Robot Adversarial Attack on Keystroke Dynamics Based User Authentication System

Abstract

Adversarial attacks on machine learning systems are an important area of study in cybersecurity. Keystroke dynamics (KD)-based user authentication systems utilize human typing behavior to distinguish between users. As robots continue to advance and become more capable at mimicking human behavior, they may increasingly pose a threat to behavioral biometric systems by performing adversarial attacks. In this study, we propose a robot adversarial attack framework to evaluate the resilience of eight commonly used classifiers and detectors in the keystroke dynamics literature against robot attacks. We invited 27 participants across three types of passwords: a complex password (CP) .tie5Roanl, a text-based password (TP) kicsikutyatarka, and a numeric password (NP) 4121937761. The results show that 1) in white-box attack scenarios, the robot achieves up to 100% Accuracy (ACC) and over 95% Equal Error Rate (EER); and 2) in grey-box attack scenarios, the results also demonstrate significant vulnerabilities, highlighting the need for robust defense strategies to enhance the security of keystroke dynamics-based authentication systems against robotic adversarial attacks.