This research reveals how Multi-Modal Large Language Models (MLLMs) can unintentionally memorize private visual content, even when that content is completely irrelevant to training tasks.

• MLLMs can extract and reproduce private information from visual content without explicit training
• Task-irrelevant content (like random text in images) can be memorized and later extracted
• The risk increases with repeated exposure during model training
• Proposed probing framework helps detect this privacy vulnerability

This research is crucial for security professionals as it highlights an emerging privacy risk in MLLMs that process visual data, requiring more robust safeguards for user images processed by multimodal AI systems.

Watch Out Your Album! On the Inadvertent Privacy Memorization in Multi-Modal Large Language Models