Projects
Research projects that I have lead and/or contributed to.
EvaluatAR
Abstract: Augmented-reality headsets continuously sense their surroundings, creating privacy risks for nearby bystanders. EvaluatAR is a headset-agnostic framework that standardizes the sensor inputs, visual stimuli, and logged outputs used to evaluate visual bystander privacy-enhancing technologies. Its record- and-replay workflow lets researchers reproduce the same privacy-relevant conditions across devices and trials, compare privacy–performance trade-offs, and revisit edge cases without repeatedly recreating live human interactions. Three case studies demonstrate cross-device evaluation, support for both implicit and explicit privacy mechanisms, and rapid debugging of failures under identical inputs.
Rethinking Privacy Indicators in Extended Reality
Abstract: Conventional XR privacy indicators, such as small recording LEDs, assume that nearby bystanders are attentive and able to notice visual cues. This project introduces the concept of situationally impaired bystanders—people who may be distracted, visually limited by the situation, or otherwise unable to perceive a conventional indicator. Four participatory focus groups with eight participants generated candidate designs, which were consolidated into five indicators spanning visual, auditory, haptic, proximity-, gaze-, and activity-triggered cues. A scenario-based evaluation with seven additional participants found that visual-only indicators were often judged less useful in impairment scenarios, while adaptable multimodal indicators were preferred in privacy-sensitive contexts.
Security Attacks in Remote Collaborative Mixed Reality
Abstract: Remote collaborative mixed reality introduces attacks that exploit three-dimensional interaction and invisible virtual geometry. This project implemented latency, click-redirection, object-occlusion, and spatial-occlusion attacks in a collaborative HoloLens 2 environment and studied how users recognized, interpreted, and responded to them. Participants were generally less likely to recognize immersive attacks that exploited MR-specific interaction mechanics than attacks resembling familiar desktop- security failures. The findings identify a gap in user awareness and motivate clearer system feedback, training, and mitigations for maintaining trust in remote immersive collaboration.
Visceral Interfaces for Eye-Tracking Privacy Awareness in VR
Abstract: Eye tracking supports interaction, rendering optimization, and adaptive experiences in VR, but gaze data can also reveal identity, behavior, and sensitive personal attributes. This project designed three visceral, in-experience interfaces intended to make eye-tracking data collection and its implications more understandable than conventional terms-of-service notices. A user study evaluated how people perceived these interfaces. Although some visualizations could be distracting, participants generally found them informative and supported making such privacy-awareness mechanisms available, even when not enabled continuously by default.
Short-Form Content Consumption in Virtual Reality
Abstract: This project investigates how immersive presentation changes the experience of consuming short-form media. A within-subjects study compared personalized short-form content in VR and on a smartphone using counterbalancing, VR acclimation, a modified Facebook Flow Questionnaire, and perceived-time measures. The analysis found higher mean flow and stronger time distortion in VR, suggesting that immersive short- form interfaces may intensify engagement and reduce awareness of elapsed time. These results have implications for safer, more transparent, and more user-aware immersive media design.
TaxData
Abstract: TaxData combined remote sensing, computer vision, and a user-facing web platform to help tax officials identify constructed land from satellite imagery. The project assembled approximately 30,000 Google Earth images, including 3,000 manually labeled examples, and trained a constructed-versus-non-constructed classifier that achieved 92% accuracy. The resulting interface allowed officials to define regions of interest and inspect current model outputs, reducing reported survey time by 70%.