Tianyun Yang (杨天韵)

I'm a final-year PhD student at the University of Chinese Academy of Sciences (UCAS), supervised by Juan Cao.

I am interested in mechanism interpretability and safety of AI models, covering topics such as hallucination mitigation, concept editing, model attribution, etc.

Email  /  Google Scholar  /  Github

profile photo

Representative Researches


Mitigating Hallucinations in Large-Vision Language Models via Modular Attribution and Intervention
ICLR, 2025
Tianyun Yang, Ziniu Li, Juan Cao, Chang Xu
Code / Paper

This work adopts a modular perspective to investigate the causes of hallucination in large vision-language models, analyzing how particular components contribute to this issue and proposing methods to mitigate it.


Model Synthesis for Zero-shot Model Attribution
IEEE Transactions on Multimedia (TMM), 2025
Tianyun Yang, Juan Cao, Danding Wang, Chang Xu
Code / Paper

This work aims to develop a generalized model fingerprint extractor capable of Zero-Shot Model Attribution that effectively attributes unseen models without exposure during training. Central to our method is a model synthesis technique, which generates numerous synthetic models that mimic the fingerprint patterns of real-world generative models.


Pruning for Robust Concept Erasing in Diffusion Models
Workshop on Safe Generative AI at Conference on NeurIPS, 2024
Tianyun Yang, Ziniu Li, Juan Cao, Chang Xu
Paper

This work designs a robust concept erasing method based on differential pruning to eliminate harmful or copyrighted concepts from diffusion models


Progressive Open Space Expansion for Open-Set Model Attribution
CVPR, 2023
Tianyun Yang, Danding Wang, Fan Tang, Xinying Zhao, Juan Cao, Sheng Tang
Code / Paper

This work presents the first study on Open-Set Model Attribution (OSMA), to simultaneously attribute images to known models and identify those from unknown ones. We propose a Progressive Open Space Expansion (POSE) solution, which simulates open-set samples that maintain the same semantics as closed-set samples but embedded with different imperceptible traces.


Deepfake Network Architecture Attribution
AAAI, 2022
Tianyun Yang*, Ziyao Huang*, Juan Cao, Lei Li, Xirong Li
Code / Paper

This work presents the first study on Deepfake Network Architecture Attribution to attribute fake images on architecture-level. Based on an observation that GAN architecture is likely to leave globally consistent fingerprints while traces left by model weights vary in different regions, we provide a simple yet effective solution named DNA-Det for this problem.

Education

2019- Institute of Computing Technology, Chinese Academy of Sciences
Ph.D. in Computer Science
Advisor: Juan Cao
2023-2024 The University of Sydney
Joint Ph.D, School of Computer Science
Advisor: Chang Xu
2015-2019 Wuhan University
B.E., School of Electrical Engineering, Excellent Engineer Class

Honors

2022 First Prize of Academic Award, University of Chinese Academy of Sciences
2021 Director's Excellence Scholarship, Institute of Computing Technology
2021 The 1st Prize in Chinese AI Competition, Deepfake Identification
2018 The 1st Prize in Mathematical Contest in Modeling, Hubei Province

Services

Reviewer: T-MM, ICML'25, ICLR’25, TMLR’24, ICLR’24, NeurIPS’24, NeurIPS’23, CVPR’23, NeurIPS’22

Design by Jon Barron, Last updated: 2025-03-10