I am a Ph.D. Candidate (2022.08-) in Computing Information & Science here at Rochester Institute of Technology (RIT) supervised by Prof. Weijie Zhao. I previously interned at Amazon (AGI, Foundation Models) and ByteDance (Audio & Speech Processing). My research interests span LLMs, AI Privacy & Security, and Scalable & Trustworthy Machine Learning.
Prior to RIT, I worked as a research assistant at the Institute of Computer Vision (ICV), Shenzhen University, China, and was supervised by Prof. Linlin Shen (Honorary Professor at the University of Nottingham, UK). I got the bachelor’s degree in Computer Science and Technology in July 2021. I was the leader of a project funded by the Ministry of Education of China for undergraduates in 2019. I received several ACM-ICPC Medals during my undergraduate.
My curriculum vitae can be found at here.
Apr. 5, 2025: I will be joining NEC Labs America as a Research Intern this summer, focusing on Trustworthy LLMs.
Apr. 4, 2025: One paper is accepted by SIGIR 2025! 🎉🎉🎉
Mar. 22, 2025: Invited as a reviewer for ACL ARR 2025.
Mar. 17, 2025: Invited as a reviewer for ICML 2025.
Feb. 3, 2025: We released two papers on ArXiv:
- DMin: Scalable Training Data Influence Estimation for Diffusion Models: The first highly scalable influence estimation method for stable diffusion models with billions of parameters, capable of identifying the most influential training samples for a given test generation within seconds.
- Online Gradient Boosting Decision Tree: In-Place Updates for Efficient Adding/Deleting Data: The first work to introduce an in-place unified incremental and decremental learning approach for GBDT, enabling real-time addition and deletion of data within the model.
Jan. 26, 2025: Invited as a reviewer for IEEE Transactions on Dependable and Secure Computing.
Jan. 22, 2025: One paper is accepted by NAACL 2025! 🎉🎉🎉
Aug. 23, 2024: Invited as a reviewer for ICLR 2025.
Jul. 21, 2024: Invited as a reviewer for ACL ARR 2024.
Jul. 12, 2024: Invited as a reviewer for KDD 2025.
May. 16, 2024: Our paper (Token-wise Influential Training Data Retrieval for Large Language Models) is accepted by ACL 2024! In this paper, we propose RapidIn, a scalable framework for estimating the influence of each training data sample on LLMs.
Apr. 16, 2024: I will be joining Amazon as an Applied Scientist Intern in Boston this summer, focusing on large language models and multimodal systems. 🎉🎉🎉
Feb. 26, 2024: We released an easy-to-run implementation for finetuning large language models (LLMs) such as llama and gemma, supporting full parameter finetuning, LoRA, and QLoRA. Please feel free to star, fork, and make your own contributions. [Github Repo]
Feb. 12, 2024: Invited as a reviewer for KDD 2024.
Oct. 02, 2023: Invited as a reviewer for NeurIPS 2023 GLFrontiers Workshop.
Aug. 05, 2023: Received the KDD’23 Student Travel Award. Thanks to KDD!
May. 16, 2023: Our paper (Machine Unlearning in Gradient Boosting Decision Trees) is accepted by KDD 2023! [Promotion Video, Poster]
Sep. 12, 2022: Our paper (Activation Template Matching Loss for Explainable Face Recognition) is accepted by the 2023 IEEE Conference on Automatic Face and Gesture Recognition (FG 2023)!
Publications & Pre-prints
2025
- Huawei Lin, Yingjie Lao, Tong Geng, Tan Yu, Weijie Zhao, UniGuardian: A Unified Defense for Detecting Prompt Injection, Backdoor Attacks and Adversarial Attacks in Large Language Models. arXiv preprint, arXiv:2502.13141, 2025.
- Huawei Lin, Jun Woo Chung, Yingjie Lao, Weijie Zhao, Online Gradient Boosting Decision Tree: In-Place Updates for Efficient Adding/Deleting Data. arXiv preprint, arXiv:2502.01634, 2025.
- Jun Woo Chung, Huawei Lin, Weijie Zhao, Locality-Sensitive Indexing for Graph-Based Approximate Nearest Neighbor Search. SIGIR 2025.
- Yanzhou Pan, Huawei Lin, Yide Ran, Jiamin Chen, Xiaodong Yu, Weijie Zhao, Denghui Zhang, Zhaozhuo Xu, ALinFiK: Learning to Approximate Linearized Future Influence Kernel for Scalable Third-Parity LLM Data Valuation. NAACL 2025.
2024
- Huawei Lin, Yingjie Lao, Weijie Zhao, DMin: Scalable Training Data Influence Estimation for Diffusion Models. arXiv preprint, arXiv:2412.08637, 2024.
- Huawei Lin, Jikai Long, Zhaozhuo Xu, Weijie Zhao, Token-wise Influential Training Data Retrieval for Large Language Models. ACL 2024.
2023
- Huawei Lin, Jun Woo Chung, Yingjie Lao, Weijie Zhao, Machine Unlearning in Gradient Boosting Decision Trees. KDD 2023.
- Huawei Lin, Haozhe Liu, Qiufu Li, Linlin Shen, Activation Template Matching Loss for Explainable Face Recognition. 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), 2023
Tech Talks
- “Influence Estimation for Large Language Models and Diffusion Models” at Rochester Institute of Technology, Apr. 1, 2025. [Slides]
- “Token-wise Influential Training Data Retrieval for Large Language Models” at ACL Virtual Poster Session, Aug. 12, 2024. [Slides, Video]
- “Toward Explainable Large Language Models via Influence Estimation” at Boston, MA, May. 23, 2024. [Poster]
- “Machine Unlearning in Gradient Boosting Decision Trees” at Long Beach, CA, Aug. 9, 2023. [Slides, Video]
- “Activation Template Matching Loss for Explainable Face Recognition” at Rochester Institute of Technology, Nov. 17, 2022. [Slides, Video]
- “Toward Explainable Face Recognition” at Shenzhen University, Oct. 28, 2021. [Slides]
- “Trust in Black-Box Models: Interpretability & Explainability for Deep Learning” at Shenzhen University, Aug. 13, 2021. [Slides]
Research Experience
UniGuardian: A Unified Defense for Detecting Prompt Injection, Backdoor Attacks and Adversarial Attacks in Large Language Models
- We define Prompt Trigger Attacks (PTA) as a unified category encompassing prompt injection, backdoor, and adversarial attacks.
- We analyze their common mechanisms and demonstrate, both theoretically and empirically, the behavioral distinctions of LLMs when processing injected versus clean prompts.
- We introduce UniGuardian, a novel training-free, inference-time detection mechanism that efficiently detects multiple attack types.
Online Gradient Boosting Decision Tree: In-Place Updates for Efficient Adding/Deleting Data
- We introduce an efficient in-place online learning framework for gradient boosting models supporting incremental and decremental learning, extensible to finetuning and transfer learning.
- We present optimizations to reduce the cost of incremental and decremental learning, making adding or deleting a small data fraction substantially faster than retraining.
- This is the first work to introduce an in-place unified incremental and decremental learning approach for GBDT, enabling real-time addition and deletion of data within the model without training from scratch.
DMin: Scalable Training Data Influence Estimation for Diffusion Models
- We introduce DMin, a scalable influence estimation framework for diffusion models that efficiently identifies the most influential training samples for a given test generation in seconds.
- DMin is the first highly scalable framework that enables stable influence estimation on diffusion models with billions of parameters.
- To overcome storage and computational limitations, DMin employs a gradient compression technique, reducing storage from around 40 GB to 80 KB per sample while maintaining accuracy, enabling feasible influence estimation on large models and datasets.
Token-wise Influential Training Data Retrieval for Large Language Models
- We present RapidIn that estimates the influence of each training data for a given LLM generation.
- We apply a collection of techniques to cache the gradients of LLMs by compressing gradient vectors by over 200,000x in the caching stage, and achieve a 6,326x speedup in the retrieval stage, enabling estimating the influence of the entire dataset for any test generation within minutes.
- We utilize multi-GPU parallelization to substantially accelerate the caching and retrieval.
Machine Unlearning in Gradient Boosting Decision Trees (GBDT)
- Propose an unlearning framework that efficiently and effectively unlearns a given collection of data without retraining the model from scratch.
- Introduce a collection of techniques, including random split point selection and random partitioning layers training, to the training process of the original tree models to ensure that the trained model requires few subtree retrainings during the unlearning.
- To the best of our knowledge, this is the first work that considers machine unlearning on GBDT.
Activation Template Matching Loss for Explainable Face Recognition
- Propose a novel method named Explainable Channel Loss (ECLoss) to construct an explainable face recognition network, which can directly explain that what face recognition networks have learned.
- To the best of our knowledge, this is the first method to construct a feature level explainable face recognition network that does not require any additional dataset or manual annotation.
Internship Experience
Amazon - AGI (Foundation Models), Boston, MA
May 2024 - Present
Applied Scientist Intern | Advisor: Rajath Kumar, Raphael Petegrosso
- Responsible for an Auto Prompting project on LLMs inference.
- Proposed an unsupervised self-improving framework for LLMs inference that enhances generation quality across various downstream tasks. The proposed framework can generate multiple diverse outputs and detect potential hallucinations by certainty score.
- The paper draft was submitted for internal review.
Shenzhen University - Institute of Computer Vision, Shenzhen, China
July 2021 - July 2022
Research Assistant | Supervisor: Prof. Linlin Shen
- Responsible for interpretability and explainability on deep learning, especially on Biometrics.
- Propose a novel method named Explainable Channel Loss (ECLoss) to construct an explainable face recognition network, which can directly explain that what face recognition networks have learned.
- To the best of our knowledge, this is the first method to construct a feature level explainable face recognition network that does not require any additional dataset or manual annotation.
ByteDance Inc. - AI Lab, Beijing, China
Oct. 2020 - July 2021
Software Engineer
- Responsible for Audio Recognition and Understanding research and development, and technical supports for ByteDance’s applications, including TikTok.
- Proposed Stream Audio Understanding Chain method which enabled real-time audio processing and achieved precise extraction of information from audios, including speakers’ genders, tones, emotions, etc.
- Designed a pipeline processing flow that significantly increased the throughput of CPUs and reduced processing time by 65%, by optimizing the usage of CPUs/GPUs using multiprocessing and multithreading.
Fellowships & Awards
Fellowships
- KDD23, Student Travel Awards Aug. 2023
- Zhou Lian Academic Scholarship (Only 1 of ~20,000 Students) Oct. 2020
- Academic Innovation and Technology Scholarship (Only 10 of ~20,000 Students) May 2020
- IAPR/IEEE Winter School on Biometrics 2020, Student Travel Grants Jan. 2020
- Academic Innovation and Technology Scholarship (Only 10 of ~20,000 Students) May 2019
Awards
- ACM-ICPC National Programming Contest (Shaanxi), Bronze Medal June 2021
- ACM-ICPC Programming Contest (Shaanxi Province), Silver Medal Sept. 2020
- ACM-ICPC National Programming Contest (Shaanxi), Bronze Medal June 2020
- ACM-ICPC National Programming Contest (Yinchuan), Bronze Medal May 2019
- ACM-ICPC Asia Regional Contest, Bronze Medal Nov. 2018
- ACM-ICPC Chinese Collegiate Programming Contest, Bronze Medal Jan. 2018
Professional Services
Conference Reviewer
- ICML 2025
- ICLR 2025
- ACL ARR 2024, 2025
- KDD: 2024, 2025
- NeurIPS 2023 GLFrontiers Workshop
Journal Reviewer
- IEEE Transactions on Dependable and Secure Computing
- IEEE Access
TECHNICAL SKILLS
- Programming Languages: C/C++, Python, Go, Java, Shell, HTML
- Deep Learning Tools: PyTorch, CUDA, Keras, TensorFlow
- Deep Learning Packages: Transformers, DeepSpeed, FSDP, PEFT, OpenAI
- Others: Slurm, MATLAB, Docker, Hadoop, Kubernetes, Kafka