Seeing Without Exposing: Adaptive Privacy Control for Open-World, Context-Hungry MLLMs

Researchers have proposed a training-free method called Anchored Privacy Drifting (APD) to protect sensitive information in multimodal large language models while preserving the contextual cues those models need to function, according to a paper posted to arXiv on 5 June 2026 [1]. The approach addresses a tension inherent in multimodal large language models (MLLMs). User-provided inputs can contain unpredictable sensitive data, yet model reasoning often depends on rich visual context that may itself be privacy-sensitive [1]. Existing protection methods rely on predefined sensitive categories and fixed obfuscation strategies, which the authors argue are insufficient for the open-world nature of MLLM inputs [1]. APD instead drifts privacy-sensitive elements toward semantically equivalent alternatives while anchoring contextual cues to the source image [1]. The technique requires no additional training [1]. To measure performance, the team built AdaptShield, a benchmark covering 22 privacy categories that combines conventional privacy metrics with MLLM-based assessments of contextual utility [1]. Across four MLLM series — Qwen2.5, Qwen3, InternVL3, and InternVL3.5 — APD delivered average gains of 10.4% on textual categories and 8.5% under MLLM-based evaluation [1]. The paper does not include direct quotes from the authors. The work appears as a preprint and has not yet been peer-reviewed [1]. The broader research landscape on arXiv shows a cluster of contemporaneous papers exploring privacy and utility trade-offs in vision-language systems, though the specific excerpts retrieved from those papers contain only metadata about code-finder and repository tools and do not provide substantive technical context [3][4][5]. Separately, the United Nations Sustainable Development Goals, adopted in 2015, highlight the cross-cutting nature of privacy-adjacent concerns such as reduced inequalities and strong institutions, though the SDGs do not directly address MLLM privacy [6]. In molecular biology, transcription factors regulate gene expression by binding to specific DNA sequences, a mechanism unrelated to the APD method but illustrative of how anchoring and drifting concepts appear across scientific domains [7].