This research exposes critical security vulnerabilities in medical multimodal large language models through model stealing attacks, which can extract valuable IP despite privacy regulations.

• Demonstrates how attackers can steal medical MLLMs using non-medical data through adversarial domain alignment
• Achieves up to 84% of the original model's performance in radiology report generation
• Proposes an effective watermarking defense mechanism that can reduce attack success by 30%
• Shows that medical AI models are vulnerable even when protected by data privacy barriers

This work highlights the urgent need for robust protection of valuable medical AI assets that incorporate both visual and language capabilities, particularly in healthcare settings where data scarcity increases model value.

Medical Multimodal Model Stealing Attacks via Adversarial Domain Alignment