This research presents innovative watermarking methods specifically designed to detect misuse of open-source large language models like Llama3.

• Addresses two key misuse scenarios: intellectual property violation and usage policy violation
• Introduces watermarking techniques that work without access to the inference process
• Demonstrates effective detection of potential misuse while maintaining model performance
• Provides a practical framework for responsible AI deployment

For security professionals, this research offers crucial tools to protect intellectual property and enforce usage policies as open-source LLMs become increasingly capable and widely available.

Mark Your LLM: Detecting the Misuse of Open-Source Large Language Models via Watermarking