
This system uses mouse movement dynamics as a behavioural biometric to verify user identity — no passwords, no tokens, no extra hardware. How it works Raw mouse events (x, y, timestamp, button, state) are captured 28+ behavioural features extracted per 256-event sliding window: Movement: velocity, acceleration, jerk, curvature, path efficiency Click: hold duration, click rate, scroll behaviour Spatial: screen region preference, quadrant entropy, x/y range Temporal: inter-event rhythm, micro-stops, idle patterns A residual neural network (MouseResNet) classifies each window Per-user decision thresholds determine the final verdict Sessions with majority 0xFFFF sentinel coordinates are instantly flagged Data quality gate Before any model inference, sessions are checked for 0xFFFF sentinel coordinates — the OS null value emitted when the cursor leaves the tracked area. Sessions where more than 50% of events are sentinel values are immediately classified as impostors without using the model at all. Dataset Trained on the Balabit Mouse Dynamics Challenge dataset. 10 users 65 training sessions ~35,000 feature windows Model performance Metric Value Mean AUC 0.957 Mean EER ~0.111 CSV format record_timestamp,client_timestamp,button,state,x,y 1234567890.123,1234567890.100,NoButton,Move,512,384 1234567890.145,1234567890.122,Left,pressed,512,384 1234567890.198,1234567890.175,Left,released,512,384 Known users in this model user12, user15, user16, user20, user21, user23, user29, user35, user7, user9
10 May 2026