Yun-Yun (Alice) Tsai

Full list on Google Scholar ↗

§Conference and Workshop Papers

1. 

   WebUIProof: Benchmarking WebUI Code Generators with UI-Agent Execution Harness

   Yun-Yun Tsai, Yuning Mao, Shiqi Wang, Junfeng Yang, Sinong Wang

   arXiv 2026 paper TL;DR

   WebUIProof is an execution-oriented benchmark for evaluating WebUI code generation beyond visual appearance or build success. It provides structured UI specifications and dense, executable interaction tests that verify whether generated websites actually behave correctly under user actions, covering both general WebUIs and interactive 3D simulations.

2. 

   RAFP: Identifying LLM Lineages via Rare-Region Fingerprints

   Yun-Yun Tsai, Jia-Hao Liang, Chuan Guo, Junfeng Yang, Laurens van der Maaten

   arXiv 2025 paper TL;DR

   RAFP is a black-box fingerprinting framework for verifying LLM lineage and ownership through carefully optimized prompt–response pairs. It constructs rare, model-specific behavioral fingerprints that remain detectable even after common model modifications such as fine-tuning, LoRA, quantization, system-prompt changes, and different decoding settings, enabling robust post-deployment model provenance verification without access to model weights.

3. 

   ARQ: Agentic CodeQL Query Refinement for C/C++ Vulnerability Detection

   Chunyi Wang, Yunfei Ke, Junfeng Yang, Yun-Yun Tsai, Penghui Li

   arXiv 2026 paper TL;DR

   ARQ is an agentic framework that automatically refines C/C++ CodeQL vulnerability-detection queries: it synthesizes near-identical witness programs, checks the query’s verdict against execution oracles to expose false positives and false negatives, and iteratively repairs the query — no labeled data or vulnerability-specific templates needed.

4. 

   Reasoning under Vision: Understanding Visual-Spatial Cognition in Vision-Language Models for CAPTCHA

   Jincen Song, Luke Tenyi Wang, Yun-Yun Tsai, Penghui Li, Junfeng Yang

   AACL 2026 Findings of the Association for Computational Linguistics paper TL;DR

   We study how vision-language models perform visual-spatial reasoning through CAPTCHA-style tasks. It uses controlled visual challenges to probe whether models can accurately perceive objects, understand spatial relationships, and reason over visual layouts, revealing where current VLMs succeed and where their visual reasoning still breaks down.

5. 

   REVEAL: Robust Evolution of Vision-Language Models for Explainable AI-Video Detection

   Yun-Yun Tsai, Qingyuan Liu, Ruijian Zha, Victoria Li, Pengyuan Shi, Chengzhi Mao, Junfeng Yang

   ICCV 2025 Workshop Knowledge-Intensive Multimodal Reasoning Workshop paper TL;DR

   We propose an explainable AI-video detection framework that improves robustness by jointly leveraging visual evidence and vision-language reasoning. It helps distinguish real from AI-generated videos while producing interpretable explanations for its predictions, aiming to make detection more reliable across different generators and manipulation conditions.

6. 

   GDA: Generalized Diffusion for Robust Test-time Adaptation

   Yun-Yun Tsai, Fu-Chen Chen, Albert Y. C. Chen, Junfeng Yang, Che-Chun Su, Min Sun, Cheng-Hao Kuo

   CVPR 2024 The IEEE / CVF Computer Vision and Pattern Recognition Conference paper TL;DR

   GDA is a diffusion-driven test-time adaptation framework that improves model robustness under distribution shifts without retraining. It uses diffusion-based conditional generation together with entropy, style, and content objectives to adapt inputs at inference time, helping pretrained models recover performance on out-of-distribution data.

7. 

   Convolutional Visual Prompt for Robust Visual Perception

   Yun-Yun Tsai, Chengzhi Mao, Junfeng Yang

   NeurIPS 2023 37th Conference on Neural Information Processing Systems paper TL;DR

   We propose CVP, a lightweight test-time adaptation method that learns convolutional input-space prompts to improve model robustness under distribution shift. By modifying visual inputs rather than updating the full model, it enables efficient deployment-time adaptation while preserving the pretrained model and improving performance on out-of-distribution data.

8. 

   Towards Compositional Adversarial Robustness: Generalizing Adversarial Training to Composite Semantic Perturbations

   Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, Tsung-Yi Ho

   CVPR 2023 The IEEE / CVF Computer Vision and Pattern Recognition Conference paper TL;DR

   Towards Compositional Adversarial Robustness extends adversarial training beyond single perturbation types to composite semantic attacks. It trains models to remain robust when multiple meaningful transformations are combined, improving generalization to more realistic and previously unseen adversarial perturbations.

9. 

   Test-time Defense against Adversarial Attacks: Detection and Reconstruction of Adversarial Examples via Masked Autoencoder

   Yun-Yun Tsai, Ju-Chin Chao, Albert Wen, Zhaoyuan Yang, Chengzhi Mao, Tapan Shah, Junfeng Yang

   CVPR 2023 Workshop AdvML Workshop paper TL;DR

   We use a pretrained masked autoencoder (MAE) as a test-time defense: its reconstruction loss flags adversarial inputs at inference, and its latent features repair the image so the classifier predicts correctly — requiring no model retraining.

10. 

    CARBEN: Composite Adversarial Robustness Benchmark

    Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, Tsung-Yi Ho

    IJCAI 2022 International Joint Conference on Artificial Intelligence paper demo TL;DR

    CARBEN is a benchmark of composite adversarial attacks that combine multiple semantic perturbations, for evaluating model robustness beyond standard single-type ℓ∞ attacks.

11. 

    Voice2Series: Reprogramming Acoustic Models for Time Series Classification

    Chao-Han Huck Yang, Yun-Yun Tsai, Pin-Yu Chen

    ICML 2021 International Conference on Machine Learning paper TL;DR

    Voice2Series reprograms pretrained acoustic models for time-series classification by transforming non-audio time-series signals into representations that speech models can understand. It enables knowledge from large pretrained speech models to transfer to diverse time-series tasks, achieving strong classification performance with limited task-specific training.

12. 

    Transfer Learning without Knowing, Reprogramming Black-box Machine Learning Model with Scarce Data and Limited Resources

    Yun-Yun Tsai, Pin-Yu Chen, Tsung-Yi Ho

    ICML 2020 International Conference on Machine Learning paper video slides code TL;DR

    We introduce a black-box model reprogramming framework that repurposes pretrained machine learning models for new tasks without access to their architecture, parameters, or training data. It learns input/output transformations through model queries, enabling knowledge transfer from inaccessible models to substantially different target tasks.

13. 

    CloudLeak: Large-Scale Deep Learning Models Stealing Through Adversarial Examples

    Honggang Yu, Kaichen Yang, Teng Zhang, Yun-Yun Tsai, Tsung-Yi Ho, Yier Jin

    NDSS 2020 Network and Distributed System Security Symposium paper code TL;DR

    We introduce an efficient black-box model-stealing attack against cloud ML services. It combines adversarial example generation, active learning, and transfer learning to selectively query informative samples near the victim model’s decision boundary, allowing an attacker to train a local substitute model with similar behavior using far fewer queries than prior extraction attacks.

14. 

    Vehicle Sequence Reordering with Cooperative Adaptive Cruise Control

    Ta-Wei Huang, Yun-Yun Tsai, Chung-Wei Lin, Tsung-Yi Ho

    DATE 2019 Design, Automation & Test in Europe Conference & Exhibition paper TL;DR

    We study reordering the crossing sequence of connected vehicles with cooperative adaptive cruise control, showing that coordinating the order vehicles pass through intersections can reduce total delay and improve traffic throughput compared with first-come-first-served scheduling.

§Patents

1. Pin-Yu Chen, Yun-Yun Tsai, Sijia Liu, Chia-Yu Chen, I-Hsin Chung, Tsung-Yi Ho. ”Transfer Learning With Machine Learning Systems”, U.S. Patent Application No: 17/029506, Application Date: Sept. 23, 2020.

2. Pin-Yu Chen, I-Hsin Chung, Bo Wu, Chuang Gan, Lei Hsiung, Yun-Yun Tsai, Tsung-Yi Ho. ”Composite Adversarial Attack Model Training For Neural Networks”, U.S. Patent Application No: 18/331,211, Application Date: May 12, 2026.