Human-Centered Computing

A VR headset renders the wearer's hands, so anything watching can read what they type. The leading defense injects decoy keystrokes to bury the real ones. Separatrix shows a prompted decoy is a cued action: it is read, located, and reached for, so it lands slower and bends off the smooth path a fluent press follows. Scoring each press on timing, reach geometry, and language, then decoding with one beam search, Separatrix recovers the secret in its top twenty guesses for most trials where a decoy-unaware attacker recovers nothing. It is the full version of Winnow (XR Security '26), in submission to USENIX Security. Full write-up on the Separatrix project page.

SignSense combines a lightweight vision model with an edge LLM to translate sign language locally. It supports real time accessible communication while keeping biometric signing data on device.

Sign Language Atlas organizes three decades of sign language recognition and translation systems across sensing, vision, neural models, privacy, and edge deployment.

Relate studies moral consideration through relational capacity and embodied interaction rather than unverifiable claims about machine consciousness. It connects posthumanist theory to AI governance.