Loading...
Super-resolution preprocessing for autoencoder-based physical-layer authentication of LEO satellites
Macias, Ivan
Macias, Ivan
Citations
Altmetric:
Authors
Other Names
Location
Time Period
Advisors
Original Date
Digitization Date
Issue Date
2026-05
Type
Thesis
Genre
Keywords
Subjects (LCSH)
Electronic dissertations
Citation
Abstract
The continued proliferation of Low Earth Orbit (LEO) satellite constellations has heightened the need for robust authentication techniques that can defend against spoofing and unauthorized access in civilian and military applications. Following defense-in-depth principles, physicallayer authentication (PLA) can serve as an additional security layer that operates outside of higherlayer cryptographic protocols, providing protection even when upper-layer defenses are compromised. PLA via radio frequency fingerprinting exploits manufacturing imperfections in transmitter hardware that produce signal characteristics unique to each device. Prior work has shown that converting raw IQ samples into two-dimensional histogram images enables both multi-class classification via convolutional neural networks and one-class authentication via autoencoders trained on a reference transmitter. Upscaling these histograms with the Enhanced Deep Super-Resolution (EDSR) network preserves hardware fingerprint detail while reducing the per-decision IQ-sample budget, but only in the multi-class classification setting. This thesis extends that line of work by evaluating super-resolution (SR) preprocessing in the autoencoder-based authentication setting, comparing EDSR, FSRCNN, ESPCN, and LapSRN against a raw-histogram baseline on publicly available Iridium IQ samples. SR preprocessing yields a 3-5 AUC-point gain over the baseline in the sample-scarce regime (N = 500 IQ samples), with the lightweight LapSRN achieving the highest mean AUC of 0.9418, above EDSR’s 0.9324 at roughly an order of magnitude lower compute. At N = 5000 all conditions saturate near AUC 0.994 and are indistinguishable. These findings suggest that if SR preprocessing is used to reduce the IQ-sample collection budget, LapSRN is the preferred network: it matches or exceeds EDSR’s authentication accuracy at a fraction of the inference cost.
Table of Contents
Description
Thesis (M.S.)-- Wichita State University, College of Engineering, School of Computing
Publisher
Wichita State University
Journal
Book Title
Series
Digital Collection
Finding Aid URL
Use and Reproduction
© Copyright 2026 by Ivan Macias
All Rights Reserved