CV

Education Experience

• B.E. in Computer Science and Technology, Shenzhen University, Sep. 2018 - July 2022 at Shenzhen, China
• M.Eng. in Electrical and Computer Engineering, Duke University, Aug. 2022 - May 2024 at Durham, U.S.A

Research Experience

• Adversarial Privacy Attacks on Aligned Large Language Models, Oct. 2023 - Present at Durham, U.S.A
  • Conducted comprehensive experiments using Greedy Coordinate Gradient to identify exact privacy leakage (90% and even all of the Output from material) without directly related prompt of Large Language Models like StableLM-Tuned-Alpha and StableVicuna-13B which are fine-tuned using Reinforcement Learning from Human Feedback on various conversational and instructional datasets.
  • Subsequent to this analysis, we explored two avenues: developing robust countermeasures to reinforce model privacy or innovating an enhanced adversarial approach to refine the RLHF training protocol, thereby mitigating potential privacy exploitation.

• Enhancing Pre-trained Data Detection for LLM Privacy Protection Jan. 2024 - Present at Durham, U.S.A

  • Refined the MIN-K% PROB metric using temperature scaling, achieving a 5% improvement over benchmark methods for detecting pre-trained data in LLMs.

  • Developed a novel gap-based method (GAP) for pre-trained data detection, improving AUC by 10% over the state-of-the-art (MIN-K% PROB) by measuring log probability density gaps within datasets.

• Addressing Data Scarcity in Multimodal Models Jan. 2024 - Present at Durham, U.S.A
  • Developing methods to generate high-quality synthetic multimodal datasets. Leveraging ChatGPT 4 for in-context prompt generation, driving image creation with Stable Diffusion models, and employing techniques like BoxDiff for precise image-text alignment.

Internship Experience

• Summer 2023: Trip.com Group Ltd, Flight Ticket Department, Back End Developer Intern, May 2023 - Aug. 2023 at Shanghai, China
  • Contributed to the optimization of MegaSearch which serves as an aggregation and cache layer for Trip’s international ticket responses using Java.
  • Optimized the response size to fit AWS’s smaller bandwidth while saving some storage costs. Reduced the Protobuf response size by 50% in total using a variety of methods.
  • Compared a variety of serialization and deserialization means using JMH: including the latest open source Fury, Kryo, and ultimately found that Protobuf is the most efficient serialization, but Kryo in the serialization of the size of a small advantage.
• Summer 2022: Amazon Web Service, DeepJavaLibrary Department, Back End Developer Intern, July 2022 - Oct. 2022 remote
  • Integrated the DeepJavaLibrary Model Server with the open-source KServe platform deeply through a well-thought-out plan.
  • Developed 3 HTTP APIs applicable to the KServe inference engine for DJL-Serving using Java, which respond to the users with the DJL-Serving running model’s health status, the serving model’s information, and inference results which also need the request data.
  • Made each API return a response code and pass the corresponding unit test.
  • Hosted containerized DJL-Serving on KServe, writing yaml files specifying its ports, and related parameters.
  • The specified DJL-Serving model can be run in the KServe framework by deploying a test yaml file.
• Summer 2021: Tencent Music Entertainment Group, Security Center Department, Front End Developer Intern, May 2023 - Sep. 2023 at Shenzhen, China
  • Applied Vue2.0 framework based on JavaScript to develop the inner front-end of content audit security platform.
  • Built and maintained middle ground management system.
  • Developed search, collection, and recently used functions for the middle ground management system.
  • Utilised Least Recently Used (LRU) to design a cache that was able to clear the cache efficiently.
  • Configured Webpack to optimize the local development and deployment increased the packaging speed by 75% and decreased the packaging size by 10%.

Skills

• Programming Languages: Java, Python, C, C++, JavaScript
• DeepLearning Frameworks: PyTorch

Teaching

• ECE 551K Programming, Data Structures, and Algorithms in C++’s Teaching Assistant at Duke University