GEO-Detective: Revealing Location Privacy Risks in Images with LLM Agents

The researchers propose an image geolocation agent, GEO-Detective, which infers where a photo was taken by imitating human reasoning and tool use, in order to reveal the location privacy risks of images on social platforms. The agent executes in four steps, adaptively selects strategies according to image difficulty, and is equipped with specialized tools such as visual reverse search, simulating the way humans collect external geographic clues. Experiments show that on country-level geolocation tasks it improves over baseline LLMs by more than 11.1%, and at finer-grained levels there is still an improvement of about 5.2%; when equipped with external clues, the proportion of "unknown" predictions drops by more than 50.6%. The authors also explore various defense strategies, and GEO-Detective shows stronger robustness to them, highlighting the necessity of developing more effective privacy protection methods.