Kinematic Features Are All You Need: Detecting Synthetic Mouse Trajectories Under Adversarial Optimization

We present a 17-feature kinematic detection framework for identifying synthetic mouse trajectories, evaluated against parametric generators including SigmaDrift: a motor-control-grounded generator we designed as the strongest parametric adversary feasible. On two public mouse dynamics datasets (29 users, 43,216 human trials), the framework achieves EER ≤ 0.001 and TPR > 99.5% at FPR < 0.1% under leave-user-out cross-validation. A 5-round Bayesian optimization adversarial loop with white-box feature access fails to produce sustained evasion, with the attacker's mean evasion score converging to 0.010 after detector retraining. Feature-family ablation reveals a structural tradeoff constraint: parametric generators cannot simultaneously satisfy all four feature families (Fitts compliance, submovement morphology, kinematic smoothness, geometry) because these properties arise from distinct neuromuscular mechanisms in human movement. We additionally identify and exclude 15 confounded features from polling-rate artifacts, establishing a cleaner evaluation methodology for mouse dynamics research.