Supervised by Prof. Lifeng Lai, UC Davis.

Graduate Student Researcher in Lai's group.

• Studied preference robustness of RLHF algorithms with online human feedback from a theoretical perspective.
• Investigated preference attacks that force online RLHF to learn a target suboptimal policy with small attack cost.
• Proposed an adversarial human-feedback poisoning strategy and proved its success in misleading online RLHF.
• Proposed a robust-defense online RLHF algorithm and proved its robustness to any attacker with bounded attack cost.
• Simulation results validate the theoretical analysis. Published in IEEE TSP 2025 [DOI].

Supervised by Prof. Lifeng Lai, UC Davis.

Graduate Student Researcher in Lai's group.

• Studied reward attacks on stochastic multi-armed bandits in non-stationary environments.
• Investigated attackers who force the victim algorithm to mostly choose a suboptimal arm with small attack cost.
• Considered three general scenarios with different environments, algorithms, and attacker information.
• Proposed three attack strategies and proved their success in terms of target-arm selection and attack cost.
• Proposed a non-stationary bandit defense method and proved its robustness to any attacker with bounded attack cost.
• Simulation results validate the theoretical analysis. Accepted to ACSSC 2023 [DOI]. Published in IEEE TSP 2024 [DOI].

Supervised by Prof. Xiaofan (Fred) Jiang, Columbia University.

Summer Research Assistant in Columbia Intelligent and Connected Systems Lab.

• An auto-discover system able to allocate sensors and actuators according to the smart home applications specified by users; deployed in multiple lab rooms.
• Designed the MCU/sensors-to-application architecture and built the hardware testbed from scratch.
• Testbed: Arduino Uno reads custom sensor data (vibration, sound, etc.) and transmits them to ESP8266 via UART. ESP8266 packs data with DT format and MQTT discovery format, and sends them to a lab computer server (HTTP with Flask) and Home Assistant Raspberry Pi server (MQTT). Integrated Google Nest camera into HA and stored data in PostgreSQL and MariaDB.
• Accepted to SenSys 2021 [DOI].

Supervised by Prof. Xiaofan (Fred) Jiang, Columbia University.

Research Intern in Columbia Intelligent and Connected Systems Lab.

• A low-cost system based on RGB-thermal cameras for continuous fever screening of multiple people without human interaction; deployed in a restaurant and hospital clinic in New York City. Featured in Columbia News.
• Implemented multiple RGB-thermal heads matching, trained and deployed YOLOV3 (head detection) and FSA-Net (head orientation regression), and estimated distance using non-identical RGB and thermal camera.
• Accepted to SenSys 2020 [DOI], CPHS 2021 [DOI], and IPSN 2022 [DOI].
• Project website: ICSL Fever Screening.

Supervised by Prof. Eugene Wu, Columbia University.

Research Intern in WuLab Columbia University.

• A server-side scheduler involving a complex optimization based on available resources, predicted user interactions, and response quality levels to maximize user-perceived interactivity and satisfaction in real-time.
• Created the simulated RL environment, which can pre-compute transition relationships or compute them dynamically at run time.
• Implemented Q-learning and SARSA-based prefetching schedulers to trade off latency and response quality with progressive encoded responses in cloud-based interactive applications.
• Added predicted user-preferred action choices to the algorithms above.
• Project GitHub: Khameleon-Scheduler-RL.

Supervised by Prof. Jiang Wu, Xi'an Jiaotong University.

Member of XJTU Information-technology Talent Program.

Research Intern in Ministry of Education Key Lab for Intelligent Networks and Network Security.

• Proposed centralized and distributed K-means methods using weighted combinations of PCA features and prior knowledge, plus parameter consensus and feature-transfer methods.
• Analyzed urban power grid data and contributed to paper writing.
• Accepted to Chinese Control Conference 2018 [DOI].