Seeing Without Exposing: Adaptive Privacy Control for Open-World, Context-Hungry MLLMs
Researchers have developed Anchored Privacy Drifting (APD), a novel training-free method to enhance privacy in multimodal large language models (MLLMs). APD addresses challenges where user inputs and visual contexts may contain sensitive information by semantically altering privacy-sensitive elements while preserving essential contextual cues. The effectiveness of APD was evaluated using AdaptShield, a new benchmark designed to assess both privacy protection and contextual utility, showing significant improvements across several MLLM series. AI
IMPACT This new method and benchmark could lead to more secure and trustworthy multimodal AI applications by better protecting sensitive user data.