This research talk delves into two significant advancements in hardware security and machine learning. Firstly, we explore the hardware security for Internet of Things (IoT) devices, presenting a novel Cost-Efficient Reliable Reconfigurable Ring Oscillator Physical Unclonable Function (CERRO PUF). CERRO PUF proves to be a promising solution, significantly reducing design overhead and power consumption while maintaining high efficiency and security. Through detailed analysis, we demonstrate its superiority over existing designs, showcasing improved challenge-response pair generation and heightened resistance against machine learning attacks.

In the second part of the talk, we shift our focus to data augmentation, a crucial strategy in overcoming the scarcity of training data for machine learning models. Generative Adversarial Networks (GANs) have emerged as powerful tools for data augmentation, producing realistic and diverse synthetic data. Our study introduces a two-player game approach to GAN training, iteratively refining the generator to create increasingly authentic samples. Furthermore, we present MAGAN, a Meta-Analysis method for GANs' latent space, shedding light on its influence on the generated image space. Quantitative results from MAGAN demonstrate its accuracy in tracing latent space changes, affirming the potential of GANs as parameterized data generators for data-driven augmentation, addressing the challenge of limited labeled datasets during model training.

Memory Bandwidth bottlenecks are the single most pressing challenges in digital computing hardware today. Memory-centric computing is an emerging non-conventional computing paradigm that promises to overcome these bottlenecks with a view to improving energy efficiency, parallel computing performance, and latency of the computing systems. The speaker's research endeavors to tackle inherent design challenges in this computing paradigm with the goal of enhancing the flexibility, efficiency, and performance of memory-centric accelerators. The speaker will discuss his previous research on developing a DRAM-based near-memory computing architecture featuring a novel look-up table (LUT) cluster-based programmable/; re-configurable processing architecture, tightly integrated within the DRAM banks for unlocking maximum data-communication bandwidth and minimizing data communication overheads. These architectural innovations in memory-centric computing have led to outstanding performance gains and energy efficiency for data-intensive and data-parallel applications, including Deep Neural Networks and Cryptography acceleration, as well as created a platform for efficient, real-time online learning within the memory device with minimal data relocation.

The speaker will also outline his future research goals of broadening the application spectrum of memory-centric architectural solutions and scaling up these systems to meet the demands of exponentially growing AI workloads, such as Generative and Multimodal AI, Graph Neural Networks, and Full Self-driving Algorithms. Additionally, the speaker will share his plans to extend his research to low-power edge computing applications and memory-oriented security of data and AI algorithms. In addition to outlining his future research directions, the speaker will discuss his strategies for securing funding and forging collaborations to pioneer cutting-edge research in the emerging field of memory-centric computing. Furthermore, he will share insights and experiences with teaching and course development alongside his future teaching interests and visions.

Modern energy systems (e.g., electric ships, aircraft, and vehicles) are increasingly designed with interconnected power converters that are tightly packaged to form an extensive network. It enables high-efficiency multi-functional operations by integrating numerous fast-switching wide bandgap (WBG) devices. However, such a complex network could create a large common-mode (CM) electromagnetic interference (EMI) flowing into common chassis or ground. This can make unprecedently large but unknown background noise. Despite increasing background noise in a network, the state-of-the-art CM EMI characterization efforts have focused mostly on single discrete devices or single packaged modules until recently. However, a limited study has been done on a large background CM EMI noise, especially under extensive networks. Furthermore, it has not been fully characterized; it has become a major technical bottleneck in electrified transportation. Due to a limited understanding of such background noise dynamics, most resort to bulky and costly passive filters with considerable tradeoffs between size, weight, and cost. These are not viable. This talk will address such emerging EMI issues that have not been discussed yet in the state of the art but are already causing problems in industry. The expected level of intended audience is with entry and intermediate backgrounds in electrical and mechanical engineering.

Dr. Seungdeog Choi is an Associate Professor in the Electrical and Computer Engineering Dept at Mississippi State University (MSU). He joined the university in Fall 2018 and is in his 6th year of service. He received B.S. from Chung-Ang University in 2004, M.S. from Seoul National University in 2006, and Ph.D. at Texas A&M University, College Station, TX, in 2010. He was a research engineer with L.G. Electronics, Seoul, Korea, 2006-2007, and Toshiba International Corp., Houston, TX, 2010-2012. He was an assistant professor at the University of Akron in 2012-2018. The keywords of his research is “Reliability, Efficiency, and Power Density.” He has published around 170 articles, including over 10 U.S. patents in the area. His research has been widely sponsored by the federal government, foundation, state government, and industry.

In today's digital era, computer systems are increasingly integrated into our daily lives and industrial operations. In pursuit of computer systems with higher computational performance and energy efficiency, the complexity of computer hardware's design and implementation has grown significantly. Such complexity, coupled with the intricate interactions between hardware and the physical world, inevitably introduces numerous hardware vulnerabilities. These vulnerabilities not only pose challenges to system security but also highlight the critical need for research focused on identifying and mitigating potential threats originating from hardware vulnerabilities. In this talk, I will present both the attacks exploiting hardware vulnerabilities and the defense strategies against hardware attacks. Specifically, I will detail a side-channel attack that exploits electromagnetic (EM) emanations from GPUs to infer computational activities on GPU and an intentional electromagnetic interference (IEMI) attack that demonstrates the potential for manipulating touchscreen-based devices via externally introduced electric fields. Furthermore, I will introduce a novel defensive technique that leverages EM side-channel information from DRAM to detect and mitigate Rowhammer hardware attacks. Finally, I will outline my future research plans, emphasizing their potential contributions toward advancing the security of future computer systems.

This seminar will introduce the concepts of self-healing and intelligence in power electronic systems for safety-critical, mission-critical, and resilient energy systems. The talk will then focus on how diagnostic, prognostic, and reconfiguration methods are applied to multi-level converter topologies that are characterized by inherent redundancy. Modeling of the converters' behavior under healthy, faulty, and reconfigured conditions will be presented. Logic-based time-domain methods, data-driven machine learning methods, along with necessary signal processing and switched control methods are introduced to demonstrate how those and other methods contribute to intelligent power electronics. Approaches to achieve self-healing and intelligence at the device-, converter-, and system-level are introduced. Various application domains that would benefit from intelligent power electronics are introduced. Those domains include electrified transportation, defense, grid integration, manufacturing, autonomous systems, and others. Relevant curriculum and research synergies will also be introduced.

Prof. Ali Bazzi is the Charles H. Knapp Associate Professor in Electrical Engineering at the University of Connecticut (UConn), and has been at UConn since 2012. He is the founder and director of the Power Electronics and Drives Advanced Research Laboratory (PEARL) at UConn. He is the co-founder and VP for Energy Conversion and Propulsion of Valcon Labs, a deep-tech start-up company that provides energy storage and energy conversion solutions for aerospace and autonomous vehicle applications. He briefly worked at Delphi Corporation and Bitrode Corporation before joining UConn. He received his PhD in 2010 from the University of Illinois at Urbana-Champaign, and his ME and BE degrees from the American University of Beirut, Lebanon, in 2007 and 2006, respectively. His research interests are in the control, design, diagnostics, and reconfiguration of power electronic systems, with focus on self-healing, intelligence, and fault-tolerance goals. He has over 130 peer-reviewed journal, magazine, and conference articles (mostly in IEEE), and 10 issued US patents. He has led or participated in over $10.5M in sponsored research funding from US Federal, State, and industry sponsors, and received the NSF CAREER Award in 2018. He was recently selected as a member of the Connecticut Academy of Science and Engineering. He received the Teaching Achievement Award and the Research Achievement Award from the UConn ECE Department in 2014 and 2023, respectively. He is an IEEE Senior Member and has served in many leadership and editorial positions in the IEEE Power Electronics Society and its various conference and publications committees.

Cyber-physical systems (CPSs), engineered systems which include sensing, processing, control, and communication in physical spaces, are ubiquitous in modern critical infrastructures including manufacturing, transportation systems, energy delivery, health care, water management, and the smart grid. As the physical, communication, and computational parts of these complex systems become increasingly interconnected and intertwined, it is important to ensure their safety and security in the presence of uncertainties and attacks. In this talk, we present three necessary components for designing safe and secure CPSs: detecting attacks, responding to attacks, and providing safety guarantees for systems with unmodeled dynamics. In particular, we introduce the moving target defense, software rejuvenation, and data-driven reachability, showing how each of these tools leverage knowledge of the underlying physical dynamics to guarantee safety and security. We illustrate our results in a number of example applications and present future research directions for data-driven CPS safety and security.

Paul Griffioen is currently a postdoctoral researcher in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley working with Murat Arcak. He received M.S. and Ph.D. degrees in Electrical and Computer Engineering from Carnegie Mellon University in 2018 and 2022, respectively, where he was co-advised by Bruno Sinopoli and Bruce H. Krogh. He received a B.S. degree in Electrical and Computer Engineering from Calvin College in 2016. His research interests include the modeling, analysis, and design of active detection techniques and response mechanisms for ensuring resilient and secure cyber-physical systems. His research interests also include data-driven analysis and design of high-performance cyber-physical systems that ensure safety while operating under computational constraints.

Rapid advances in sensing and computing technologies have led to the proliferation of Cyber-Physical Systems (CPS). However, increasing device complexity, shrinking technology sizes, and shorter time to market have resulted in significant challenges in ensuring the reliability, safety, and security of CPS\null. Significant efforts have been made using techniques such as run-time verification, monitoring, and anomaly detection. However, these approaches cannot maintain high detection accuracy, considering complex system dynamics and unpredictable human behaviors in the control loop, and often detect the occurrence of hazards late, which may not leave enough time for successful mitigation. In addition, there is often a gap between the safety properties checked at run-time and the safety requirements specified at design time, which are usually based on ad-hoc and fixed rules and do not account for the multi-dimensional context in the CPS, including physical processes, the environment, the cyber components that affect the physical processes, and their interactions in both temporal and spatial domains.

In this talk, I will present my research addressing these challenges through a hybrid knowledge and data driven approach to context-aware safety assurance in CPS. First, I will discuss vulnerabilities in safety-critical CPS and introduce a formal framework for control-theoretic specification of safety requirements. Then I will present two combined knowledge and data driven approaches to refining safety specifications for run-time safety monitoring and hazard mitigation, and design-time safety validation. Finally, I will illustrate my visions for the future, engineering the next generation of CPS that are safer and more robust through automatic, adaptive, and trustworthy safety assurance.

Xugui Zhou is a Ph.D. candidate in Electrical and Computer Engineering at the University of Virginia, advised by Prof. Homa Alemzadeh. Before Joining UVA in 2019, he was a project manager and senior engineer at State Grid and researched power grid protection technology. Prior to that, he obtained his B.Eng in Automation and M.Eng in Control Science and Engineering from Shandong University, China, in 2012 and 2015, respectively. His research interests are at the intersection of computer system security and control system engineering by drawing techniques from formal methods and machine learning. His work has appeared at top-tier venues, including AAAI, DSN, and IEEE TDSC\null. He has received Rising Star Award in CPS 2023, Carlos and Esther Farrar Graduate Fellowship Award, Google Ph.D. Fellowship Internal Selection, and GTRI Focus Fellowship Award, and is the inventor of three international patents.

Federated learning is a distributed machine learning approach that enables multiple clients (e.g., smartphones, IoT devices, and edge devices) to collaboratively learn a model with help of a server, without sharing their raw local data. Due to its potential promise of protecting private or proprietary user data, and in light of emerging privacy regulations such as GDPR, federated learning has become a central playground for innovation. However, due to its distributed nature, federated learning is vulnerable to poisoning attacks. In this talk, we will discuss local model poisoning attacks to federated learning, in which malicious clients send carefully crafted local models or their updates to the server to corrupt the global model. Moreover, we will discuss our work on building federated learning methods that are secure against a bounded number of malicious clients.

Minghong Fang is a Postdoctoral Associate in the Department of Electrical and Computer Engineering at Duke University. He earned his Ph.D. in the Department of Electrical and Computer Engineering at The Ohio State University. His research interests lie broadly in the span of machine learning, security, privacy, with a recent focus on the intersection among them. He is also interested in the distributed optimization for learning and networking. His research has been published in top-tier security, machine learning and networking venues, such as USENIX Security, NDSS, ICLR, The Web Conference (WWW), MobiHoc, etc. His USENIX Security 2020 paper has been selected as one of the “Normalized Top-100 Security Papers since 1981”.

With the rapid increase in data collected from various edge devices across distributed networks, there is a pressing need for innovative solutions to harness intelligence at the network edge. Traditional cloud-based centralized learning methods won't suffice. Instead, federated learning, an emerging approach, keeps data local at its source, avoiding the need for centralization on a cloud server. This method pushes model updates to the edge and aggregates local updates to train a global model on a shared parameter server. However, federated learning presents challenges, including poor model convergence compared to centralized learning and device lag due to heterogeneity and network unreliability. To empower distributed edge intelligence efficiently, optimizing machine learning models to utilize decentralized data, adapting to diverse device capabilities, and complying with network constraints are crucial. In this talk, I will delve deeper into these insights and share my research on bridging the gap between centralized and federated learning. My strategy encompasses three key areas: (i) enhancing processor utilization through relaxed synchronization and tackling memory-efficient problems in distributed networks, (ii) developing parallel algorithms that accelerate model learning via data summaries and facilitate linear scaling in decentralized machine learning, and (iii) co-designing energy-efficient systems that make AI accessible at the edge, promoting green AI. To wrap up my talk, I will share some intriguing directions for future research.