Shujiang Wu (Johns Hopkins University), Pengfei Sun (F5, Inc.), Yao Zhao (F5, Inc.), Yinzhi Cao (Johns Hopkins University)
Browser fingerprints, while traditionally being used for web tracking, have recently been adopted more and more often for defense or detection of various attacks targeting real-world websites. Faced with these situations, adversaries also upgrade their weapons to generate their own fingerprints---defined as adversarial fingerprints---to bypass existing defense or detection. Naturally, such adversarial fingerprints are different from benign ones from user browsers because they are generated intentionally for defense bypass. However, no prior works have studied such differences in the wild by comparing adversarial with benign fingerprints let alone how adversarial fingerprints are generated.
In this paper, we present the first billion-scale measurement study of browser fingerprints collected from 14 major commercial websites (all ranked among Alexa/Tranco top 10,000). We further classify these fingerprints into either adversarial or benign using a learning-based, feedback-driven fraud and bot detection system from a major security company, and then study their differences. Our results draw three major observations: (i) adversarial fingerprints are significantly different from benign ones in many metrics, e.g., entropy, unique rate, and evolution speed, (ii) adversaries are adopting various tools and strategies to generate adversarial fingerprints, and (iii) adversarial fingerprints vary across different attack types, e.g., from content scraping to fraud transactions.