Research
Privacy-preserving wearable vision-language models
November 2025–present · Advisor: Jingtong Hu
I study privacy risks in wearable AI, where intermediate visual representations can reveal sensitive attributes even when final answers appear harmless. I developed a split-inference approach that transforms visual tokens before they leave the trusted device. In a 3,221-record privacy–utility benchmark, private-attribute inference accuracy fell from 56.7% to 7.4% while relaxed utility remained at 74.4%.
I also developed SAGE, a hardware-aware approach to protected inference on real smart glasses. It lowers visual input cost, runs shape-specialized INT8 graphs, and sends compact protected packets. On a matched Qwen workload, SAGE reduced glasses inference latency by 39.18% and protected packet size by 96.56%.
Fairness in AI-assisted mobile healthcare
November 2023–November 2025 · Advisor: Jingtong Hu
I worked on medical image synthesis and downstream diagnosis with demographic balance and clinician-informed preferences. FairGen improved fairness by 95.9% on skin images, 80.0% on chest radiography, and 35.2% on brain MRI while maintaining competitive diagnostic accuracy.
Efficient real-time object detection on the edge
March–November 2023 · Advisor: Yanzhi Wang
I contributed to a deployment framework that coordinates video keyframe selection, CPU/GPU dynamic voltage and frequency scaling, and DNN pruning. It maintained object-detection accuracy within 0.28% mAP of processing every frame while reducing system power by up to 96.7%.
Digital avatars on mobile devices
August–November 2023 · Advisor: Yanzhi Wang
I built a pipeline connecting speech recognition, LLM reasoning, text-to-speech, facial animation, and audio-video streaming for responsive mobile digital humans.
DNN data management on hybrid memory systems
February–September 2021 · Advisor: Dazhao Cheng
I studied placement and migration of DNN data across memory tiers, contributing to performance evaluation using ReDL.