A Multimodal AI Approach for Predicting Personal Privacy Preferences in Photo Sharing

Yuqi Niu, Kejia Zhang, Weidong Qiu, Nadin Kökciyan

Keywords: Privacy, Photo-sharing, Social Media

Abstract: Social media users frequently share photos containing others’ faces without explicit consent, raising significant privacy risks. Existing face privacy protection methods face a major limitation: they cannot reliably predict individuals’ diverse privacy preferences, particularly when only limited training data are available. To address this challenge, we analyze a large-scale public survey dataset, examining both the influence of contextual factors on privacy decisions and the variability of personalized privacy needs across individuals. Based on these insights, we propose a novel framework that leverages multimodal large language models to predict users’ privacy preferences for public photo sharing using in-context learning with few-shot and zero-shot prompts, without model fine-tuning. Our framework achieves strong performance on the dataset, reaching 84.45\% accuracy and improving over baselines by up to 18.77\%. Moreover, for 87.5\% of the 486 users, the model correctly predicts at least 12 out of 16 scenarios. These results demonstrate the potential of our approach to advance automated privacy protection on social media by incorporating individual preferences and enabling face-level privacy management.

Area: Coordination, Organisations, Institutions, Norms and Ethics (COINE)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1369