Poster title
Bootstrapping Trust from Zero-Trust for IoT Devices Using In-Band Wireless Pairing

Investigator
Shucheng Yu

Abstract
IoT devices are increasingly deployed in emerging smart systems. However, resource-constrained IoT devices also become security bottlenecks of complex systems which can lead to more severe cyberattacks. Due to the complex supply chain of IoT devices, configuring their security in prior is difficult if not impossible. In this research, we address the problem of establishing initial trust among ad hoc wireless IoT devices using wireless physical signals. Specifically, we designed an authenticated key agreement (AKA) protocol that assumes no prior trust among IoT devices and can withstand strong attacks such as RF signal modification attacks. Our protocol only requires common wireless communication interfaces of IoT devices and is therefore compatible with commodity-off-the-shelf devices. It is lightweight and can also serve as an efficient continuous authentication protocol for IoT devices.

Poster title
CirSTAG: Circuit Stability Analysis via Graph Neural Networks

Investigator
Zhuo Feng

Abstract
Circuit stability (sensitivity) analysis aims to estimate the overall performance impact due to the perturbations of underlying design parameters (e.g., gate sizes, capacitance variations, etc.), which remains challenging since many time-consuming circuit simulations are typically required. On the other hand, graph neural networks (GNNs) have proven to be effective in solving many chip design automation problems, including circuit timing prediction, parasitics prediction, gate sizing, and device placement. This paper presents a novel approach (CirSTAG) that exploits GNNs to analyze the stability (robustness) of modern integrated circuits (ICs). CirSTAG is based on a spectral framework for analyzing the stability of GNNs leveraging input/output graph-based manifolds: when two nearby nodes on the input manifold are mapped (through a GNN model) to two distant nodes (data samples) on the output manifold, it implies a significant distance mapping distortion (DMD) and thus poor GNN stability. CirSTAG computes a stability score equivalent to the local Lipschitz constant for each node/edge, considering both graph structural and node feature perturbations, which immediately allows for identifying the most critical (sensitive) circuit elements that would significantly alter the circuit performance. Our empirical evaluations on various timing prediction tasks with realistic circuit designs show that CirSTAG can truthfully estimate each circuit element’s stability under various parameter perturbations, offering a scalable approach for assessing the stability of large IC designs.

Poster title
Combining Block-Based Programming with Robotics Kits to Support a Middle School Computing Curriculum

Investigator
Bernard Yett

Abstract
Goals of past research were to develop a comprehensive curriculum with hands on components, progressively teach students CT, networking, and network security, develop systematic methods to evaluate the effectiveness of the curriculum through learning analytics (LA) and results from assessments, and encourage student growth in computing interest and confidence, task value, and other attitudes. Next steps could include improvements to the existing curriculum to further scaffold student learning and better study the robot's role in student learning and engagement. Alternative directions would explore different content areas such as space sciences while still utilizing the block-based programming platform and various included tools.

Poster title
Computationally Efficient Neural Networks Architecture for Robust Classifiers

Investigator
Cristina Comaniciu

Abstract
The rapid advancement of neural network technologies has catalyzed transformative changes across various application domains. Despite these successes, neural networks often grapple with significant challenges related to robustness and reliability. We introduce our novel Pseudo-Invertible Neural Networks (Psi-NN) architecture, which incorporates components and training procedures that enable bidirectional data processing. Psi-NN are a class of neural network components and training procedures which construct models that simultaneously implement an approximate inverse operation to their forward function. This capability allows a single model to learn a task along with its approximate inverse, which usually requires two separately parameterized models. Our generalized approach allows this method to be applied to a wide variety of neural network architectures and application areas because, as opposed to a true inverse, our pseudoinverse layers can be added as drop in layers in an existing network architecture which was not designed for such functionality. In particular robust classifiers based on autoencoder structures can be implemented using Psi-NN architectures, cutting the number of trainable parameters in half.

Poster title
Computational Imaging and Object Detection with Quantum LiDAR

Investigator
Hong Man

Abstract
The aim of this project is to achieve a good quality 3D image reconstruction with lower photons per pixel requirement, shorter sensing time, lower optical power consumption, and improved noise resilience. We are also developing a practical solution based on start-of-the-art machine learning techniques for objection detection and recognition in Quantum LiDAR data. In particular, our focus is to support the training of the deep learning algorithm using only synthetic point cloud data from publically available CAD models

Poster title
Dissipativity-Based Controller and Topology Co-Design for DC Microgrids

Investigator
Shirantha Welikala

Abstract
The rising demand for DC loads, such as electric vehicles, renewable energy sources, and data centers, has led to increased recognition and use of DC microgrids (MGs) compared to AC microgrids, as DC MGs offer greater efficiency, scalability, and reliability. Most existing work on designing controllers for DC MGs typically treats controller design and communication topologies separately. This poster presents a co-design strategy for distributed controllers and communication topology in DC MGs using a dissipativity-based approach. In this approach, the optimization of the design variables representing controllers and topology of the DC MG is formulated as a linear matrix inequality (LMI) problem. Then, both centralized and decentralized approaches to solve this LMI problem are discussed. Unlike the existing works, in which the controllers are determined while treating the topology as fixed and predefined, our proposed approach determines the controllers and topology simultaneously. The simulation results show the effectiveness of this approach in optimizing data communication while ensuring voltage restoration and proper current sharing. The ongoing research explores the decentralized co-design of controllers and topology for AC and Hybrid MGs.

Poster title
Enhancing Energy Efficiency and Smart Grid Development For Energy Sustainability, Reliability, and Resilience

Investigator
Lei Wu

Poster abstract
The increasing need for an affordable, clean, and reliable electric energy supply has necessitated grouping power generation from heterogeneous resources. However, facing the changing resource mix, the rapid deployment of these resources would significantly transfer the landscape of the emerging power grid and bring new planning, operation, and control challenges. There is an urgent need to develop next-generation tools with innovative models and scalable algorithms that can solve the planning, operation, and control challenges of future power grids with heterogeneous supply and demand resources. This poster presents several research and development projects by our team on: (i) long-term planning on EV charging station planning with user equilibrium and driver behavior; (ii) short-term operation with enhanced flexibilities to boost grid reliability with renewables; (iii) real-time coordinated control of VSC-HVDC integrated offshore wind farms; and (iv) real-time control of microgrids with 100% clean energy.

Demo title
Extremely High-Temperature mm-Wave Fiber Communications

Investigator
Rod Kim

Demo abstract
We will demonstrate the application of millimeter-wave (mm-wave) ceramic fiber made of low-loss alumina (ϵr = 9.8 and tan δ = 0.003) exhibiting excellent thermal tolerance and isolation from room temperature to 1100°C. Such high-temperature mm-wave interconnects will provide an additional dimension to the extreme environment electronics used across many industries, including aerospace, avionics, and oil industries, to mention a few. In particular, we will pair ceramic fibers with a 60-GHz CMOS transmitter (Tx) and receiver (Rx) to demonstrate extreme temperature communications. We implemented an automatic gain control (AGC) loop embedded in a front-end low-noise amplifier (LNA) to compensate for the temperature-dependent variation across a wide temperature range, achieving a 5-Gb/s data rate via 60 cm ceramic fibers.

Poster title
Full Radio Spectrum Awareness Using Deep Learning

Investigator
Yu-Dong Yao

Abstract
With the advancement of AI, machine learning (ML), and deep learning (DL) research, widespread applications have been found in various engineering domains. Our research focuses on AI/ML/DL applications in wireless communications and mobile computing in terms of radio spectrum awareness. Specifically, we investigate wireless systems identification and classification issues, which include radio spectrum sensing, fading channel statics estimation, signal modulation format identification, signal transmission protocol classification, and cellular network identification. The research of radio spectrum awareness finds many applications in adaptive transmission for spectrum efficiency and reliability and in wireless networks for security and access control.

Poster title
Large-scale Deep Reinforcement Learning with Serverless Computing

Investigator
Hao Wang

Abstract
Deep reinforcement learning (DRL) algorithms, are widely applicable in many different areas like scientific simulations, robotics, autonomous driving, and large language model (LLM) development. However, DRL training is computationally expensive, requiring numerous trials and errors, and consuming substantial computing resources and time. From an algorithmic perspective, the stochastic nature of environment dynamics can cause some actors to complete episodes sooner, resulting in idle periods while waiting for other actors to finish. From a systems perspective, actors remain idle during the policy update by the learner, significantly wasting computing resources and amplifying training costs. We propose to leverage serverless computing architectures to address the fundamental challenges in large-scale distributed DRL training. Serverless computing, also known as Function-as-a-Service (FaaS), is a cloud computing model that employs lightweight containers as execution units. The instant execution and auto-scaling capabilities of serverless computing naturally meet the highly dynamic resource demands of DRL training.

Poster title
Maximized Detection of Bees Using Deep Learning

Investigator
Kevin Lu

Poster abstract
In Fall 2023, Harsha Tangirala of the Applied Artificial Intelligence (AAI) Masters Program explored the application of object detection on various images of bees, specifically utilizing computer vision models like YOLOv5 and Faster R-CNN detection techniques. It aims to identify all the bees on the images so that the model is able to track them in a given commercial apiary, with the end goal being counting the number of bees for a single image snapshot. Counting bees can have many potential applications for gathering statistics with video tracking, so manual labor of beekeepers can be alleviated. This poster discusses various topics such as the field background, dataset preparation, image augmentation, hardware considerations, existing works in the field, research findings with YOLOv5 and Faster R-CNN, challenges encountered, future research directions, and finally potential real-world applications. The results show that YOLOv5 has an average precision score of 94.2%, with dataset augmentations yielding up to 95.9%, and 48.4% with Faster R-CNN. The future directions have also been identified.

Poster title
Misinformation Detection

Investigator
K.P. Subbalakshmi (Yupeng Cao)

Abstract
Misinformation has the potential to cause significant damage to public health, law and order, and democracy. This poster will discuss three topics (1) A method for detecting misinformation on COVID related misinformation and a dataset; (2) A claim detection architecture that emphasize relevant information and (3) LLM pipelines for detecting scientific misinformation "in the wild".

Poster title
ML-Aided Channel Prediction and Effective Beamforming for High Throughput UAV Networks

Investigator
Min Song

Abstract
Channel prediction at the transmitter side can significantly enhance communication efficiency in unmanned aerial vehicle (UAV) networks. Previous research in this area has been constrained by limitations in channel prediction and beamforming-based transmission, often relying on either perfect or outdated channel state information supplemented by feedback from pilot sequences, which leads to considerable overhead. This paper addresses these limitations by introducing a novel machine learning-assisted channel prediction and beamforming strategy for autonomous multiple-input multiple-output (MIMO) UAV networks. We present an Attention-assisted Long Short-Term Memory-based Channel Prediction (ALSTM-CP) model that utilizes a comprehensive time-series dataset, capturing the dynamic UAV environment to minimize prediction errors. Additionally, we propose a joint beamforming strategy employing an adaptive loss integration technique to improve system performance. A detailed analysis of system complexity across various configurations is also provided. Simulation results affirm that our integrated ALSTM-CP and beamforming approach outperforms benchmark strategies in prediction accuracy and overall transmission rate performance.

Poster title
Multitrack Read Channel Designs for Modern Two-Dimensional Magnetic Recording

Investigator
Elnaz Banan Sadeghian

Abstract
Two-dimensional magnetic recording (TDMR) is an emerging storage technology that aims to achieve areal densities on the order of 10 Tb/in2, mainly driven by innovative channels engineering with minimal changes to existing head and media designs within a systems framework. We develop efficient noise prediction, synchronization, and symbol detection algorithms that jointly detect multiple data tracks from one or more readback signals, in order to significantly increase the areal density and the throughput over the current industry’s single-track detection designs. We approach this problem from every promising direction, from purely communication-theoretic to machine learning models, according to our prior achievements and the latest research trends.

Poster title
Passive Localization of RF-Silent Mobile Objects in Multipath Environments

Investigator
Hongbin Li

Abstract
Many wireless communication signals, WiFi, cellular, radio/TV, and satellites, can be utilized to provide passive radio frequency (RF) sensing services as by-products, including object and event detection, localization, healthcare monitoring, etc. Passive RF sensing is capable of locating non-cooperative device-free objects, which are usually harder to locate than cooperative objects equipped with an active RF device (e.g., a mobile phone). Most prior passive solutions are non-coherent processing based, using signal strength or channel state information. Non-coherent methods suffer poor resolution, small coverage, and fluctuating performance due to their intrinsic physical limitations. They are ineffective in exploiting the larger bandwidth of newer wireless signals. Aimed to surpass the limitations of state-of-the-art non-coherent methods, this project is focused on computationally efficient coherent high-resolution techniques to locate passive RF-silent mobile objects in complex indoor and outdoor multipath environments, by harnessing bandwidth-rich wireless signals, multi-static sensing geometry, and new graph-theoretic techniques.

Poster title
Rock-SAT-C – Sounding Rocket Payload Experiment

Investigator
Joe Miles

Abstract
The goal of the 2022 to 2025 project is to develop a payload capable of obtaining atmospheric gas samples at predetermined time intervals and investigating the noble gas fractionation of the sample.

Mission Statement: In order to support and evaluate collection methods and atmospheric models for a concept sample collection mission to Venus, VATMOS-SR (Venus Atmospheric Sample Return Mission Concept), the primary goal of the project is to develop a payload capable of obtaining atmospheric gas samples at predetermined time intervals after Mach 1. Pressure sensors and a series of various valves will be used to flow atmospheric air into three sample tanks. The valves and a bleed system will ensure that each sample is as unique as possible and fills the tank to the highest extent. After the samples are returned to ground, the noble gas fractionation of the samples is to be investigated through mass spectroscopy via a research lab. The data can then be used to validate models to be used for the Venus mission.

Poster title
Semantic Map Based Robot Navigation with Natural Language Input

Investigator
Yi Guo

Abstract
There is an increasing demand for robots to assist humans in daily tasks, where the robot is expected to understand humans and respond to human instructions using natural languages. We present a new semantic map based robot navigation system in the paper. The system takes human voice input, processes multi-modal data including natural languages and RGB-D images, and generates semantic maps for robot navigation. Making use of recent development in image segmentation tools, we integrate robot mapping and localization with a customized real-time object detection model, so that the semantic and navigation layers are efficiently combined for robot navigation purpose. We demonstrate the performance of our developed algorithms in both simulation and real robot experiments. Compared with existing works, we demonstrate applicability to real robot system and superior performance in terms of success rate.

Poster title
SleepSentry: Early Detection of Sleep Apnea through Convolution Neural Network (CNN)

Investigator
Md Abu Sayeed

Abstract
Sleep Apnea is one of the sleep disorders that affect 5 to 10 percent of the world’s population; the number is astounding; records suggest that about 39 million adults suffer from sleep apnea only in the United States alone. It disrupts sleep either by partial stoppage or complete blockage of breathing. Current diagnostic solutions include polysomnography (PSG), which is an in-lab study while the patient sleeps. This is a very costly and time-consuming process that is also inaccessible to many. Hence, to address this problem, the research focuses on using machine learning algorithms to make early detections. As soon as the detection is made, the system integrates Internet of Things (IoT) to notify the concerned health officials, record all the readings for further studies, and provide comprehensive solutions for patient healthcare. The use of Convolution Neural Network (CNN) after filtering the test data, which is of a varied nature, makes the solution highly optimized and fast, and integration of smart-IoT solutions with Things Speak makes sure that the data is securely pushed to the servers for further analysis and notification purposes.

Poster title
Towards Understanding and Handling Problems Due to Coexistence of Multiple Internet of Things Platforms

Investigator
James Xiaojiang Du

Abstract
Many Internet of Things (IoT) platforms allow users to manage heterogeneous devices and install flexible automation logic for their smart environments, such as homes, offices, and laboratories. There are now over one hundred IoT platforms, and each has its own strengths and drawbacks. To get the best of each, users in a smart environment may use multiple IoT platforms. However, when multiple IoT platforms take effect in the same physical space, due to their coexistence, many unique problems arise due to their coexistence, which we call Multi-Platform Coexistence (MPC) problems. When multiple IoT platforms have views and control over IoT devices in the same smart environment, a variety of problems can arise.

For example, multiple platforms do not necessarily share consistent observations about the states of IoT devices. As a result, automation applications (apps, for short) installed on these platforms may issue incorrect or conflicting commands. As another example, under certain conditions, one app installed on a platform triggers the execution of another app on a different platform, causing confusing, undesired, or even hazardous automation. Existing works mostly assume the usage of a single IoT platform in a smart environment and, thus, cannot handle or reveal these problems. To obtain in-depth understanding of MPC problems, the proposed research will systematically describe, categorize, and measure them.

Poster title
Trustworthy AI

Investigator
K.P. Subbalakshmi (Yupeng Cao)

Abstract
We present two aspects of trustworthy AI in this poster. (1) Information Theoretic look into explainability of attention weights: In this work, we explore how certain attention mechanisms can act as proxies for explanations under specific circumstances. By applying an information-theoretic framework, we demonstrate that attention values can, in some cases, provide meaningful insights into model behavior.

(2) Causal T-GAN: Addressing the challenges of privacy leakage and data scarcity, this research presents Causal-TGAN, a generative model that captures underlying causal relationships in tabular data. Unlike traditional models, Causal-TGAN generates synthetic data that reflects the causal dynamics between variables, leading to more accurate and useful datasets.