AI Data Poisoning Red Teaming

소개

This skill provides a comprehensive framework for red teaming machine learning training pipelines by simulating advanced adversarial attacks. It enables security analysts to evaluate how models respond to label flipping, backdoor triggers, and sophisticated clean-label attacks that are often undetectable by standard validation. By integrating this skill into the development lifecycle, teams can identify critical weaknesses in data ingestion, fine-tuning, and RLHF processes, ultimately mapping vulnerabilities to industry-standard frameworks like OWASP LLM and MITRE ATLAS.

주요 기능

• Maps findings to OWASP LLM Top 10 and MITRE ATLAS frameworks
• Simulates label flipping and backdoor injection attacks
• 1 GitHub stars
• Tests LLM fine-tuning and RLHF pipeline integrity
• Implements sophisticated clean-label poisoning scenarios
• Provides automated impact assessment and attack success metrics

사용 사례

• Testing fine-tuned LLMs for resilience against instruction and preference poisoning
• Validating the security of crowdsourced or externally scraped training datasets
• Evaluating the effectiveness of data provenance and integrity checks in MLOps