The CSE Best PhD Dissertation Award for 2023/24

The CSE Best PhD Dissertation Award Seminar and Certificate Presentation Ceremony was held successfully on April 14,2025 in HKUST Campus. We were glad to invite Prof. Xiaofang Zhou, Chair Professor Head of Department, as our distinguished guest to present the awards to our deserving awardees.

We are pleased to announce that the recipient of the CSE Best PhD Dissertation Award Dr. Yuanyuan Yuan for his work on “Side Channel Analysis for AI Infrastructures”. He also shared valuable insights and knowledge in this field during the seminar.

About the Seminar

His research explores how non-functional characteristics of systems can leak sensitive information, undermining security even in rigorously secured algorithms. He focuses on two main types of secrets: user inputs (like images and text) and the intellectual property of neural networks. The research highlights potential threats in AI systems which lead to risks of input inference and unauthorized access to neural networks. Moreover, Dr. Yuanyuan Yuan critiques existing security measures and proposes a universal detection method for identifying these secret leakages.

In addition, we would like to recognize the following students for their outstanding dissertations, which were selected as Honorable Mentions:

1. Dr. Hao Zhang, for his work on "From Query to Prompt: Towards Open-World Perception," advised by Prof. Heung-Yeung Shum and Prof. Lionel Ni
2. Dr. Haotian Li, for his work on "Bridging Data Analysis and Storytelling with Human-AI Collaborative Tools," advised by Prof. Huamin Qu
3. Dr. Yuandao Cai, for his work on "Making Call Graph Construction More Practical for Program Analysis in the Real World," advised by Prof. Charles Zhang

We would like to congratulate all our award recipients and express our gratitude to all the students who submitted dissertations for consideration. We are proud of our students' exceptional work and appreciate their contributions to our curriculum.

Seminar conducted by Dr. Yuanyuan Yuan

CSE Best PhD Dissertation Awardee, Dr. Yuanyuan Yuan
(From left to right): Prof. Kai Chen, Dr. Shuai Wang, Dr. Yuanyuan Yuan and Prof. Xiaofang Zhou

Honorable Mention Awardee, Dr. Haotian Li
(From left to right): Prof. Kai Chen, Prof. Huamin Qu and Prof. Xiaofang Zhou

About Yuanyuan Yuan's Dissertation and His Sharing

Biography:

Yuanyuan Yuan is currently a postdoctoral researcher in the Department of Computer Science at ETH Zurich. He obtained his Ph.D. in Computer Science and Engineering from The Hong Kong University of Science and Technology (HKUST) in 2024 under the supervision of Prof. Shuai Wang. Before that, he received his B.S. in Computer Science from Fudan University in 2020. Yuanyuan's research focuses on the safety and security of AI systems. His long-term goal is to strengthen AI systems' safety and security across a range of conventional and emerging scenarios. Over the past several years, he has been pursuing this goal primarily from software and hardware perspectives, with results published at prestigious research venues, such as IEEE Sample, USENIX Security, CCS, NDSS, ICSE, and ICLR. Yuanyuan's research is also recognized by the industry and presented at Black Hat USA and Black Hat Europe, two world-renowned industrial security conferences.

Dr. Yuanyuan Yuan

Overview of his research:

His PhD thesis is titled “Side Channel Analysis for AI Infrastructures”. Side channels indicate secret leakages through non-functional characteristics (e.g., execution time) of a system, which can break the security guarantee even if the system's algorithms are rigorously proved secure. As AI systems are increasingly adopted in security- and privacy-critical scenarios, his thesis presents the first comprehensive study of side-channel leakages in AI systems.

His thesis considers two key types of secrets: 1) user inputs such as images, text, audio, videos, etc., and 2) neural networks, the intellectual property of AI systems. It examines the leakage of these secrets in the entire workflow of AI systems and covers threats posed by all participants involved. Specifically, when multiple users are using an AI system, a malicious user may infer other users' inputs; meanwhile, the system's owner can also directly access these inputs. Moreover, when deploying an AI system on an untrusted host platform, the host may duplicate the neural network to copy the intellectual property. While mature solutions (e.g., trusted execution environments) have been implemented in modern AI systems to mitigate the above threats, his thesis breaks their security guarantees through different side channels. To identify secret leakages in AI system infrastructures, his thesis also proposes a universal detection approach.

Key breakthrough in research:

The main contributions of Yuanyuan's thesis are from three aspects. First, new threats and research directions: his thesis for the first time uncovers severe secret leakages due to side channels in AI systems, which consequently bring privacy breaches and enable adversaries to downgrade or manipulate AI systems. These findings call for a revisit of AI security from a more holistic, system-level perspective.

Second, new and automated secret recovery techniques: his thesis achieves the first successful recovery of neural network weights and diverse types of user inputs from different side channels, which were deemed hardly doable previously. The recovered secrets accurately match the original ones, and the recovery pipelines are automated, paving the way for future research on secret leakage in AI systems.

Third, new and universal leakage detection paradigm: his thesis recasts the paradigm of detecting side-channel leakages and presents a universal and scalable detection framework. It for the first time supports analysing infrastructures (e.g., data processing libraries, runtime environments like PyTorch) in modern AI systems. The detected leakages and the universal nature of his framework can help developers harden future AI systems.

Recent works:

After his graduation, Yuanyuan continues to study security threats in emerging AI systems — in particular, the recent large foundation model (LFM) systems. He has identified new secret leakages introduced by optimizations specifically designed for LFM computations.

In addition to the security research, Yuanyuan also actively improves the safety and reliability of AI systems with software analysis techniques. Recently, he and his colleagues have proposed the concept of dynamic benchmark to deeply assess the reasoning capabilities of large language models.

Thanks to the department and supervisor:

Yuanyuan believes that he could not have these achievements without the support from his supervisor and the department. He expresses his deepest gratitude to his supervisor, Dr. Shuai Wang, from whom he learned almost all his research skills. Dr. Wang's guidance during the first few years of Yuanyuan's Ph.D. study was invaluable, which shaped his research taste and the way of identifying, analyzing, and solving problems. Dr. Wang also provided substantial support for Yuanyuan's career development, such as opportunities to visit top research groups, collaborations with top researchers, and training on how to mentor junior students. Yuanyuan is grateful for Dr. Wang's patience, encouragement, and support. He is glad that he has gradually become a more mature and independent researcher under Dr. Wang's supervision. Without any doubt, Dr. Wang is the best advisor Yuanyuan could ever have and sets an exceptional role model for him.

Yuanyuan would also like to extend his sincere gratitude to the Department of Computer Science and Engineering. As one of the leading institutions in computer science research, the department offered him generous resources on research. For instance, the Research Travel Grant (RTG) which supported him in attending conferences and presenting his research; it was at these conferences that Yuanyuan got research ideas and established connections with other researchers. Besides, Yuanyuan is thankful for the excellent research atmosphere in the department and his research lab — the Cybersecurity lab; the insightful talks hosted by the department and the inspiring discussions with his lab members are invaluable for his research.

Feeling:

Being recognized with this award is an incredibly meaningful moment for Yuanyuan. It reflects the dedication he has put into his dissertation research and encourages him to aim higher in his future work. He is deeply appreciative of the opportunities and resources he received from the department. He also thanks his supervisor, mentors, colleagues, and lab members for their constant support in achieving this milestone in his PhD journey.

About Haotian Li's Dissertation and His Sharing

Biography:

Haotian Li is currently a Researcher in Social Computing Group at Microsoft Research Asia. Previously, he received his PhD in Computer Science and Engineering from The Hong Kong University of Science and Technology (HKUST) under the supervision of Prof. Huamin Qu in HKUST VisLab. Before that, he received his bachelor degree in Computer Engineering with first class honors from HKUST in 2019.

His recent research focus is to understand, enhance, and envision effective and safe human-AI collaboration from a human-centered perspective with techniques from human-computer interaction (HCI) and artificial intelligence (AI). He is passionate about integrating the power of humans and AI in both productivity and creativity tasks, such as data analysis and data storytelling. Previously, he has also worked on visual analytics and E-learning.

Dr. Haotian Li

Overview of his research:

Nowadays, data has transformed the norms of various fields, from scientific discovery to finance and business. In practice, data workers often perform data analysis tasks, such as seeking patterns inside the datasets. Then, they author data stories to communicate their insights and knowledge from data to other stakeholders with narrative techniques, such as authoring and presenting slide decks. To finish these tasks, data workers need to finish various tasks with a wide spectrum of skills. The workflow and skills pose considerable challenges to data workers. The collaboration between human users and AI agents has been considered one of the solutions since they have shown merit in lowering the skill barriers and reducing the efforts from data workers. To unleash the power of AI techniques through effective collaboration, this thesis enhances the theoretical foundations for human-AI collaboration in data analysis and storytelling and applies the theories for real-world applications. Furthermore, we hope to contribute our insights to related fields, such as cooperative work and data science, and serve as the cornerstone for future human-AI symbiosis.

Key breakthrough in research:

In this thesis, they contribute:

The first framework to characterize human-AI collaboration in data storytelling tools with the roles of human and AI collaborators and the stages where collaborators work together. The framework has been recognized by a Best Paper Honorable Mention Award in ACM CHI 2024.

1. A novel aspect to delineate the connection between data story pieces with meta relations.
2. The first systematic investigation of challenges and expectations in human-AI collaboration.
3. Two human-AI collaborative tools to facilitate data storytelling.

Thanks to the department and supervisor:

Prof. Qu was the one who gave me the chance to enter the world of research and taught me the “first principles” of surviving and establishing myself in the research community. First, he has always asked us to be unique and not to follow a similar path as others. I benefited greatly from the philosophy when I made many decisions, such as my research direction, my visit to Berkeley, and my future career. I am also grateful that Prof. Qu had always supported my decision and provided every help when needed. Second, I am always motivated by his spirit of “just doing it”. As a PhD student, there were lots of “the first times”, such as the first time to attend a conference, the first time to participate in proposal writing, and the first time to apply for awards. The spirit of “just doing it” has encouraged and will always encourage me not to be afraid of the challenges brought by these new challenges. Besides, there were so many cases of his insightful ideas on educating researchers, such as systematic skill training and project management methods. All of his sayings and ideas have shaped part of me and will benefit me for the rest of my life as a researcher.

Feeling:

It is a great honor to receive with the award. I feel thrilled to receive this award as a recognition for my past research but it is more important to serve as a motivation for me to continue the research direction and conduct more high-quality and cut-edge research to push forward the boundaries of how humans can live and work together with AI in the future world. At the same time, I would like to express my gratitude to all people who have helped and supported my research, especially my supervisor, Prof. Huamin Qu, my thesis committee, and all my mentors. Without them, it is impossible for me to receive the award.

About Hao Zhang's Dissertation and His Sharing

Biography:

Hao Zhang completed his Ph.D. in Computer Science and Engineering at the Hong Kong University of Science and Technology (HKUST) in 2024 under the guidance of Professors Heung-Yeung Shum and Lionel Ni. He was mainly working on open-world perception and has 'DINO' series works on vision foundation models. He published a series of papers in CVPR, ICLR, ICCV, ECCV, and NeurIPS on computer vision and multimodal learning. His research interests include open-set detection/segmentation and vision-language multi-modal foundation models. Hao is currently a Research Scientist at NVIDIA Research, where he continues to push the boundaries of AI and computer vision.

Dr. Hao Zhang

Overview of his research:

Hao Zhang's dissertation, From Query to Prompt: Towards Open-World Perception, explores innovative approaches to improve the efficiency and adaptability of perception models. By introducing novel query designs and integrating multimodal prompts, his work addresses key challenges in object detection, segmentation, and visual grounding. This research is pivotal in advancing autonomous systems and enabling AI to adapt and operate in dynamic, real-world environments.

Key breakthrough in research:

During Hao Zhang's Ph.D., he made several notable contributions:

1. Denoising Queries for Transformer Models: I introduced designs like DN-DETR and DINO, which leverage local priors to improve both training efficiency and performance. These models have become benchmarks in the community, widely adopted in both academia and industry.
2. Integration of Visual and Language Prompts: He developed models like Semantic-SAM and SEEM, which push the boundaries of open-world perception by combining visual and language prompts to tackle complex tasks such as zero-shot segmentation and interactive vision.
3. Multimodal Vision and Language Models: He proposed LLaVA-Grounding, which integrates large language models with perception systems to provide deeper understanding of objects, enabling a new level of interaction between humans and machines.

Recent works:

Hao Zhang's work has had a significant impact:

Firstly, his DINO model has been widely adopted as a top-performing detection head, achieving state-of-the-art results on the COCO detection leaderboard.

Secondly, his publications have collectively received over 4,900 citations, with key papers like DINO and DN-DETR shaping the future of object detection and segmentation.

Thirdly, practical tools he has developed, including Segment Everything Everywhere All at Once and Grounded Segment-Anything, have empowered researchers globally and gained significant attention in the field.

Thanks to the department and supervisor:

Hao Zhang is deeply grateful for the support and resources provided by the Department of Computer Science and Engineering at HKUST. His supervisors, Professors Heung-Yeung Shum and Lionel Ni offered invaluable mentorship and guidance throughout his Ph.D. journey. The department created a nurturing environment for innovation and provided access to cutting-edge facilities that enabled his research to thrive.

Feeling:

Receiving the CSE Best PhD Dissertation Award is an incredible honour and a defining moment in my academic journey. It represents not only the culmination of years of hard work but also the collective effort of his mentors, collaborators, and the supportive research community. This recognition inspires me to continue advancing AI and computer vision, tackling the challenges of tomorrow with renewed energy and determination.

About Yuandao Cai's Dissertation and His Sharing

Biography:

Yuandao Cai, after obtaining his PhD from HKUST in 2023, joined the HUAWEI Hong Kong Research Centre through the Talented Young Program. His research is centered on advancing practical program analysis and verification techniques to establish the reliability and security of complex modern codebases. His research has been published in prestigious conferences including PLDI, OOPSLA, USENIX Security, and ASPLOS. His research prototype tools have been adopted by Ant Group and Huawei for real-world usage. Furthermore, his work has been recognized by academia and industry, receiving the ASPLOS Best Paper Award, HUAWEI Distinguished Collaborator Award, and Ant Group Distinguished Research Project Award.

Dr. Yuandao Cai

Overview of his research:

The explosively increasing code size and high concurrency make modern software increasingly complex and error-prone. On the other hand, the ever-growing complexity of modern codebases has made many conventional program analysis tasks impractical for safeguarding today’s software. As one of the most fundamental tasks of program analysis, call graph construction for large-scale code has become either imprecise or unscalable, which impairs the effectiveness of many downstream program analysis applications. In addition, as one of the most popular tasks of program analysis, bug detection faces significant challenges in effectively detecting concurrency bugs. To address these challenges, his thesis contributes to making fundamental call graph construction and popular concurrency bug detection more practical.

Key breakthrough in research:

They introduce two groundbreaking methods that characterize two kinds of function pointers, abandoning the conventional analysis mode of using uniform pointer analyses and type analyses without distinguishing various function pointers. By evaluating the call graphs through the lens of several various downstream program analysis clients, their approaches can dramatically promote the clients' effectiveness for better vulnerability understanding, hunting, and reproduction.

They also advocate a series of static approaches for detecting concurrency bugs, such as race-like bugs and deadlocks, capable of scaling to large million-line programs with high precision. Their experimental results show that their concurrency bug detectors can beat many popular open-source and commercial static concurrency bug detectors, such as Clang Static Analyzer (CSA) and Facebook/Meta Infer, regarding efficiency and precision.

Recent works:

Yuandao Cai’s research prototype tools have successfully caught over 300 security bugs across a variety of complex open-source systems and closed-source IoT devices. This work has resulted in the assignment of 76 CVE IDs and commendations from software vendors. Additionally, his call graph construction tools have been integrated into HUAWEI's program testing platform, enhancing code coverage and product quality. Furthermore, his concurrency bug detectors have been deployed at both HUAWEI and Ant Group, detecting and preventing numerous bugs in practice. These efforts have been recognized with distinguished research project awards and have contributed to a multimillion-Hong Kong dollar collaboration on the Trustworthy Parallel Software project between Huawei and our research laboratory.

Thanks to the department and supervisor:

Yuandao Cai’s supervisor has been instrumental in guiding him with valuable advice and consistently encouraging him to tackle challenging research problems. Additionally, his supervisor has facilitated many opportunities for internships and interactions with industry practitioners, thereby exposing me to real-world issues and requirements. These experiences have significantly broadened my understanding of practical challenges and demands in the field.

Feeling:

Yuandao Cai is deeply honored and grateful for the recognition of the Honorable Mention in the Best PhD Thesis Award. He attributes this achievement not only to his own efforts but also to the unwavering support and assistance of my supervisors and friends from HKUST and HUAWEI. Their guidance and encouragement have been invaluable on his PhD journey. This award serves as a strong motivation for him to persist in his research endeavors, particularly in the field of program analysis, and to strive toward making further meaningful contributions to the academic and industry community.