The CSE Best PhD Dissertation Award for 2022/23

The CSE Best PhD Dissertation Award Seminar and Certificate Presentation Ceremony was held successfully on March 6, 2024 in HKUST Campus. We were glad to invite Prof. Dit-yan Yeung, Acting Department Head, 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. Wei DONG for his work on "Practical Data Analytics under Differential Privacy". He also shared valuable insights and knowledge in this field during the seminar.

About the Seminar

His research focuses on practical DP SQL query engines and AI solutions. In the era of big data, organizations face the challenge of balancing privacy and meaningful analytics. Differential privacy (DP) has emerged as the gold standard, ensuring indistinguishable query results while preserving individual privacy. However, many DP methods lack practical utility due to consistent high error levels. To address this, Wei Dong proposes a novel concept: instance-optimal error, minimizing error for specific instances.

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

1. Dr. Chaoliang ZENG, for his work on "Towards High-performance Datacenter Systems with Application-oriented Optimizations" advised by Prof. Kai CHEN
2. Dr. Tengfei WANG, for his work on "High-Quality Visual Content Creation With Foundation Generative Models" advised by Dr. Qifeng CHEN
3. Dr. Zhibo LIU, for his work on "Towards Assessing and Enhancing Modern Software Reverse Engineering" advised by Dr. Shuai WANG

We would like to extend our congratulations to all our award recipients, as well as our sincere thanks to all of the students who submitted their dissertations for consideration. We are proud of the excellent work being done by our students and are grateful for their contributions to our program.

Seminar conducted by Dr. Wei Dong

CSE Best PhD Dissertation Awardee, Dr. Wei Dong
(From Left to Right) Prof. Ke Yi, Dr. Wei Dong and Prof. Dit-yan Yeung

Honorable Mentions Awardee, Dr. Chaoliang Zeng
(From Left to Right) Prof. Kai Chen, Dr. Chaoliang Zeng and Prof. Dit-yan Yeung

Honorable Mentions Awardee, Dr. Zhibo Liu
(From Left to Right) Dr. Zhibo Liu and Prof. Dit-yan Yeung

About Wei Dong's Dissertation and His Sharing

Biography:

Wei Dong is currently a Postdoctoral Fellow in the Department of Computer Science at Carnegie Mellon University. He earned his Ph.D. from the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology in 2018 under the supervision of Prof. Ke Yi. His general areas of interest span data security and privacy, database theory and algorithm, and machine learning. Wei's research has been recognized by the academic community and has been presented at top conferences, such as SIGMOD, S&P, PODS, CCS, VLDB, NeurIPS, and KDD. Notably, he received the prestigious Best Paper Award at SIGMOD 2022(1/514) and SIGMOD 2023 research highlight award. Additionally, he is the recipient of the 2023 PhD Research Excellence Award from the School of Engineering of the Hong Kong University of Science and Technology and the 2018 Hong Kong Ph.D. Fellowship.

Wei Dong

Overview of his research:

He studies deploying private data analysis in the database and machine learning fields. In the big data era, organizations continuously collect vast amounts of sensitive information. A key challenge is to get meaningful analytical results without compromising privacy. One recent breakthrough in this field is a privacy definition called differential privacy (DP). As a gold standard for private data analysis, DP has been deployed by big companies like Apple, Google, Microsoft, Amazon, and institutions such as the U.S. Census Bureau.

Informally speaking, DP ensures that query results are indistinguishable regardless of whether any particular individual's data is included or not in the database thus we cannot infer any individual's information through the query result. Noise injection is inherently necessary for this goal. Although DP is widely researched in all kinds of data science areas, many DP mechanisms do not provide a practical utility (error level) in real-world applications. In practice, the errors introduced by some DP mechanisms can greatly surpass the actual query results, making these results private but also meaningless. His research has been dedicated to developing DP solutions with good utilities in both theory and practice. He mainly studies two problems. The first one is to build a practical DP SQL query engine. The second one is to design practical AI solutions.

Key breakthrough in research:

Traditional DP mechanisms often incur large errors because they impose a global limit on individual contributions to queries and add the noise proportional to that limit. This approach results in the same error level across all datasets, which can often be very large. Furthermore, because this mechanism considers only the global worst-case scenario, the resulting error is called worst-case error. He proposes a 'paradigm shift' in DP mechanism design, which aims at achieving instance-optimal error. Roughly speaking, instead of only considering the worst-case, they achieve optimal error at every dataset.

For example, in aggregating individual deposits, the traditional DP mechanisms would set an upper limit U for deposits and introduce noise proportional to U. To guarantee privacy, such U should be set as the deposit of the richest man in the world. This often results in impractical utilities due to such high deposit values seldom appear. A more desired approach would achieve instance optimal error, which is proportional to the maximum deposit present in the dataset. From this example, we can see instance optimal error can be much smaller than worst-case error.

He is one of the first researchers who study the instance optimal DP mechanism. In parallel with the research team led by Professor John C. Duchi at Stanford University, they have jointly established a theoretical foundation in this area. In the theoretical part, He first showed the traditional notation of instance optimality does not work in the DP literature and introduced some new notations. In the application part, He designed the instance optimal DP solutions to answer SQL queries and built machine learning models.

Recent works:

After his graduation, he continues to deploy practical private data analytics in more new problems and new areas. First, he explores how to deploy private data analytics in more realistic setting, such as distributed setting. Second, he combines the differential privacy with other secure computation models to provide stronger privacy protections in real-world data analytics. Third, he deploys DP in traditional scientific areas like bioinformatics and neuroscience.

Thanks to the department and supervisor:

The academic achievements he achieved today are inseparable from his supervisor Prof. Ke Yi and the Department of Computer Science at the Hong Kong University of Science and Technology.

Prof. Yi is the purest scholar and best supervisor Wei Dong has ever met. As the supervisor, he has been exceptional. His guidance during Wei Dong's early years as a Ph.D. student has been crucial in shaping my research skills, from identifying potential problems to designing solutions and writing academic papers. He is always very patient to correct my mistakes. Every decision he makes is totally from his consideration in students' development. As a pure researcher, his hard work and enthusiasm for research have been truly inspiring, setting an excellent example for Wei Dong to follow. As a friend, he has offered invaluable suggestions and warm-hearted encouragement whenever faced challenges. Without any exaggeration, he considers being supervised by Prof. Yi as one of the greatest blessings in his life.

As one of top institutions in computer science areas, department offers substantial support for his research. First, he must emphasize the valuable role that the department's courses have played in his academic development. For instance, the knowledge he gained from the "Combinatorial Optimization" course, taught by Prof. Sunil Arya, was important in the success of his SIGMOD best paper. Second, the department hosts many insightful talks each semester, significantly broadening his view and providing numerous research inspirations.

Feeling:

He is both honored and excited to receive this award. This recognition means a great deal to him, marking a milestone in his academic journey. This serves a big encouragement to him to continue to do the research. He is deeply grateful to the department for the support and resources, which have been crucial to his journey. This award also motivates him to keep pushing boundaries in the field.

About Chaoliang Zeng's Dissertation and His Sharing

Biography:

Chaoliang Zeng is now working on BitIntelligence in building scalable and performant transport ASICs and their software stack to support millions of GPUs or AI ASICs interconnect. As a researcher, his research interests include datacenter networking, hardware acceleration, and machine learning systems. He obtained his Ph.D. degree in 2023 from the Hong Kong University of Science and Technology (HKUST), supervised by Prof. Kai Chen. Before that, he received his B.S. degree with Outstanding Graduate Award in 2018 from University of Science and Technology of China (USTC).

Caoliang Zeng

Overview of his research:

His dissertation is "Towards High-Performance Datacenter Systems with Application-Oriented Optimizations". This dissertation advocates the first principle and application-oriented principle to design full-stack performant datacenter systems for big data, including the datacenter gateway, the serving system, and the analytics system. The systems proposed in the dissertation are groundbreaking and proven effective in production environments. They are accepted by the top networking and system conferences (NSDI and OSDI) and have been deployed in ByteDance datacenters to provide acceleration services for ByteDance businesses, including gateway, personal recommendation, advertisement, etc.

Breakthroughs of research:

His research achievements can be summarized from two aspects. First, novel hardware architecture design from the first principle. A hardware- accelerated gateway for stateful layer-4 load balancing, which maps different load balancing functionalities with different characteristics to the most suitable hardware. This hybrid architecture can benefit generic gateway functions, e.g., firewall and DDoS protection. Second, an FPGA-accelerated embedding-based retrieval system with performance-optimal architecture design. This architecture can accelerate generic vector search services.

The second aspect is application-oriented optimization. The dissertation proposes a novel scheduling method to accelerate neural recommendation training, which leverages the characteristics of data accesses during the training to reduce data transmissions. This scheduling method is applicable to general embedding model training under data parallelism.

Thanks to the department and supervisor:

Chaoliang would like to express his gratitude to his supervisor, Prof. Kai Chen, and the department. They provided him with the necessary resources and support, including funding, computing resources, guidance, and external connections, to conduct his research.

Recent work and feeling:

He feels honored to receive this award. It is a recognition of his research achievements and a motivation for his future career. After graduation, he now is designing and building scalable and high-performance transport processors for large-scale AI clusters.

About Tengfei Wang's Dissertation and His Sharing

Biography:

Tengfei Wang is currently a researcher at Shanghai AI Lab. He received his PhD in computer science from the Hong Kong University of Science and Technology in 2023, supervised by Prof. Qifeng Chen. His research interests focus on generative modeling and applications, including image generation and 3D content generation. Some of his works have been published in top-tier conferences such as CVPR and ICCV as Highlights and Oral presentations. He serves as reviewer for conferences and journals such as CVPR, ICCV, ECCV, SIGGRAPH, NeurIPS, ICLR, TPAMI, TOG. He served as assistant coach of HKUST ACM teams from 2019-2023.

Tengfei Wang

Overview of his research:

His thesis focuses on AI-based high-quality visual content creation. The demand for high-quality visual content, such as 2D images and 3D models, has increased significantly in recent years, driven by applications such as virtual reality, video games, animation, and interactive design. Unfortunately, creating such content can be a time-consuming and labor-intensive process that requires artistic expertise and proficiency in 2D painting or 3D modeling pipelines. To address this challenge, recent advances in deep generative models endows machine with the ability to create a diverse range of visual content.

Breakthroughs of research:

Inspired by the success of pretraining in visual understanding and natural language processing, this thesis focuses on a new generative paradigm that cultivates and leverages the knowledge captured by well-trained generative models to boost visual content creation. These generative models are learned to capture the manifold of natural images or 3D models in the pretraining stage, where large-scale training data are readily available. They can then serve as foundation generative models to boost versatile visual creation tasks. By leveraging the powerful capacity of these foundation generative models, we can unify various visual synthesis tasks and achieve unprecedented generation performance. This thesis highlights the benefits of using well-trained foundation generative models as a generative prior and demonstrating the results of various visual content creation tasks.

Thanks to the department and supervisor:

Tengfei would like to express his thankfulness to his supervisor, Prof. Qifeng Chen, and the CSE department to provide him many opportunities and a borader vision for his research.

"Qifeng has consistently been a genuinely warm and nice person, offering me an abundance of help and support throughout my academic journey. I still remember the moment four years ago when Qifeng extended the Ph.D. offer to me. At that time, I had little experience in research, yet he saw my potential and generously provided me with an incredible opportunity to grow and learn. Also, the CSE department has gave me many opportunities and a borader vision for my research."

Recent work and feeling:

He feels honored to receive this award. This serves as a big encouragement to him to continue to do the research. He is now focusing on the foundation 3D generative model, which can be applied in asset design, simulation and XR. They made the first 3D diffusion model that successfully creates high-quality 3D avatar in an automative way. Further, they extend this architecture to general objects generation, aiming for a foundation model for 3D generation.

About Zhibo Liu's Dissertation and His Sharing

Biography:

Zhibo Liu's research is centered on basic research on binary-level software security analysis and reverse engineering technology. With reverse engineering technology as the core driver, he is particularly invested in developing and enhancing software security analysis techniques. His work has been recognized by academia and industry, receiving the ACM SIGSOFT Distinguished Paper Award and presenting his work at the Black Hat USA conference.

Zhibo Liu

Overview of his research:

Software reverse engineering techniques, as one of the paramount cornerstones of cyber security, have been studied and developed for decades. However, some challenging problems are still fundamentally hard to solve. Zhibo's thesis aims to provide an up-to-date understanding of state-of-the-art reverse engineering tools that are indispensable in many security-related scenarios, including modern C decompilers and binary lifters, and shed light on further academic research and industrial development. Besides, with the rise of new deep-learning compilation technology, this thesis, for the first time, explores migrating traditional reverse engineering techniques to the emerging field of compiled DNN executables. Specifically, this thesis proposes the first decompilation work against DNN executables, pioneered software security analysis in the DNN executables field.

Breakthroughs of research:

His thesis provides a comprehensive understanding of state-of-the-art reverse engineering tools, including modern C decompilers and binary lifters. Besides, with extensive testing and assessing and considerable manual efforts, this thesis provides deep insights into the challenges and limitations of modern C decompilers and binary lifters. These results can be used to guide the following research work and practical industrial applications of reverse engineering tools.

Moreover, this thesis proposes the first decompiler for compiled DNN executables, enabling extraction of function DNN models from low-level binary coed that shows identical behaviors as the original models. As a seed work in the field of security software analysis of compiled DNN executables, his work will inspire a series of subsequent attacks and defenses and promote the progress of security research in the field.

Thanks to the department and supervisor:

Zhibo would like to express his gratitude to Prof. Shuai Wang and the members of the Cybersecurity lab.

"I would like to thank Prof. Shuai Wang for his dedicated guidance. He is not only a creative researcher but also an excellent educator. He inspired my research enthusiasm and academic pursuit of binary analyses and reverse engineering. The CSE department provides an excellent research atmosphere, and many academic lectures have broadened my research vision and helped me a lot in my research. I would also like to thank all the members of the Cybersecurity lab, with whom I have had very enlightening and inspiring discussions."

Recent work and feeling:

"I am very grateful to receive the honorable mention of the best dissertation award, which is a recognition of all my hard work and persistence during my PhD study. Also, I would like to thank my supervisor and friends for their support and help, without which I could not have achieved this achievement. This award will motivate me to continue my in-depth research on binary security and make more contributions to the academic community."

Zhibo's recent work provides the first side channel analyses targeting deep learning compilation techniques. He demonstrates the first cache side-channel attack against compiled DNN executables. This work poses a severe security threat to deep learning services deployed on clouds and will alert the DNN compilation community further to promote the development of related side channel defense techniques.