PrivAR uses vision-language models with chain-of-thought prompting to detect privacy risks in AR by understanding visual context

PrivAR leverages vision language models (VLMs—AI systems that understand images and text together) with chain-of-thought prompting to infer sensitive information types from visual scenes, such as identifying password notes in office environments through contextual reasoning.

The system detects and obfuscates textual content to prevent exposure of sensitive information while preserving contextual cues needed for VLM inference. Experiments on a real-world AR dataset show PrivAR achieves accuracy of 81.48% and F1-score of 84.62% compared to baselines, while reducing privacy leakage rate to 17.58%.

User studies evaluated contextually-informed warning interfaces to enhance privacy awareness in AR design, providing insights into effective privacy-aware AR interfaces.