ANVIL uses LLMs for anomaly-based vulnerability detection

Researchers have developed ANVIL, a novel approach to identifying software vulnerabilities by reframing the problem as anomaly detection. Unlike traditional supervised methods that struggle with limited labeled data, ANVIL leverages Large Language Models (LLMs) trained on vast unlabelled code. The system works by having an LLM reconstruct masked code segments, scoring deviations from the original as anomalies. This method has demonstrated superior performance compared to existing supervised detectors on the PrimeVul dataset, achieving up to twice the Top-3 accuracy and significantly improving ROC-AUC. Furthermore, ANVIL's integration with fuzzers has successfully uncovered two previously unknown vulnerabilities, highlighting its practical application in enhancing software security. AI

IMPACT This research could lead to more effective and efficient methods for detecting software vulnerabilities, improving overall code security.

RANK_REASON This is a research paper detailing a new method for vulnerability identification using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• ANVIL
• Large Language Models
• LineVD
• LineVul
• LLMAO
• LLMs
• PrimeVul
• Weizhou Wang

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →