Event
Ph.D. Research Proposal Exam: Haoying Dai
Tuesday, June 22, 2021
10:00 a.m.
https://umd.zoom.us/j/7136415498
Emily Irwin
301 405 0680
eirwin@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Haoying Dai
Committee:
Professor NAME (Chair) Yanne K.Chembo
Professor NAME Thomas E.Murphy
Professor NAME Kevin Daniels
Date/time: Jun. 22 10am-12am
Location: https://umd.zoom.us/j/7136415498
Title:RF Fingerprinting with Narrowband Optoelectronic Oscillator based Reservoir Computing
Abstract: RF fingerprinting refers to a type of method that recognizes transmitters with their hardware-level characteristics like fabrication imperfection. This technology often serves as a physical-layer protection measure for communication networks. Currently, neuromorphic computing techniques like convolutional neural networks (CNNs) have been used to tackle this problem. However, the signals that transmit through the channels are I-Q modulated signals that operating as tens to hundreds of Gigabits while the most conventional machine learning algorithms are running on GPU or TPU, a digital circuit that operates at only several Gigabits. Thus, the I-Q modulated signals have to be demodulated and processed into compatible formats before applying to these platforms, which inevitably slows down the processing speed. This could be fatal to application scenarios like RF fingerprinting where real-time processing is needed.
Comparing to deep learning technologies, reservoir computer has a relatively simple yet powerful structure and reach state-of-the-art in several benchmarks. It’s known that wideband OEO is able to perform reservoir computing (RC) for the classification of signals. However, this hardware architecture is operating at baseband, making it not suitable for processing modulated RF signals for tasks like RF fingerprinting. In this proposal, we demonstrated a reservoir computer based on narrowband OEO that could be utilized to directly classify I-Q modulated signals without the need for demodulation. We show that with much less training data comparing to CNNs (less than 1%), it could still achieve about 90% classification accuracy for recognizing 10 different transmitters.
Name: Haoying Dai
Committee:
Professor NAME (Chair) Yanne K.Chembo
Professor NAME Thomas E.Murphy
Professor NAME Kevin Daniels
Date/time: Jun. 22 10am-12am
Location: https://umd.zoom.us/j/7136415498
Title:RF Fingerprinting with Narrowband Optoelectronic Oscillator based Reservoir Computing
Abstract: RF fingerprinting refers to a type of method that recognizes transmitters with their hardware-level characteristics like fabrication imperfection. This technology often serves as a physical-layer protection measure for communication networks. Currently, neuromorphic computing techniques like convolutional neural networks (CNNs) have been used to tackle this problem. However, the signals that transmit through the channels are I-Q modulated signals that operating as tens to hundreds of Gigabits while the most conventional machine learning algorithms are running on GPU or TPU, a digital circuit that operates at only several Gigabits. Thus, the I-Q modulated signals have to be demodulated and processed into compatible formats before applying to these platforms, which inevitably slows down the processing speed. This could be fatal to application scenarios like RF fingerprinting where real-time processing is needed.
Comparing to deep learning technologies, reservoir computer has a relatively simple yet powerful structure and reach state-of-the-art in several benchmarks. It’s known that wideband OEO is able to perform reservoir computing (RC) for the classification of signals. However, this hardware architecture is operating at baseband, making it not suitable for processing modulated RF signals for tasks like RF fingerprinting. In this proposal, we demonstrated a reservoir computer based on narrowband OEO that could be utilized to directly classify I-Q modulated signals without the need for demodulation. We show that with much less training data comparing to CNNs (less than 1%), it could still achieve about 90% classification accuracy for recognizing 10 different transmitters.
