This research introduces a novel approach to determine if specific data was used to train large language models, addressing a critical security gap in AI evaluation and use.

• Proposes Fine-tuned Score Deviation (FSD) technique that improves detection accuracy
• Demonstrates that fine-tuning smaller models on known data creates better detection capabilities
• Achieves significantly better results than traditional methods like perplexity testing
• Provides a practical tool for auditing LLMs and reducing ethical risks

This work matters for security professionals as it offers a robust method to verify model training compliance, detect potential copyright violations, and ensure fair model comparisons - crucial steps toward more transparent and trustworthy AI systems.

Fine-tuning can Help Detect Pretraining Data from Large Language Models