A Survey on Adversarial Machine Learning for Code Data: Realistic Threats, Countermeasures, and Interpretations

Code AI systems face three major security threats: Code Language Models are vulnerable to poisoning attacks (corrupting training data), evasion attacks (tricking the model at deployment), and privacy attacks (stealing confidential information), mirroring the standard security framework of confidentiality, integrity, and availability.

Real-world software development creates unique risks: Unlike general AI systems, CLMs used for code generation and understanding operate in "extremely adversarial environments" where malicious actors can inject poisoned code into training data or manipulate inputs, potentially compromising the integrity of software systems that rely on these tools.

Current defenses are incomplete: While the survey identifies various countermeasures across the 79 papers reviewed, the research landscape reveals gaps in protection strategies, suggesting that existing safeguards may be insufficient for deploying CLMs safely in critical software engineering applications.

Explainability is key to trustworthiness: The survey highlights the importance of eXplainable AI (XAI) as a novel perspective for understanding and defending against adversarial attacks on code models, implying that making AI decision-making transparent is essential for identifying when systems have been compromised or are behaving unexpectedly.

Interdisciplinary collaboration is essential: Addressing these security challenges requires coordination across AI research, computer security, and software engineering communities, as vulnerabilities in code models span technical, operational, and systemic dimensions that no single field can address alone.