UCLA
I am a Ph.D. Candidate at UCLA Computer Science Department, advised by Prof. Cho-Jui Hsieh. Before that, I received my B.Eng. degree from the CST department at Tsinghua University where I worked with Prof. Minlie Huang.
My research interest is machine learning, and my primary research focus is on trustworthy machine learning and the robustness of machine learning models. My works include:
General and scalable formal verification for neural networks: I developed formal verification methods for general neural network-based models, from Transformers, to general computational graphs and higher-order computational graphs. These efforts have been integrated into our scalable open-source software, auto_LiRPA (originally proposed in our NeurIPS 2020 paper) for general perturbation analysis and verified bound computation. Based on auto_LiRPA and complete verification algorithms including complete verification for models with general nonlinearities, we also developed a complete verification toolbox named alpha-beta-CROWN. We have won the International Verification of Neural Networks Competitions (VNN-COMP) for three years from 2021 to 2023, with demonstrated applications in power systems, computer systems, etc.
Adversarial robustness in NLP: I proposed methods for generating adversarial examples for NLP by leveraging modern language models. To generate word-substitution attacks with high-quality synonyms that are compatible with context, I proposed to leverage a masked BERT model, use a instructional-following model such as ChatGPT, or use a text-completion model such as LLaMA. I proposed to red team large language model detectors by LLM-generated word substitution attacks or instructional prompts. We also recently proposed to defend against LLM jailbreaking attacks by backtranslation.
Evaluation of out-of-distribution robustness: I proposed a new understanding and evaluation on the effective robustness of multimodal pre-trained models especially CLIP models. I found that pre-training data in CLIP can interfere the previous evaluation of OOD effective robustness rather than improve effective robustness, and suggested that CLIP is not effectively more robust than traditional models.
Efficiently training robust neural networks: I developed methods for faster certified robust training by an improved interval bound propagation-based training.
TA at UCLA: