This research introduces the Parameter Learning Attack (PLA) method to protect large vision-language models from unauthorized usage and copyright infringement.

• Creates adversarial images that can be embedded within model parameters
• Enables copyright tracking without affecting normal model performance
• Provides a technical solution to detect fine-tuned or stolen models
• Addresses growing security concerns in the expanding LVLM ecosystem

As vision-language models become more commercially valuable, this research offers critical protection mechanisms for AI developers and companies investing in these technologies.

Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Images