Yuxin "Myles" Liu

Research

"Seeing is believing" no longer holds in an era saturated with generative AI and autonomous agents. I build full-stack systems that restore trust by establishing a verifiable chain of custody, from raw physical capture to consumption and actuation. My work is grounded in digital provenance and extends that principle to physical AI, where safe autonomy demands verifiable sensor data, authenticated commands, and contract-bound execution.

Verifiable Digital Provenance

Can we trust what a camera saw? I build provenance systems that let anyone verify where media and data came from, from trusted computing bases inside the camera to defenses against recapture attacks and provenance for machine learning data.

ProvCam FPGA prototype with camera sensor board — ProvCam FPGA prototype

3D concept render of ProvCam as a stacked camera module — ProvCam ASIC concept (render)

Scoop pipeline: a captured photo's learning-based point cloud is compared with the ground-truth depth point cloud, and the flat recaptured region is highlighted in red — Scoop in action: comparing learned and measured depth exposes a recaptured photo

Scoop: detecting recapture attacks, where fake footage shown on a screen is filmed by a provenance-enabled camera. USENIX Security'25 · ACM GetMobile Research Highlight [Paper] [GetMobile Highlight] [Code] [Poster]
ProvCam: a camera module with a self-contained trusted computing base that produces verifiable videos. MobiCom'24 [Paper] [Code]
Vronicle: verifiable provenance for videos captured and edited on mobile devices. MobiSys'22 · Best Poster Award Runner-up [Paper] [Code] [Poster]
Ongoing Semantic provenance for live video chat; confidential, verifiable provenance for machine learning data.

Trustworthy Physical AI

As robots driven by foundation models move into public spaces, they must verify whom they obey and bound what they do. I bring provenance and hardware-rooted trust to robot perception, command, and actuation.

Quadruped robot testbed fitted with secure RF ranging boards, a microphone array, and a LiDAR — RoboAuth quadruped testbed

RoboAuth: authenticating emergency robot overrides by cross-checking secure RF ranging, depth sensing, and acoustics, each signed at capture. ACSAC'26 Poster [Poster Proposal]
Ongoing Contract-bound safety enforcement for robots driven by foundation models.

Hardware & OS Foundations

The trusted hardware and operating system mechanisms that the work above builds on.

Patchlings: safety-preserving hotpatching for real-time automotive microcontrollers, in collaboration with Bosch Research. arXiv [Paper]
UnderRansom: real-time ransomware detection and mitigation in the OS kernel, developed at Bosch Research.
Tabellion: secure legal contracts on mobile devices. MobiSys'20 [Paper] [Code]
Ongoing A physically isolated TEE design with high performance and utilization.