I am a postdoctoral fellow advised by Dr. Zhangyang Wang at the Institute for Foundations of Machine Learning (IFML) and the Wireless Networking and Communications Group (WNCG) at UT Austin. I obtained my Ph.D. in Computer Science and Engineering from Michigan State University, where I was advised by Dr. Jiayu Zhou. I hold a B.S. in Physics and an M.S. in Computer Science from the University of Science and Technology of China. I was honored as one of the MLCommons Rising Stars in 2024 and VLDB 2024 best paper finalist.
My long-term research vision is to establish Holistic Trustworthy AI for Healthcare.
My recent research is driven by the emergent challenges in AI for Dementia Healthcare, and centers around Privacy-Centric Trustworthy Machine Learning toward Responsible AI where I pursue fairness, robustness, security and inclusiveness under privacy constraints.
Most of my work (including GenAI GAI) follows these principles to trade-off efficiency, utility and privacy: edge-edge (federated learning) FL and edge-cloud (pre-training fine-tuning, transfer learning) EC collaboration.
- Dementia [Healthcare] β¨ Generative AI:
- Disease modeling and from distributed devicesFL.
- Conversational AI for dementia simulation and prevention GAI.
- [Privacy] β¨ Collaborative ML β¨ Large Vision/Language Models:
- Tools: Efficient auditing tools, and adaptive private learning.
- Foundation Models: Privacy risks of diffusion modelsEC and private offsite prompt tuning of ChatGPTEC GAI.
- Applications for distributed vision, finance and disease modeling.
- [Trustworthy ML] β¨ Privacy:
- Benchmark the trustworthiness of compressed LLMs and privacy GAI.
- AI Safety via automatic guardrail GAI.
- Fair unsupervised trainingEC and multi-group learningFL.
- Robustness acquired from collaboration, against adversarial samplesFL and OoD inferenceFL.
- Security in data-free distillationFL and model copyrightEC.
- Inclusive efficient system for for low-end devices by offsite prompt tuningECGAI, on-device model adaptationEC, private outsourcing trainingEC and customizable modelsFL.
Check my curricula vitae and feel free to drop me an email if you are interested in collaboration.
News
- August, 2023 π Our privacy benchmark paper (LLM-PBE) gets into the best paper finalist at VLDB 2024.
- July, 2024 π₯ We are exciting to organize the GenAI for Health: Potential, Trust and Policy Compliance workshop (GenAI4Health 2024) at NeurIPS 2024.
- June, 2024 π New benchmark on LLM privacy is accepted to VLDB!
- June, 2024 New paper on safeguarding LLM agent is online!
- June, 2024 π Thrilled to co-organize The LLM and Agent Safety Competition 2024 at NeurIPS 2024!
- June, 2024 π Grateful to receive API grants from the OpenAIβs Researcher Access Program!
- May, 2024 π Thrilled to co-organize The NeurIPS 2024 LLM Privacy Challenge! Join us for the competition!
- May, 2024 π I am honored to be selected one of ML and Systems Rising Stars by ML Commons for my work in health and trustworthy ML, 2024!
- May, 2024 π Two benchmark papers are accepted at ICML 2024: how to obtain trustworthy compressed LLMs (models at huggingface) and how to optimize LLMs with less memory.
- April, 2024 π€ I am honored to give a invited talk on conversational AI for dementia health at UT school of information.
- March, 2024 π€ I am honored to give a talk on the new privacy risk of GenAI at UT Good System Symposium 2024.
More
- March, 2024 We are exciting to organize the International Joint Workshop on Federated Learning for Data Mining and Graph Analytics (FedKDD 2024).
- March, 2024 Our benchmark work, Decoding Compressed Trust, has been accepted to SET LLM @ICLR. A curated set of compressed models are available at huggingface.
- Feb, 2024 New benchmark preprint on zeroth-order optimization for LLMs.
- Jan, 2024 π Three papers are accepted: The first local privacy-preserving prompt tuning as Spotlight at ICLR, robust watermarking from one image as poster at ICLR, the generalization of unsupervised pretraining at AISTATS!
- Dec, 2023 πΎ Our paper on amplifying privacy risks via fine-tuning (Shake-To-Leak) is accepted to SaTML.
- Nov, 2023 π Grateful to be selected as Top Reviewer at NeurIPS 2023.
- Dec, 2023 Our new preprint on private prompt engineering for close-source LLMs is online.
- Dec, 2023 βοΈ π· I will be at New Orleans for presenting our recent work on understanding gradient privacy (NeurIPS'23 ) and tracking IP leakage in FL (NeurIPS-RegML).
- Nov, 2023 π€ We are releasing a set of compressed LLMs at compressed-llm for public benchmarks.
- Nov, 2023 Our work on tracking IP leakage in FL is accepted to NeurIPS'23 Workshop on Regulated ML (NeurIPS-RegML).
- Sep, 2023 Our work on understanding gradient privacy via inversion influence functions is accepted to NeurIPS'23.
- Sep, 2023 Our new work on watermarking models using one image is online.
- August, 2023 π₯ We are organizing a KDD workshop on federated learning for distributed data mining (FL4Data-Mining) on August 7th at Long Beachπ΄.
- July, 2023 I am going to travel for ICML 2023 at Hawaii πΊ. Come and talk to me about data-free backdoor!
- July, 2023 π Honored to receive Research Enhancement Award for organizing FL4DataMining workshop! Thank you to MSU Graduate School!
- July, 2023 π I successfully defended my thesis. Many thanks to my collaborators, advisor and committees.
- May, 2023 My new website is online with released junyuan-academic-theme including many cool new features.
- April, 2023 One paper on data-free backdoor got accepted to ICML'23.
- March, 2023 π Our ILLIDAN Lab team just won the 3rd place in the U.S. PETs prize challenge. Media cover by The White House, MSU EGR news and MSU Office of Research and Innovation.
- Jan, 2022 Two papers got accepted to ICLR'23: OoD detection by FL (splotlight!), memory-efficient CTA.
- Sep, 2022 Our work on federated robustness sharing has been accepted to AAAI'23 (oral).
- Nov, 2022 Two papers got accepted to NeurIPS'22: outsourcing training, backdoor defense.
- May, 2022 Our work on connection-resilient FL got accepted to ICML'22.
Interests
- Healthcare
- Privacy
- Trustworthy Machine Learning
- Federated Learning
- Large Language/Vision Models
Education
-
PhD in CSE, 2023
Michigan State University
-
MSc in Computer Science, 2018
University of Science and Technology of China
-
BSc in Physics, minor in CS., 2015
University of Science and Technology of China
Projects
.js-id-Selected
DiRP Trustworthy LLM
Directed Reading Program (DiRP) on trustworthy large language models.
Holistic Trustworthy ML
Instead of isolated properties, we target on a holistic trustworthiness covering every properties in one solution.
Federated Learning
On the need of data privacy and more data, we strive to join the knowledge from a fair amount of users to train powerful deep neural networks without sharing data.
Privacy in Collaborative ML
On the concern of data privacy, we aim to develop algorithms towards learning accurate models privately from data.
AI for Dementia Healthcare
We aim to early detect and intervene dementia diseases leveraging the power of (Generative) AI.
Subspace Learning
Supervised learning on subspace data which could model real data like skeleton motion.
Publications
.js-id-Selected
VLDB (Best Paper Finalist) 2024
LLM-PBE: Assessing Data Privacy in Large Language Models.
PDF Website π Competition
PDF Website π Competition
ICML 2024
Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression.
PDF Models Website
PDF Models Website
ICML 2024
Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark.
PDF Code π¨βπ«Tutorial
PDF Code π¨βπ«Tutorial
LLM Agents 2024
A-CONECT: Designing AI-based Conversational Chatbot for Early Dementia Intervention.
PDF Website π€Demo
PDF Website π€Demo
SaTML 2024
Shake to Leak: Fine-tuning Diffusion Models Can Amplify the Generative Privacy Risk.
PDF Code
PDF Code
ICLR (Spotlight) 2024
DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer.
PDF Code
PDF Code
NeurIPS-RegML 2023
Who Leaked the Model? Tracking IP Infringers in Accountable Federated Learning.
PDF
FL4DM 2023
A Privacy-Preserving Hybrid Federated Learning Framework for Financial Crime Detection.
PDF Code π PETs ποΈ White House
PDF Code π PETs ποΈ White House
TMLR 2023
How Robust is Your Fairness? Evaluating and Sustaining Fairness under Unseen Distribution Shifts.
PDF
ICLR (Spotlight) 2023
Turning the Curse of Heterogeneity in Federated Learning into a Blessing for Out-of-Distribution Detection.
PDF Code
PDF Code
AAAI (Oral) 2023
Federated Robustness Propagation: Sharing Adversarial Robustness in Federated Learning.
Preprint Code Poster
Preprint Code Poster
NeurIPS 2022
Outsourcing Training without Uploading Data via Efficient Collaborative Open-Source Sampling.
PDF Poster Slides
PDF Poster Slides
NeurIPS 2022
Trap and Replace: Defending Backdoor Attacks by Trapping Them into an Easy-to-Replace Subnetwork.
PDF Code
PDF Code
FAccT 2022
Dynamic Privacy Budget Allocation Improves Data Efficiency of Differentially Private Gradient Descent.
Preprint Slides
Preprint Slides
ICLR 2022
Efficient Split-Mix Federated Learning for On-Demand and In-Situ Customization.
PDF Code Slides Video
PDF Code Slides Video
AAAI 2021
Learning Model-Based Privacy Protection under Budget Constraints.
PDF Slides Video Supplementary
PDF Slides Video Supplementary
AD 2020
Detecting MCI using real-time, ecologically valid data capture methodology: How to improve scientific rigor in digital biomarker analyses.
PDF
TKDD 2019
Variant Grassmann Manifolds: A Representation Augmentation Method for Action Recognition.
PDF DOI
PDF DOI
KDD (Oral) 2018
Disturbance Grassmann Kernels for Subspace-Based Learning.
PDF Cite Video DOI Supplementary
PDF Cite Video DOI Supplementary
Featured Publications
We develop a chatbot for early dementia prevention and leverage LLMs to build digital twins to evaluate chatbots.
We make local LLMs to engineer privacy-preserving prompts that are transferrable for cloud models.
We develop a hybrid federated learning for learning financial-crime predictive models from horizontal and vertical federated data structures.
We propose a new privacy-preserving learning framework, outsourcing training to cloud without uploading data, which provides more data without injecting noise into gradient or samples.
Efficient and federated learning for heterogeneous clients with different memory sizes
A distributed domain/group debiasing framework for unsupervised domain adaptation or fairness enhancement.
Professional Activities
Experience
- Postdoctoral Fellow with Dr. Zhangyang Wang, VITA group@UT Austin, IFML and WNCG, 2023-Now.
- Research Intern with Dr. Lingjuan Lyu, Sony AI, 2022
Awards
- Best Paper Finalist, VLDB 2024
- ML and Systems Rising Stars, ML Commons, 2024
- Top Reviewer, NeurIPS, 2023
- Research Enhancement Award, MSU, 2023
- The 3rd place in the U.S. PETs prize challenge, 2023
- Dissertation Completion Fellowship, MSU, 2023
- Carl V. Page Memorial Graduate Fellowship, MSU, 2021
Fundings
I am grateful that our research is supported by the multiple programs.
- OpenAI Researcher Access Program early-dementia prevention via A-CONECT chatbot.
- IFML Postdoctoral Fellowship for my research in trustworthy LLMs.
Media Coverage
- At Summit for Democracy, the United States and the United Kingdom Announce Winners of Challenge to Drive Innovation in Privacy-enhancing Technologies That Reinforce Democratic Values, The White House, 2023
- Privacy-enhancing Research Earns International Attention, MSU Engineering News, 2023
- Privacy-Enhancing Research Earns International Attention, MSU Office Of Research And Innovation, 2023
Talks
- ‘Building Conversational AI for Affordable and Accessible Early Dementia Intervention’ @ AI Health Course, The School of Information, UT Austin, April, 2024: [paper]
- ‘Shake to Leak: Amplifying the Generative Privacy Risk through Fine-Tuning’ @ Good Systems Symposium: Shaping the Future of Ethical AI, UT Austin, March, 2024: [paper]
- ‘Foundation Models Meet Data Privacy: Risks and Countermeasures’ @ Trustworthy Machine Learning Course, Virginia Tech, Nov, 2023
- ‘Economizing Mild-Cognitive-Impairment Research: Developing a Digital Twin Chatbot from Patient Conversations’ @ BABUΕ KA FORUM, Nov, 2023: [link]
- ‘Backdoor Meets Data-Free Learning’ @ Hong Kong Baptist University, Sep, 2023: [slides]
- ‘MECTA: Memory-Economic Continual Test-Time Model Adaptation’ @ Computer Vision Talks, March, 2023: [slides] [video]
- ‘Split-Mix Federated Learning for Model Customization’ @ TrustML Young Scientist Seminars, July, 2022: [link] [video]
- ‘Federated Adversarial Debiasing for Fair and Transferable Representations’, @ CSE Graduate Seminar, Michigan State University, October, 2021: [slides]
- ‘Dynamic Policies on Differential Private Learning’ @ VITA Seminars, UT Austin, Sep, 2020: [slides]
Teaching
- Mentor@Directed Reading Program (DiRP) on Trustworthy LLM, UT Austin, 2023
- Teaching Assistant@CSE 847: Machine Learning, MSU, 2021
- Teaching Assistant@CSE 404: Introduction to Machine Learning, MSU, 2020
Mentoring
Co-mentored with my advisor:
- Zhangheng Li, PhD student at UTAustin, GenAI Privacy (SaTML'24)
- Haobo Zhang, PhD student at MSU, DL Privacy (NeurIPS'23)
- Shuyang Yu, PhD student at MSU, AI Security (ICLR'24, ICML'23, ICLR'23 Spotlight)
Services
- Organizers: GenAI4Health@NeurIPS, The Competition for LLM and Agent Safety 2024, The NeurIPS 2024 LLM Privacy Challenge, FL4Data-Mining Workshop@KDD 2023 (Lead Chair), FedKDD Workshop 2024 (Co-Lead Chair)
- External Reviewer: NeurIPS 22-23 (Top Reviewer, 2023), ICML 22-24, ICLR 23-24, KDD 22-24, ECML-PKDD 23, AISTATS 23, WSDM 22, AISTATS 22, AAAI 21-23, IJCAI 19, NeuroComputing, TKDD, TKDE, JAIR, TDSC, ACM Health
- Volunteer: KDD 18, 21