Continuous User Authentication Using Mouse Dynamics, Machine Learning, and Minecraft

Machine and Deep Learning Applications to Mouse Dynamics for Continuous User Authentication

“…Note that variable mouse action block lengths have also been proposed, with mouse action block lengths usually a great order of magnitude larger than 10. However, using a smaller mouse action block length and observing similar results indicates that our dataset encodes more information in a smaller amount of data, which is what we were able to conclude in [12]. Additional features such as velocity, acceleration, jerk, angular velocity, angle and length of mouse…”

“…In tandem with the use of machine and deep learning, mouse dynamics could potentially serve as an inexpensive, non-intrusive behavioral biometric for user authentication. Thus, the scope of this paper revolves around evaluating three machine learning and three deep learning algorithms on an improved novel dataset introduced in our previous work [12]. This dataset contains the mouse dynamics data of 40 users as they engaged in a highintensity task on a desktop computer for 20 minutes.…”

“…This paper builds on the results of [12], where we hypothesized that collecting data in an environment where users have more variability and freedom to choose their actions would naturally encode more distinct features in each user's data. We proceeded by collecting data from 10 users while they played their own 20-minute game of Minecraft on a designated lab computer, then trained binary random forest classifiers for user authentication.…”

“…A positive instance was defined as a genuine action that should warrant continued user access, while negative instances were imposter mouse actions that should result in additional monitoring/blocking of the user's access. Many data processing and model building decisions in [12] were inspired by the original paper proposing mouse dynamics as a reliable behavioral biometric [15]. Authors of [15] introduced mouse dynamics as a novel human-computer interaction behavior that could serve as an alternative security medium.…”

“…For example, due to the spatiotemporal attributes of the data, mouse dynamics can be framed as a time-series problem yet integrate deep learning through different modalities. There are multiple ways to leverage CNNs for mouse dynamics [23], such (a) as using the X and Y coordinates from the raw data to create user mouse maps, as observed in [12], and evaluate their spatial features using a two-dimensional CNN [24]. By focusing on and extracting the spatial features of mouse dynamics data, the image processing capabilities of CNNs can be similarly applied to user authentication.…”