Multi-Modal Adversarial Activity Detection Using Keyboard and Mouse Dynamics

Behavioral biometrics is a rapidly advancing field that leverages human-computer interaction patterns such as typing dynamics, mouse movements, and interaction timing for user authentication and anomaly detection. Unlike traditional biometric systems based on physiological traits, behavioral biometrics offer a non-intrusive and continuous security layer that passively operates during regular system usage. Traditional approaches also use aggregated features losing intricate information of user behavior. In this research, we present a multi-modal approach to behavioral biometric classification by fusing keystroke and mouse activity data. We introduce a temporal alignment framework that segments and synchronizes fixed-size windows of keyboard input with corresponding mouse movement sequences, preserving temporal relationships between modalities and retaining the rich nuances of user behavior. Deep learning models like Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks are employed to classify user sessions as benign or adversarial. Our results show that integrating keyboard and mouse modalities significantly improves classification performance over uni-modal systems, achieving up to 99% accuracy on a dataset of 100 users, evenly split between benign and adversarial behavior. This work highlights the value of multi-modal, non-intrusive behavioral biometrics in enhancing the robustness of modern threat and intrusion detection systems.