Transmitter Identification with machine learning: Future radio spectrum management techniques will use dynamic information to allocate spectrum resources, rather than using a fixed number of bands. Transmitter identification is critical to accuracy gauge demand as well as enforce existing policies. In this project, you will quantify the efficacy of a variety of machine learning approaches to transmitter identification. You will collect training data from a variety of transmitter types in the orbit grid and compare the accuracy of different types of neural networks ability to identify the type and number of transmitters.

High speed training using binary neural networks: Training machine learning systems is currently very slow. Recent work has shown promise by using simpler representations of numbers than the commonly used floating point ones. In this project, you will create and measure neural networks which use only binary or integer numbers for both training and inference.

Distributed spectrum sensing and channel assignment: use a collection of software defined radios (SDRs) to detect spectrum occupancy and perform channel usage coordination among number of radio systems.

Characterizing 5G Leakage on Passive Weather Sensing Spectrum Bands: Perform RF interference characterization experiments at 23-24 GHz by using mmWave radios and RF measurement equipment

FPGA implementation of object detection: This project aims to study real-time object detection on the FPGA board. FPGA is a programmable chip that inherently supports parallel processing. Building object detection, which is a key component for autonomous driving and drone, on FPGA can provide task-flexible time-efficient energy-efficient solutions for practical object detection.

Smart Intersection Situational Awareness: Use inputs from lidars, 2d and 3d cameras and other sensors to create a 3D point-cloud of the intersection area.

Quantifying the transformation of Information to Energy: Understanding the lower bounds of energy use in communications and computing is not well understood. In this project, you will build a software version of Maxwell’s demon, which converts information into energy. You will measure the demon’s effectiveness by modeling the relationship between the demon and the physical system as a discrete channel with noise, and measuring the energy extracted as a function of the conditional entropy.

Analyzing Social Distancing Based on Sensory Inputs: Use computer vision and other machine learning techniques to identify and measure the effects of social distancing policies on the population in an urban setting

Using Machine Learning for IoT: Develop a framework for building intelligent environment applications through ambient sensing and machine learning

AR Memorial: Develop augmented reality obelisk/memorial that can grow in all dimensions allowing users can contribute paintings, photos, videos or notes and leave them for other visitors

Adversarial Machine Learning on Voice Assistant Systems and IoT Devices: Driven by the advancement of machine learning, voice-user interfaces in voice assistant systems (e.g., Alexa, Google Assistant, Siri) and IoT devices have widely incorporated machine learning models as building blocks for speech recognition and speaker identification. However, the machine learning models are inherently vulnerable to adversarial examples, which are well-designed perturbations leading to wrong predictions while being imperceptible to human beings. This project aims to study the vulnerability of voice-user interfaces under such adversarial attacks by developing audio adversarial examples.

Mobile User Authentication via Deep Learning: User authentication is a critical process in many mobile and IoT applications (e.g., unlocking devices, mobile payment). This project aims to achieve user authentication by applying deep learning techniques to the wireless signals generated by mobile/IoT devices. The insight is that the user’s unique behaviors can lead to distinctive patterns in the wireless signals. By using wireless signals as input, we can develop deep learning techniques to extract unique behavioral biometrics to perform user authentication in real-time.