PEBench introduces the first specialized dataset to evaluate machine unlearning capabilities in Multimodal Large Language Models, addressing critical privacy and security concerns.

• Creates a controlled environment to test how effectively MLLMs can 'forget' specific information
• Enables systematic assessment of unlearning techniques across different models
• Provides a foundation for developing more privacy-preserving AI systems
• Establishes metrics for measuring unlearning efficacy in multimodal contexts

This research is vital for security as it directly addresses the growing challenge of protecting personal data within AI systems while maintaining model performance. PEBench offers a standardized way to evaluate and improve privacy safeguards in increasingly powerful multimodal AI technologies.

PEBench: A Fictitious Dataset to Benchmark Machine Unlearning for Multimodal Large Language Models