Nina-Huang/SopriBench

Paper: https://arxiv.org/abs/2606.06784

SopriBench is the first large-scale multimodal benchmark dedicated to evaluating privacy inference risks on social media platforms. It contains 50 user profiles and 1,569 images with fine-grained annotations of 28 personal attributes across 4 dimensions: basic information, socioeconomic status, lifestyle, and sensitive information.

The dataset is designed to facilitate research on privacy-preserving AI, adversarial privacy attacks, and multimodal representation learning. All data is synthetically generated and does not contain any real user information.

This repository hosts a 3-user preview subset to demonstrate the dataset structure and annotation quality. The complete dataset will be released upon paper publication.

The complete version is available at: https://huggingface.co/datasets/Nina-Huang/SopriBench-full