Behavioral biometrics use mouse dynamics as a method for user authentication and bot detection. Real mouse movement data collection at scale proves difficult because of privacy issues, time requirements and diversity limitations.

This research investigates the application of Generative Adversarial Networks (GANs) to produce artificial mouse movement trajectories that mimic human like behavior. The WGAN-GP model used LSTM networks to generate synthetic mouse movement data from sequential movement information. The model received training from the Bogazici Mouse Dynamics dataset before undergoing evaluation through statistical metrics, classifier performance assessments and visual inspection.

The generated data matches real samples in both spatial and temporal characteristics. The generated data proved difficult for shallow classifiers to differentiate but deep models like LSTMs demonstrated the ability to detect differences through their analysis of timing patterns.

The research shows GANs can effectively produce realistic behavioral data while showing that generating human like motion in sequential models remains an on going challenge.