Detecting Web Bots via Mouse Dynamics and Communication Metadata
Abstract
The illegitimate automated usage of Internet services by web robots (bots) is an ongoing problem. While bots increase the cost of operations for service providers and can affect user satisfaction, e.g., in social media and games, the main problem is that some services should only be usable by humans, but their automated usage cannot be prevented easily. Currently, services are protected against bots using visual CAPTCHA systems, the de facto standard. However, they are often annoying for users to solve. Typically, CATPCHAs are combined with heuristics and machine-learning approaches to reduce the number of times a human needs to solve them. These approaches use request data like IP and cookies but also biometric data like mouse movements. Such detection systems are primarily closed source, do not provide any performance evaluation, or have unrealistic assumptions, e.g., that sophisticated bots only move the mouse in straight lines. Therefore we conducted an experiment to evaluate the usefulness of detection techniques based on mouse dynamics, request metadata, and a combination of both. Our findings indicate that biometric data in the form of mouse dynamics performs better than request data for bot detection. Further, training a mouse dynamic classifier benefits from external and not only website-specific mouse dynamics. Our classifier, which differentiates between artificial and human mouse movements, achieves similar results to related work under stricter and more realistic conditions.