SpecEES: DISCOVER: Device Identification for Spectrum-optimization using COnVolutional nEural netwoRks

Northeastern University, Boston MA

Investigators

Abstract

The research objective of this project, called DISCOVER, is to harness the power of deep machine learning (ML) algorithms to communicate wirelessly with high spectral efficiency and low power consumption. The techniques will result in advanced networking protocols that consume minimal resources by securely identifying the devices that are active in the surrounding environment without (or with minimal) control signaling. Apart from learning channel usage and device activity, DISCOVER will allow for rapidly deploying these algorithms in hardware, so that real-time inferences can be made. Thus, DISCOVER is directly aligned with the US President's executive order from February 2019 'Maintaining American Leadership in Artificial Intelligence' that seeks to prioritize research and development of America's artificial intelligence (AI) capabilities. DISCOVER aims to bring together industry, academia and government stakeholders through collaborative workshops towards identifying high priority challenges, limitations of available data sources, and identify a list of candidate machine learning solutions that will shape the next generation of wireless technologies. The open source release of signal datasets and simulation code will foster new interactions of wireless researchers with core machine learning domain experts. DISCOVER has three goals for optimizing spectrum utilization with overlapping interests of either energy saving or resilience to identity spoofing through the use of deep learning architectures: 1. It aims to explore deep convolutional neural network (CNN) architectures that will allow highly accurate device classification and demonstrate how to eliminate identifier-related protocol fields. This approach of reducing packet headers will achieve quantifiable spectrum utilization improvements, especially for large-scale deployment of the Internet of Things (IoT). 2. It aims to demonstrate the first learning-in-the-loop radio frequency (RF) system where spectrum-driven decisions are enabled through real-time deep learning algorithms implemented directly on the device hardware. This will result in significant energy savings for embedded IoT devices. 3. The emulation engine developed in the project will empower users to create custom-signals to train ML algorithms. Furthermore, it will create community RF signal datasets that will ensure means of standardized validation for the larger research community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.