
- A quantitative evaluation framework combines an auxiliary LLM and four metrics to assess how target LLMs analyze obfuscated code.
- Experiments on 7 state-of-the-art LLMs (e.g., GPT-4o, DeepSeek, Qwen, Llama, Gemma) show Acoda achieves up to 70% attack success rate (ASR), with strong cross-model transferability and minimal runtime overhead while preserving semantics.
- This work provides a new perspective for code protection and LLM security defense in the LLM era.