This research introduces a sample-level unlearning difficulty framework for Large Language Models that enables more precise and interpretable privacy protection.

• Challenges the assumption that all data is equally difficult to unlearn from LLMs
• Proposes a neuro-inspired interpretation to measure unlearning difficulty
• Demonstrates that samples with higher perplexity require more unlearning effort
• Enables more effective privacy protection strategies through selective unlearning

Security Impact: This framework provides organizations with a more nuanced approach to removing sensitive information from deployed AI systems, improving compliance with privacy regulations while maintaining model performance.

A Neuro-inspired Interpretation of Unlearning in Large Language Models through Sample-level Unlearning Difficulty