MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models

Xiongtao Sun, Hui Li, Jiaming Zhang, Yujie Yang, Kaili Liu, Ruxin Feng, Wen Jun Tan, Wei Yang Bryan Lim
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:116792-116819, 2026.

Abstract

Modern Vision-Language Models (VLMs) pose significant individual-level privacy risks by linking fragmented multimodal data to identifiable individuals through hierarchical chain-of-thought reasoning. However, existing privacy benchmarks remain structurally insufficient for this threat, as they primarily evaluate privacy perception while failing to address the more critical risk of privacy reasoning: a VLM’s ability to infer and link distributed information to construct individual profiles. To address this gap, we propose MultiPriv, the first benchmark designed to systematically evaluate individual-level privacy reasoning in VLMs. We introduce the Privacy Perception and Reasoning (PPR) framework and construct a bilingual multimodal dataset with synthetic individual profiles, where identifiers, such as faces and names, are linked to sensitive attributes. This design enables nine challenging tasks spanning attribute detection, cross-image re-identification, and chained inference. We conduct a large-scale evaluation of over 50 open-source and commercial VLMs. In our controlled benchmark, 60% of widely used VLMs can perform individual-level privacy reasoning with up to 80% accuracy, suggesting a significant potential threat to personal privacy. The benchmark is available at https://github.com/CyberChangAn/MultiPriv-PII.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-sun26t, title = {{M}ulti{P}riv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models}, author = {Sun, Xiongtao and Li, Hui and Zhang, Jiaming and Yang, Yujie and Liu, Kaili and Feng, Ruxin and Tan, Wen Jun and Lim, Wei Yang Bryan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {116792--116819}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/sun26t/sun26t.pdf}, url = {https://proceedings.mlr.press/v306/sun26t.html}, abstract = {Modern Vision-Language Models (VLMs) pose significant individual-level privacy risks by linking fragmented multimodal data to identifiable individuals through hierarchical chain-of-thought reasoning. However, existing privacy benchmarks remain structurally insufficient for this threat, as they primarily evaluate privacy perception while failing to address the more critical risk of privacy reasoning: a VLM’s ability to infer and link distributed information to construct individual profiles. To address this gap, we propose MultiPriv, the first benchmark designed to systematically evaluate individual-level privacy reasoning in VLMs. We introduce the Privacy Perception and Reasoning (PPR) framework and construct a bilingual multimodal dataset with synthetic individual profiles, where identifiers, such as faces and names, are linked to sensitive attributes. This design enables nine challenging tasks spanning attribute detection, cross-image re-identification, and chained inference. We conduct a large-scale evaluation of over 50 open-source and commercial VLMs. In our controlled benchmark, 60% of widely used VLMs can perform individual-level privacy reasoning with up to 80% accuracy, suggesting a significant potential threat to personal privacy. The benchmark is available at https://github.com/CyberChangAn/MultiPriv-PII.} }
Endnote
%0 Conference Paper %T MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models %A Xiongtao Sun %A Hui Li %A Jiaming Zhang %A Yujie Yang %A Kaili Liu %A Ruxin Feng %A Wen Jun Tan %A Wei Yang Bryan Lim %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-sun26t %I PMLR %P 116792--116819 %U https://proceedings.mlr.press/v306/sun26t.html %V 306 %X Modern Vision-Language Models (VLMs) pose significant individual-level privacy risks by linking fragmented multimodal data to identifiable individuals through hierarchical chain-of-thought reasoning. However, existing privacy benchmarks remain structurally insufficient for this threat, as they primarily evaluate privacy perception while failing to address the more critical risk of privacy reasoning: a VLM’s ability to infer and link distributed information to construct individual profiles. To address this gap, we propose MultiPriv, the first benchmark designed to systematically evaluate individual-level privacy reasoning in VLMs. We introduce the Privacy Perception and Reasoning (PPR) framework and construct a bilingual multimodal dataset with synthetic individual profiles, where identifiers, such as faces and names, are linked to sensitive attributes. This design enables nine challenging tasks spanning attribute detection, cross-image re-identification, and chained inference. We conduct a large-scale evaluation of over 50 open-source and commercial VLMs. In our controlled benchmark, 60% of widely used VLMs can perform individual-level privacy reasoning with up to 80% accuracy, suggesting a significant potential threat to personal privacy. The benchmark is available at https://github.com/CyberChangAn/MultiPriv-PII.
APA
Sun, X., Li, H., Zhang, J., Yang, Y., Liu, K., Feng, R., Tan, W.J. & Lim, W.Y.B.. (2026). MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:116792-116819 Available from https://proceedings.mlr.press/v306/sun26t.html.

Related Material