Scroll Behavior Analysis Proposed for Bot Detection

Limitations of Previous Bot Detection

Previous research used burstiness (B) and memory (M) coefficients to differentiate human and bot sessions based on request timings. This method identified sessions as human if they originated from a browser loading CSS and JS files, allowing advanced bots using headless browsers to appear human. The approach distinguished between crafted requests and requests from a real browser, rather than between bots and humans.

Scroll Behavior as a New Metric

A new approach suggests that human scrolling patterns are distinctly bursty and non-linear, unlike typical bot scrolling. This bursty pattern, observed when humans search for information on a page, is believed to be difficult for current scraping bots to imitate effectively. The proposal is to apply burstiness and memory analysis to the inter-event times of scroll events.

Testing the Predictive Power

The next step involves testing if burstiness and memory applied to scroll event timings can serve as predictive variables in a machine learning model. This model would aim to distinguish between human users and bots. The dataset from the paper "FP-Agent: Fingerprinting AI Browsing Agents" by Ethan Wang, et al., which recorded data from different agents and humans navigating a controlled website, will be used for this testing.