Catch Me If You Can: Deceiving Stance Detection and Geotagging Models to Protect Privacy of Individuals on Twitter

Abstract

The recent advances in natural language processing haveyielded many exciting developments in text analysis and lan-guage understanding models; however, these models can alsobe used to track people, bringing severe privacy concerns. Inthis work, we investigate what individuals can do to avoid be-ing detected by those models while using social media plat-forms. We ground our investigation in two exposure-riskytasks, stance detection and geotagging. We explore a varietyof simple techniques for modifying text, such as inserting ty-pos in salient words, paraphrasing, and adding dummy socialmedia posts. Our experiments show that the performance ofBERT-based models fine-tuned for stance detection decreasessignificantly due to typos, but it is not affected by paraphras-ing. Moreover, we find that typos have minimal impact onstate-of-the-art geotagging models due to their increased re-liance on social networks; however, we show that users candeceive those models by interacting with different users, re-ducing their performance by almost 50%.