This research introduces a dual-purpose token technique that mitigates membership inference attacks (MIAs) in large language models without compromising performance.

• Addresses critical privacy concerns in LLMs by preventing attackers from identifying if specific data was used in training
• Innovates by embedding special tokens that serve both learning and unlearning functions
• Demonstrates effective defense against privacy leaks while maintaining model utility
• Provides a practical solution that accounts for the sequential nature of text data

This work is significant for security professionals as it offers a computationally efficient way to enhance privacy protections in deployed language models, helping organizations mitigate data breach risks.

Tokens for Learning, Tokens for Unlearning: Mitigating Membership Inference Attacks in Large Language Models via Dual-Purpose Training