Faculty Presentations

The widespread adoption of Artificial Intelligence (AI) has led to a plethora of security and privacy vulnerabilities that need to be understood and mitigated. In this talk, we will explore some of the most significant privacy issues associated with AI. Specifically, the talk will examine the privacy challenges of federated learning, a collaborative machine learning approach where multiple participants train a model without sharing their data. Although federated learning is a promising technique for preserving privacy and promoting data decentralization, our recent research has revealed that private information can still be leaked through shared gradient information, even under certain defense settings (e.g., local differential privacy). By examining the emerging attack surfaces, we aim to raise public awareness about the vulnerabilities that exist in current AI-based systems, and emphasize the need for prioritizing their mitigation to prevent potential attacks.

Novel connected and automated vehicle (CAV) technologies are transforming the way drivers, vehicles, and traffic infrastructure interact in real-world environments. CAVs use vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications to coordinate traffic interaction to minimize congestion, maximize fuel efficiency, and reduce collisions. Despite the fact that the power of AI and machine learning are revolutionizing the potential impact of CAVs, much more research is required to understand how data-driven approaches interact with model-based components to better enable traffic mobility and energy savings. This two-year StART project aimed to bridge this gap by studying a novel traffic control framework intended to reconcile data-driven traffic signal phase and timing (SPaT) control and model-based speed control of CAVs. In the first part of the project, we aimed to investigate the possibility of generating optimal SPaT information using historical and real-time traffic data. State-of-the-art deep learning and reinforcement learning techniques allow us to develop scalable, adaptive, data-driven algorithms that predict optimal SPaT using extracted information from vehicular networks. In the second part of the project, we aimed to study efficient approaches to optimizing CAV movements and providing optimal speed advisories based on the generated SPaT data. Recent advances in computational optimal control and convex optimization allow us to exploit the SPaT data and develop efficient speed control strategies to improve the mobility and fuel economy of CAVs. The developed models and algorithms were demonstrated via simulations using the Simulation of Urban MObility (SUMO) simulator.

Approximate computing represents a fresh approach to computation, aiming to yield less precise outcomes by easing the demand for absolute precision or complete determinism. In the realm of quantum computing, approximate computing offers the opportunity to compromise circuit accuracy in favor of optimizing parameters like noise fidelity, quantum depth, and resource utilization. This strategy proves especially advantageous in quantum computing application domains where exact results are unnecessary, such as machine learning, signal processing, and image processing. This presentation will delve into the realm of approximate quantum arithmetic circuits, which serve as fundamental components for complex applications. Additionally, we will provide an overview of a design automation tool capable of generating approximate quantum circuits that surpass the accuracy of baseline exact quantum circuits.

Vehicles are generating more data than ever, and this study seeks to find a way to leverage rich vehicle kinematics data in combination with driver biometrics and information about the surrounding traffic environment to create a tool for reliable estimation of the driver’s state of distraction while operating a vehicle. The tests were conducted in a safe and immersive virtual reality environment where participants driving a simulated vehicle (using a head-mounted-display and natural controls, such as accelerator, brake pedals and steering wheel) were visually distracted by an increasingly complex task. The visual distraction was meant to emulate use of cell phones or operating a touchscreen interface and achieved by having the driver repeatedly find the randomized location of a differently-oriented arrow in a grid of otherwise similarly-directed arrows. Environmental data (such as relative distance and relative speed of traffic around the ego vehicle, deviation from the lane center-line), vehicle kinematics data, (such as speed, acceleration, steering input, accelerator and brake effort) as well as driver-biometric data (such as gaze direction, galvanic skin response, pulse rate) were collected in real time and synchronized. This combined time-series data was then used to train a multivariate time series feature extractor (WEASEL+MUSE) and a logistic regression classifier. Relative importance of the features in the dataset was examined by permuting through and quantifying the classification accuracy of combinations of features. Most importantly, the data shows that by combining the different modalities, a higher predictive accuracy can be attained than when using any one modality.

Radiation that damages DNA can be both dangerous to healthy tissues and used as a cancer therapeutic. To develop therapeutics that destroy cancer while minimizing damage to surrounding tissues, it is critical to understand, and even manipulate, how different cell types respond to radiation. A cell’s sensitivity to radiation is in part affected by how the DNA is arranged in 3D inside the nucleus. In the first year of our StART grant, we have combined computational models (ORNL) with experimental work (UTK) to investigate how differences in chromosome folding influence the amount of damage caused by radiation in different cell types. To investigate the influence of the initial 3D chromosome folding state on radiation damage, we have pretreated healthy and cancerous epithelial cells (MCF-10A vs. MCF-7) with a drug (histone deacetylase inhibitor, TSA) that decondenses chromosome structure. The drug by itself caused no significant cellular toxicity in either cell line and did not induce any DNA damage. However, pre-treatment with this drug causes a radiosensitizing effect: cells experience more DNA damage when exposed to X-rays after TSA treatment than without this pre-treatment. This effect is more pronounced in cancer cells vs. the healthy cells, consistent with a higher initial level of open, acetylated chromatin in MCF-7 cells after TSA treatment. Our future work will measure the long term impact of this combined treatment, as well as additional chromatin structure disrupting treatments with X-rays, on cell survival, and compare these results with computational models. We will also compare the impact of X-rays with alpha particles (ORNL), which will be used in future targeted radiotherapeutics.

Alzheimer’s Disease (AD) is a complex neurodegenerative disorder in which the underlying mechanisms for disease initiation and progression are not understood. Researchers have identified two hallmarks of AD, neurofibrillary tangles of hyperphosphorylated tau and amyloid plaques of amyloid-β (Aβ), yet there are no effective FDA-approved treatments. Importantly, plaques occur in elderly patients without AD diagnosis, and elderly patients without plaques can have impaired cognition. This underscores the need for a basic science approach to understanding the molecular mechanisms of AD and integrating molecular studies with observable cognitive deficits. Additionally, while the risk and prevalence of AD increase with age and two-thirds of patients with AD are women, the impact of age and sex on brain biochemistry remains unclear.Without a fundamental understanding of the intersectional effects of age, sex and genotype, an effective treatment for AD and other dementia-related neurological diseases will remain elusive. The overall goal is to use a multimodal protocol with explainable-AI to map phenotypic signatures of the brain with observed cognitive deficits to identify molecular changes due to aging, sex, or genetics. We expect to uncover molecular changes not previously associated with AD, aging, or sex. We will use computational algorithms to integrate data from cognitive testing followed by mass spectrometry imaging, liquid chromatography coupled with ion mobility spectrometry mass spectrometry of aggregated cells, and spatial RNA sequencing of brain sections to create a 2D molecular map. We have found significant differences in lipids in aged vs. young mouse brains across genotypes. We propose to fill these critical gaps in fundamental knowledge by taking an unbiased molecular approach to investigating the spatial distribution of brain molecules combined with translationally-relevant behavioral indicators of cognitive decline. This approach will provide new information on the initiation and progression of biomolecular changes across brain regions underlying cognitive decline in AD.

A fundamental understanding of materials response to coupled extreme environments is necessary for everything from long term manned interstellar missions to the design and production of advanced fusion and fission energy systems. The Tennessee Ion Beam Materials Laboratory (TIBML) is a University of Tennessee core facility that has undergone a rapid evolution over the last year. TIBML has a long and rich history as a world-class Ion Beam Analysis (IBA) facility for elemental analysis of surface and subsurface regions. Over the last year, the facility has undergone a rapid growth phase to expand the ability to explore the response of materials to coupled extreme environments. These environments include ion irradiation, ion implantation, heating, cryogenic cooling, and soon mechanical testing. Work is underway to couple these extreme environments. This presentation will highlight the evolution of the facility, detail the current capabilities, and discuss the potential applications in materials ranging from biological samples through quantum devices to additive manufactured refractory high entropy alloys. The presentation will end with how ion beam modification can be used for tailored far-from-equilibrium nanoscale structures, as will be demonstrated with kryptonite as an exemplar.