Year: 2026
Type: dissertation
Abstract: The Metaverse is widely understood as a persistent, shared, immersive virtual space that integrates physical and digital realities. As the Internet evolves toward this paradigm, the convergence of spatial computing and machine learning is reshaping human-computer interaction. The resulting ecosystem rests on two complementary capabilities: a spatial interaction layer that captures continuous, embodied user kinematics, and a cognitive intelligence layer that synthesizes this data through predictive and generative models. Together they enable immersive experiences but also introduce system-level risks. Without careful design, the Metaverse may evolve into a pervasive sensing and inference infrastructure with expanded attack surfaces, opaque data practices, and entrenched centralization across both the spatial and cognitive layers. This dissertation reexamines the security and privacy foundations of Metaverse applications around this dual-pillar architecture. Rather than relying on reactive, point-wise defenses, it advocates for a proactive approach that prioritizes trustworthiness at the architectural level. For the spatial interaction layer, we bridge vulnerability discovery with structural reform. We empirically uncover a novel class of side-channel threats showing that seemingly benign VR kinematics expose sensitive user intentions. To understand why such vulnerabilities persist, we conduct an in-depth interview study with 23 professional XR developers and surface ecosystem-level root causes. Finally, instead of treating decentralization as a binary switch between keeping the central server and removing it entirely, we redistribute each role the central operator plays—identity, shared state, behavior auditing—to system participants one role at a time. We ground this design in a 22-participant interview study of user-stated needs, and a 20-participant within-subjects double-blind comparison against a centralized baseline returns no evidence of a perceptual gap on the constructs participants themselves consider load-bearing for a usable collaborative experience. For the cognitive intelligence layer, we establish rigorous privacy boundaries using model-centric attacks. We show that predictive spatial-temporal models leak precise user trajectories, and that fine-tuned zero-shot Text-to-Speech models expose biometric voice features at both speaker-level identity and record-level verbatim granularities, with attack performance grounded in an information-theoretic characterization of the optimal query. Together, these contributions move the Metaverse toward an architecture that is secure, accountable, and user-sovereign by design.
Author: Cai, Kunlin
Institution: University of California, Los Angeles
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7171480670