Leveraging Large Language Models to Obscure Code Stylometry: A Comparative Study of GPT-3.5 and GPT-4

Researchers demonstrate that Large Language Models like GPT-3.5 and GPT-4 can effectively obscure programmer code stylometry while maintaining functionality, challenging the reliability of authorship attribution techniques used in cybersecurity. The study reveals that structured, multi-shot prompting strategies outperform single-shot approaches in evading detection by traditional machine learning classifiers.

This research addresses a critical vulnerability in code-based forensic techniques that cybersecurity professionals rely upon to attribute malicious code to specific threat actors or developers. As LLMs become increasingly sophisticated, their capacity to transform code while preserving its execution logic introduces meaningful risks to attribution pipelines that depend on stylometric fingerprinting—the unique patterns that betray a programmer's identity.

The broader context involves an arms race between defensive security measures and AI-assisted obfuscation. Authorship attribution has historically served as a cornerstone for incident response, threat intelligence, and law enforcement investigations. The ability of GPT models to defeat these techniques at scale fundamentally undermines confidence in current methodologies.

For the software security industry, this creates immediate implications. Organizations cannot assume their existing stylometry-based detection systems will identify compromised code or attribute attacks with historical accuracy. Security vendors and researchers must accelerate development of multi-factor authentication approaches for code analysis that combine stylometry with behavioral, contextual, and cryptographic verification methods.

The functionality preservation challenge noted in the study highlights a secondary concern: LLMs can introduce subtle bugs or security weaknesses while refactoring code, potentially creating exploitable vulnerabilities alongside successful obfuscation. Looking ahead, the field requires advances in robust code analysis that remain resilient against AI-assisted evasion, including dynamic execution analysis, supply chain verification, and enhanced logging mechanisms that operate independently of static code inspection.

• →LLMs can effectively obscure code stylometry signatures while maintaining functionality, undermining traditional authorship attribution methods.
• →Multi-shot prompting strategies demonstrate superior effectiveness compared to single-shot approaches in evading Random Forest classifiers.
• →Functionality preservation remains a technical challenge when using LLMs to alter code, introducing potential vulnerabilities.
• →Current cybersecurity attribution pipelines relying on code stylometry face significant robustness challenges from advanced AI capabilities.
• →Security teams must develop multi-factor code analysis approaches that extend beyond stylometric fingerprinting for reliable threat attribution.