Vulnerability of machine learning models for gender recognition in Virtual Reality

Virtual Reality (VR) systems continuously capture fine-grained behavioral signals such as head motion, hand trajectories, and gaze dynamics. These spatio-temporal signals have been shown to contain distinctive patterns enabling accurate gender classification through machine learning models. While predictive performance under nominal conditions is often high, the robustness of such models to structured behavioral perturbations remains largely unexplored. In this paper, we present a systematic robustness analysis of VR-based gender classification models under a comprehensive catalog of realistic behavioral adversarial attacks. We evaluate multiple model families, including ensemble-based tabular classifiers and neural architectures, using statistical and dynamic motion features extracted from public VR datasets. More than one hundred perturbation scenarios targeting metric coherence, global motion style, multimodal synchronization, and latent behavioral structure are assessed using balanced accuracy, flip rate, and confidence stability metrics. Our results reveal significant vulnerability to coordinated, structurally consistent attacks, particularly those affecting global motion properties or metric integrity, while localized noise-like perturbations exhibit limited impact. These findings demonstrate that high nominal accuracy does not guarantee robustness and highlight the necessity of robustness-aware evaluation frameworks for VR-based behavioral inference systems.