��ݮ��Ƶ

A DPU is a network interface card (NIC) with programmable compute and memory resources. It sits on the system bus, PCIe, which is the fastest path to access SSDs, and it directly connects to the network. It therefore can process storage requests as soon as they arrive at the NIC, rather than passing them through to the host. DPUs are widely deployed in public clouds and will soon be ubiquitous.

In this talk, we’ll describe DPU-Optimized Disaggregated Storage (DDS), our software platform for offloading storage operations from a host storage server to a DPU. It reduces the cost and improves the performance of supporting a database service. DDS heavily uses DMA, zero-copy, and userspace I/O to minimize overhead and thereby improve throughput. It also introduces an offload engine that can directly execute storage requests on the DPU. For example, it can offload GetPage@LSN to the DPU of an Azure SQL Hyperscale page server. This removes all host CPU consumption (saving up to 17 cores), reduces latency by 70%, and increases throughput by 75%. This is joint work with Qizhen Zhang, Badrish Chandramouli, Jason Hu, and Yiming Zheng.

Vector databases have recently emerged as a hot topic due to the widespread interest in LLMs, where vector databases provide the relevant context that enables LLMs to generate more accurate responses. Current vector databases can be broadly categorized into two types: specialized and integrated. Specialized vector databases are explicitly designed for managing vector data, while integrated vector databases support vector search within an existing database system. While specialized vector databases are interesting, there is a significant customer base interested in integrated vector databases for various reasons, such as reluctance to move data out, the desire to link vector embeddings with their source data, and the need for advanced vector search capabilities. However, integrated vector databases face challenges in performance and interoperability.

In this talk, I will share our recent experience in building integrated vector databases within two important classes of databases: Relational Databases and Graph Databases. I will show how we address the performance and interoperability challenges, resulting in much more powerful database systems that support advanced RAGs. Next, I will present other challenges in vector databases along with our ongoing work. Finally, I will discuss the broader role of database systems in the era of LLMs and explore how to build future databases that extend beyond vector databases to better support LLMs.

Jianguo Wang is an Assistant Professor of Computer Science at Purdue University. He obtained his Ph.D. from the University of California, San Diego. He has worked or interned at Zilliz, Amazon AWS, Microsoft Research, Oracle, and Samsung on various database systems.

His current research interests include database systems for the cloud and LLMs, especially Disaggregated Databases and Vector Databases. He regularly publishes and serves as a program committee member at premier database conferences such as SIGMOD, VLDB, and ICDE. He also served as a panel moderator for the VLDB'24 panel on vector databases. His research has won multiple awards, including the ACM SIGMOD Research Highlight Award and the NSF CAREER Award.��

More information can be found at

Invisible Yet Powerful: Watermarking to Protect Datasets and Models in Machine Learning

The rapid advancement of AI has transformed both datasets and models into valuable assets, yet they remain vulnerable to unauthorized use, theft, and replication. Watermarking provides a promising solution by embedding verifiable ownership signals to establish ownership protection. Traditional database watermarking techniques assume that attackers seek to preserve query utility, which inherently restricts the extent of modifications they can apply to the data. However, this assumption does not hold for machine learning, where models can maintain predictive performance even when trained on significantly altered datasets. As a result, adversaries can heavily modify a dataset or distill a model while preserving its learning utility, which enables much stronger watermark removal attacks than those in traditional database watermarking.

How can we design watermarking methods that safeguard AI-related assets against these threats while maintaining their usability?��

This talk presents our recent research on addressing the novel challenges in watermarking tabular datasets and deep learning models in the context of machine learning. First, I will introduce TabularMark, a non-blind watermarking framework that embeds verifiable ownership signals into tabular datasets while ensuring that models trained on watermarked data retain high predictive performance. Second, I will discuss blind watermarking for numerical tabular datasets, which enables watermark verification without requiring access to the original data, making it more practical for real-world data-sharing scenarios. Third, I will introduce a robust model watermarking approach that embeds ownership signals into deep neural networks to withstand ensemble distillation attacks. Finally, I will conclude with open challenges and future directions.

Lingyang Chu is an Assistant Professor in the Department of Computing and Software at McMaster University. He received his Ph.D. in Computer Science from the University of Chinese Academy of Sciences. Before joining McMaster University, he was a postdoctoral fellow at Simon Fraser University and a Principal Researcher at Huawei Technologies Canada.

His research focuses on data mining, explainable machine learning, and trustworthy computing, with a growing focus on data security in database systems. Some of his recent works explore AI-related data watermarking techniques to ensure data integrity and provenance in large-scale systems and data markets. He is an Associate Editor of ACM Transactions on Knowledge Discovery from Data (TKDD) and he also served as a program committee member and reviewer for conferences and journals including SIGMOD, VLDB, KDD, ICDE, ICDM, CIKM, CVPR, NeurIPS, ICML, ICLR, ACM Multimedia, TKDE, TMM, etc.