This research demonstrates how simple pruning techniques can significantly reduce unwanted data memorization in LLMs, offering an effective defense against privacy attacks.

• Pruning effectively decreases the extent to which LLMs reproduce memorized training data
• Positions model pruning as a practical approach to mitigate membership inference attacks
• Provides a straightforward yet powerful technique for enhancing LLM privacy and security
• Demonstrates important tradeoffs between model utility and memorization reduction

For security professionals, this research offers a viable technique to address privacy concerns in AI deployment without requiring complex architectural changes or expensive retraining processes.

Pruning as a Defense: Reducing Memorization in Large Language Models