CV
TIANQI CHEN
🏛️ Education
University of Texas at Austin Austin, TX
Aug 2021 - May 2026 PhD in Statistics
University of Michigan Ann Arbor, MI
Sept 2019 - Apr 2021 Master of Science in Applied Statistics
Fudan University Shanghai, PRC
Sept 2015 - June 2019 Bachelor of Science in Mathematics and Applied Mathematics
📖 Publications and preprints
ASK: Adversarial Soft k-Nearest Neighbor Attack and Defense [arxiv]
under review
- We proposed a novel information-theoretic loss function as a differentiable surrogate of kNN classification error, based on which we developed a new attack method, ASK-Attack, that outperforms existing kNN attacks by a large margin.
- We further devised ASK Defense, a regularized adversarial training strategy built upon ASK loss. To best of our knowledge, it is the first algorithm that effectively defends against kNN attacks on hidden layers of DNNs.
- We conducted extensive experiments on hyperparameter sensitivity as well as provided detailed ablation study on all the components used in both ASK Attack and ASK Defense.
Immuno-mimetic Deep Neural Networks (Immuno-Net) [arxiv]
accepted by the 2021 ICML Workshop on Computational Biology
- We proposed a general biomimetic evolutionary algorithm intended for robustification of deep neural networks (DNNs).
- The algorithm can further improve the classification performance of state-of-the-art methods on typical machine learning datasets such as SVHN & CIFAR10 under adversarial setting.
RAILS: A Robust Adversarial Immune-inspired Learning System [arxiv] [ieee access]
published in IEEE Access
- We proposed a novel adversarial learning framework inspired by mammalian immune system, which is agnostic to backbone neural network architectures and adaptive to unseen attacks.
- We also empirically showed that RAILS can improve the robustness of DNN classifiers with different model structures (e.g., VGG16 and ResNet18) and various SOTA adversarial attacks (e.g., CW, PGD, Square and HopSkipJump).
🚀 Projects
GARD: Guaranteeing AI Robustness Against Deception
Research Assistant July, 2020 – Present
- Participated in the proposal of a novel immune system inspired adversarial learning algorithm, implemented and tested the architecture in pytorch;
- Experimented on the model initialization mechanisms, affinity measurements and multiple hyperparameter settings, accelerated and stablized the affinity maturation simulation process;
- Compared RAILS and DkNN defenses both under white-box attacks (PGD) on CNN and under black-box attacks (ZOO & HopSkipJump) on the whole architecture, the novel algorithm outperformed its counterparts in adversarial cases and preserved accuracy given the clean inputs.
UM-OIG Project: High-dose Opioid Transaction Prediction and U.S. Medicaid Pharmacy Fraudulence Risk Evaluation
Research Assistant May 2020 – Sept. 2020
- Processed ∼500M transaction data from the Michigan Department of Health and Human Service, derived predictor variables including morphine milligram equivalents and opioid-involved overdose death rate by CDC prescription guideline and Michigan demographic data.
- Investigated the relationship between high-dose opioid transaction with factors such as pharmacy ownership, geographical location and patient composition with weighted least square linear models.
- Built three models using standard logistic regression, random forest and non-ignorable missing data methods (Ibrahim and Lipschitz algorithm with Firth penalty) and compared the model performances on the hold-out set.
💼 Work Experiences
Department of Information, Risk and Operations Management, University of Texas at Austin
Teaching Assistant Sept 2021 - Present Student Technician June 2021 - Aug 2021
Department of Statistics, University of Michigan
Research Affiliate May 2021 - May 2022 Graduate Student Instructor Aug 2020 - May 2021 Course Grader May 2020 - Aug 2020
⚙️ Skills
Programming
proficient in Python, R, SQL
ML / DL
familiar with NumPy, pandas, scikit-learn, experienced in TensorFlow and PyTorch
Language
Chinese (native), English (fluent), Japanese (intermediate)