research

ABSTRACT

Synthetic mouse trajectory generators are a core component of aimbots and other automation tools that circumvent interactive security systems. Modern generators ground their output in motor-control science—Fitts' Law timing, lognormal submovements, physiological tremor—yet no prior work rigorously evaluates whether these biomechanical embellishments actually survive detection under adversarial pressure. We present a kinematic detection framework comprising 17 trajectory-shape features across four families (Fitts compliance, submovement morphology, kinematic smoothness, and geometry). On two public mouse dynamics datasets, the framework achieves EER ≤ 0.001 and TPR > 99.5% at FPR < 0.1% under leave-user-out cross-validation against the strongest parametric generator we could construct.

We evaluate adversarial robustness through a 5-round Bayesian optimization loop in which the attacker has white-box access to the feature set, unlimited queries, and a 17-dimensional parameter space. After a single round of detector retraining, the attacker's mean evasion score drops from 0.999 to 0.010 by round 5. Feature-family ablation reveals that this robustness arises from a tradeoff constraint: the parametric generator's single parameter set cannot simultaneously match all feature families, whereas the corresponding properties of human movement arise from distinct neuromuscular mechanisms.