CATCHA: When Cats Track Your Movements Online

Abstract

Any website can record its users’ mouse interactions within that site, an emerging practice used to learn about users’ regions of interests usually for personalization purposes. However, the dark side of such recording is that it is oblivious to the users as no permissions are solicited from the users prior to recording (unlike other resources like webcam or microphone). Since mouse dynamics may be correlated with users’ behavioral patterns, any website with nefarious intentions (“cat”) could thus try to surreptitiously infer such patterns, thereby compromising users’ privacy and making them prone to targeted attacks. In this paper, we show how users’ personal information, specifically their demographic characteristics, could leak in the face of such mouse movement eavesdropping. As a concrete case study along this line, we present CATCHA, a mouse analytic attack system that gleans potentially sensitive demographic attributes—age group, gender, and educational background—based on mouse interactions with a game CAPTCHA system (a simple drag-and-drop animated object game to tell humans and machines apart).

CATCHA ’s algorithmic design follows the machine learning approach that predicts unknown demographic attributes based on a total of 64 mouse dynamics features extracted from within the CAPTCHA game, capturing users’ innate cognitive abilities and behavioral patterns. Based on a comprehensive data set of mouse movements with respect to a simple game CAPTCHA collected in an online environment, we show that CATCHA can identify the users’ demographics attributes with a high probability (almost all attributes with more than 85%), significantly better than random guessing (50%) and in a very short span of interaction time (about 14 s). We also provide a thorough statistical analysis and interpretation of differentiating features across the demographics attributes that make users susceptible to the CATCHA attack. Finally, we discuss potential extensions to our attack using other user interaction paradigms (e.g., other types of CAPTCHAs or typical web browsing interactions, and under longitudinal settings), and provide potential mitigation strategies to curb the impact of mouse movement eavesdropping.

A. Neupane and K. Satvat—Work done at UAB.

Appendix

Fig. 4.

An instance of game CAPTCHA. Targets (left) are static and moving objects (right) are mobile. The task of the user is to drag-drop a subset of moving objects to their corresponding target locations.

Table 6. Single game challenge. Performance (all presented in %) of two classification models - (a) cross-validation, and (b) train-test, corresponding to various demographic attributes when using single CAPTCHA game. The figures within the parenthesis (“Random”) in the last column show the random guessing accuracy of the classifier.