CSE Colloquium: Dr. Xing Hu

A HOLISTIC VIEW: DOMAIN-SPECIFIC ARCHITECTURES— SECURITY AND PERFORMANCE

Abstract

Domain-specific architectures (DSAs) emerge as an important research direction because of huge impetus from emerging applications and new opportunities introduced by new devices and technologies. Focusing on both the security issue and performance optimization, this talk introduces the architecture and system specializations for domain-specific applications.

Deep neural networks (DNNs), one of the most important domain applications, achieved great success in computer vision, speech recognition, and language processing, etc. Despite the rising opportunities for DNN techniques to benefit our life, the DNN attacks are very pernicious and could cause serious consequences. Therefore, DNN system security has emerged as an urgent and severe problem. Exploring the vulnerabilities is the initial step of establishing robust DNN systems. This talk will take a look at the existing attack models, such as model extraction, adversarial attack, and Trojaning attack. Along the way, we explore and reveal the potential vulnerabilities of DNN systems from the algorithm, system stack, and hardware perspectives and discuss the possible defensive solutions. Additionally, this talk briefly introduces the new computing and storage paradigm to boost the performance of other important domain applications, such as database, graph analytics, and bio-informatics.

Speaker’s Biographical Sketch

Xing Hu is currently a Postdoc working at Scalable Energy-Efficient Architecture Lab (SEAL), Department of Electrical and Computer Engineering, University of California, Santa Barbara. Dr. Hu received her Ph. D. from University of Chinese Academy of Sciences in 2014. Before she joined the University of California, Santa Barbara, she worked in Huawei Technologies on the topics of future memory systems. Her research interests reside in the intersection of domain-specific architectures, deep learning, and security. The highlight of her research is that it rethinks the domain-specific architectures in a holistic view and cooperates the optimizations across the entire stack including emerging device/circuit, architecture, system, and algorithms. Her research studies have been published and presented at top conferences or journals in Computer Systems (ASPLOS), Computer Architectures (MICRO, HPCA), Electrical Design Automation (DAC, DATE), and Algorithms areas (ICLR, Neural Networks).