Who Are You in Virtual Reality? Suppressing Demographic Leakage via Progressive Adversarial Privacy

Virtual reality (VR) systems continuously capture multimodal sensor streams, including head motion, gaze direction, and physiological signals, to enable different interactions and immersive experiences. However, this multimodal telemetry carries persistent biometric signatures that adversaries can exploit to re-identify users and infer sensitive demographic attributes such as age and gender, even from anonymized datasets. Existing privacy-preserving approaches, including noise injection and anonymization, degrade the temporal structure of the data, which is essential for different downstream utilities (i.e., application utility) in VR. To address this gap, we present a Progressive Privacy Preserving method which is designed specifically for multimodal VR telemetry while maintaining downstream utility. Rather than optimizing against all demographic attributes jointly, our method separates physiological and kinematic feature domains and applies adversarial optimization sequentially. This progressive strategy prevents gradient interference between demographic traits that originate from distinct physiological signal sources, enabling targeted suppression of sensitive attributes while preserving the behavioral patterns required for application utility. We evaluated our proposed method on three publicly available VR datasets: Mazed and Confused, Simulation21, and WhoIsAlyx. Results show that demographic inference accuracy drops to near random chance (approximately 50% for binary attributes), while downstream utility task performance, including cybersickness severity classification, cognitive load classification, physical load classification, and action intent classification, remains strong, with reductions of at most 14.5% compared to the unprotected data. On average, the method modifies only approximately 11% of the original signal magnitude, confirming that the underlying temporal structure of the datasets remains largely intact. Collectively, our findings demonstrate that sensitive demographic attributes embedded in VR data can be substantially suppressed through progressive adversarial optimization while retaining the signal fidelity necessary for downstream utility.