隐私泄露图像地理定位数据集
收藏资源简介:
本研究构建了一个包含50张真实世界图像的数据集,这些图像展示了个人和与隐私相关的环境元素,用于系统分析。数据集通过手动收集并构建,捕捉了现实生活中的隐私敏感场景。该数据集旨在评估ChatGPT o3在图像地理定位任务中的隐私泄露风险。通过实验评估发现,ChatGPT o3能够在60%的情况下以高精度预测用户位置,实现街道级别的准确度(在1英里以内)。该数据集对于研究隐私保护和图像地理定位技术具有重要意义。
This study constructs a dataset consisting of 50 real-world images that depict individuals and privacy-related environmental elements for systematic analysis. The dataset is manually collected and compiled, capturing real-life privacy-sensitive scenarios. It is designed to assess the privacy leakage risks of ChatGPT o3 in the image geolocation task. Experimental evaluations demonstrate that ChatGPT o3 can predict user locations with high accuracy in 60% of cases, achieving street-level accuracy (within 1 mile). This dataset is of great significance for research on privacy protection and image geolocation technologies.

- 1Doxing via the Lens: Revealing Privacy Leakage in Image Geolocation for Agentic Multi-Modal Large Reasoning Model佐治亚大学, 威斯康星大学麦迪逊分校, 南加州大学 · 2025年



