Multi-user XR apps let people meet, play, and work together in shared virtual spaces. To meet stringent latency requirements, many apps are client-authoritative: each client computes and broadcasts its own state, which the server relays without validation. We presented the first systematic analysis of the security risks of this client-side trust (USENIX Security 2026). Across 20 popular apps on four platforms with 150M+ combined installs, we found exploitable trust boundaries in every app and demonstrated eight attacks spanning surveillance, object tampering, safety circumvention, disruption, and impersonation. To enable auditing of this attack surface at scale, we designed an agentic framework in which an LLM coding agent locates and subverts each app’s client-side enforcement logic (XR Security Workshop 2026, Best Paper Award). We also explored how users experience client modifications in practice and how they perceive the defenses platforms deploy against them, through a large-scale analysis of Reddit discussions (ACM CCS 2026).
XR headsets increasingly integrate eye tracking for foveated rendering, gaze-based interfaces, and adaptive content. Yet gaze is a continuous, largely subconscious signal that can reveal far more than where a user is looking. We showed that eye-tracking data from commercial XR headsets can be exploited to infer user activity, classifying which of seven video genres a user is watching from gaze alone with 75% average accuracy on unseen videos (MobiSys 2025, Best Poster Runner-Up). Gaze-driven systems also rely on scanpath prediction models, which forecast where users will look and are often fine-tuned from public datasets or third-party pretrained weights, exposing them to training-time poisoning attacks. We presented the first backdoor attacks on scanpath prediction models, which redirect the predicted search toward an attacker-chosen object or prolong it with extra fixations, reaching up to 93.5% attack success with as little as 2.5% of the training data poisoned (WACV 2027).
Emerging technologies such as conversational AI and social XR introduce new security and privacy risks, but how users perceive and respond to these risks often remains unexplored. We study this at scale through online user discourse, leveraging applied machine learning and qualitative analysis techniques. We presented one of the first works on users’ security and privacy concerns about conversational AI platforms, where we found that users doubt their data is ever truly deleted once used for training, fear it may be regurgitated to other users, and question whether using these platforms for medical or legal work violates confidentiality obligations such as HIPAA. Users also report dark patterns, such as default opt-in to model training and withholding features from those who opt out, and their concerns spike after events like Microsoft’s investment in OpenAI and a bug that exposed other users’ conversation titles (IEEE S&P 2025). We also studied the harms users experience from modified clients in social XR and their perceptions of the defenses platforms deploy against them, finding that users perceive avatar theft as a violation of their identity and report that attackers bypass blocking and other safety features to stalk and harass others. User attitudes toward anti-cheat engines are mixed: some give them credit for reducing harms after deployment, while others frame them as a contested defense whose tradeoffs (privacy, performance, usability) fall disproportionately on users who never installed a mod (ACM CCS 2026).
Latency-sensitive mobile applications such as AR/VR and cloud gaming rely on edge computing, but suffer service interruptions when users move and their session state must migrate to a new edge site. We designed EdgeWarp, which proactively mirrors application state at edge sites predicted to serve the user next and prioritizes latency-critical flows during 5G handovers, cutting mobility-induced downtime by up to 15.4× (ACM/IEEE SEC 2025).
I have also built or contributed to a number of other cool projects, ranging from a federated learning testbed of real mobile devices to a Raspberry Pi sound alert system for deaf and hard-of-hearing users, which you can find on my GitHub.