VR/metaverse headsets stream continuous 6-DoF pose data (head + two hand controllers) at 60–120 Hz. Instead of authenticating once at login, a continuous behavioral-biometric layer treats this motion stream itself as a signature: a sliding window of pose samples is reduced to a feature vector
f = [ v̄, σ_v, σ_tilt, r̄_reach, 1/T_period, v_inst ]
(mean speed, speed variance, head-tilt jitter, average hand-reach radius, dominant reach-cycle rate, and instantaneous speed) — a compact summary of *how* a specific person moves, not just *what* they do.
On enrollment the platform stores a baseline vector f₀. Every window afterward is compared with cosine similarity:
match = (f · f₀) / (‖f‖ ‖f₀‖)
anomaly = 1 − match
session locked if match < threshold
- Re-enroll baseline — captures a fresh f₀ from the current (legitimate) motion pattern.
- Spoof / replay attack — swaps the live stream for a different motion signature (a different person, a bot script, or a replayed recording), simulating an impostor trying to ride an already-open session.
- Spoof deviation — how far the attacker's motion parameters are pushed away from the enrolled user's own values (see note below).
- Match threshold — the similarity floor the detector accepts; raising it catches subtler impersonation but also legitimate motion drift (fatigue, different controller grip).
- Tracking noise — sensor/tracking jitter added to every sample, showing why thresholds can't be set to 1.0.
Note on this rebuild: the original 3D engine generated the attacker's motion profile from an independent random seed and only widened its own internal randomness range as "deviation" increased — two different seeds are already dissimilar regardless of the slider, so the control wasn't a reliable monotonic knob on distance-from-baseline. Here the attacker profile is built as the enrolled legit profile's own parameters plus a signed offset scaled directly by the deviation slider, so "spoof deviation" now deterministically controls how far the impostor's signature sits from the real one.
This is the same principle used by real continuous-authentication research for XR headsets (motion/gait signatures identifying users from head and controller kinematics alone) and is purely defensive: it flags impersonation of an already-logged-in avatar, it is not a way to fabricate one.