ePASS: Enhanced Physical-layer Authentication with Super-resolution for Satellites

Official implementation of "Lightweight Physical-Layer Authentication via IQ Sample Super-Resolution in LEO Satellite Networks"

Authors: Ivy Selorm Dogbey†, Yongjae Lee†, Jihwan Moon, Taehoon Kim, and Inkyu Bang

† These authors contributed equally to this work.

Published in: IEEE Communications Letters (Early Access, Feb. 2026)

DOI: 10.1109/LCOMM.2026.3659912

---

Overview

ePASS is a practical and lightweight physical-layer authentication (PLA) framework for LEO satellite communication networks. It leverages Enhanced Deep Super-Resolution (EDSR) techniques to enhance the resolution of IQ sample-based images, enabling accurate satellite authentication with a minimal number of received packets.

Key Features

• IQ Sample to Image Conversion: Transform raw IQ data into 2D histogram images
  • Cartesian coordinates (224x224) for standard processing
  • Polar coordinates (256x256) for AMC-style comparison
• Super-Resolution Enhancement: EDSR x4 upscaling (224x224 → 896x896)
• Multi-Architecture Support:
  • CNN: Modified ResNet-18 for 66-class satellite classification
  • RNN: LSTM and GRU for sequential feature extraction
• Spoofing Attack Detection: Robust authentication using sparse autoencoder

Performance Highlights

• 90.3% authentication accuracy with only 500 IQ samples (~5 packets)
• 94.77% authentication accuracy with 1,000 IQ samples
• Significantly outperforms Raw ResNet baseline and AMC methods
• Suitable for real-time operation on edge AI hardware (e.g., NVIDIA Jetson Orin Nano)

---

Project Structure

ePASS/
├── src/
│   ├── gen_img_raw.py         # Generate Cartesian IQ images (224x224)
│   ├── gen_img_amc.py         # Generate Polar (AMC-style) images (256x256)
│   ├── train_cnn.py           # CNN training (ResNet-18) with PyTorch
│   ├── train_rnn.py           # RNN training (LSTM/GRU)
│   └── upscale_images.py      # EDSR/Bicubic x4 upscaling (CUDA accelerated)
├── models/
│   └── EDSR_x4.pb             # Pre-trained EDSR model (x4 scale)
├── figures/                   # Paper figures
├── results/                   # Experiment results
└── data/sample/               # Sample data

---

Installation

Requirements

pip install -r requirements.txt

Dependencies

Python Environment (3.12+)

• PyTorch >= 2.0.0
• torchvision >= 0.15.0
• opencv-python >= 4.5.0 (with CUDA support for EDSR acceleration)
• opencv-contrib-python >= 4.5.0
• numpy, pandas, matplotlib
• scikit-learn, tqdm, Pillow, openpyxl

MATLAB Environment (R2025a)

• Deep Learning Toolbox
• Image Processing Toolbox

Hardware (Recommended)

• NVIDIA GPU with CUDA support
• For deployment: NVIDIA Jetson Orin Nano or equivalent edge AI hardware

---

Usage

1. Generate Images from IQ Data

# Cartesian coordinates (224x224) - Standard ePASS input
python src/gen_img_raw.py

# Polar coordinates (256x256) - AMC-style comparison
python src/gen_img_amc.py

Configuration (in scripts):

• Sample sizes: 500, 1000, 2000, 5000
• Normalization: Per-satellite min-max scaling to [-1, 1]

2. Apply Super-Resolution

python src/upscale_images.py

Output:

• images_896_edsr/: EDSR x4 upscaled images (896x896)
• images_896_bicubic/: Bicubic x4 upscaled images (896x896)

Note: CUDA acceleration is enabled by default. Requires OpenCV built with CUDA support.

3. Train Classification Models

# CNN (ResNet-18) - Multiple modes and sample sizes
python src/train_cnn.py --mode edsr bicubic --samples 500 1000 5000 --epochs 10

# RNN (LSTM/GRU)
python src/train_rnn.py --model lstm gru --mode edsr --samples 5000 --epochs 10

Command Line Arguments

Argument	Description	Default
--mode	Data mode: base, amc, edsr, bicubic	edsr, bicubic
--samples	Sample sizes per image	500, 1000
--epochs	Training epochs	1
--limit	Images per class	200
--batch_size	Batch size	32
--split_strategy	Data split: random, sequential	random

Training Parameters (ResNet-18)

Parameter	Value
Optimizer	SGD with Momentum (0.9)
Initial Learning Rate	1e-3
Learning Rate Drop Factor	0.1 (every 5 epochs)
Mini-batch Size	32
Max Epochs	10
Normalization	ImageNet mean/std

---

Results

Intra-Satellite Authentication (Hit Rate)

Method	0.5K	1K	2K	5K	10K
Raw + ResNet-18	73.83%	82.27%	85.30%	96.74%	97.31%
Bicubic + ResNet-18	86.82%	96.74%	-	-	-
AMC + ResNet-18	46.44%	60.11%	67.12%	91.33%	94.73%
EDSR + LSTM	1.52%	1.52%	-	-	-
EDSR + GRU	1.52%	1.52%	-	-	-
ePASS (EDSR + ResNet-18)	90.30%	94.77%	-	-	-

Spoofing Detection Performance

Sample Size	Scenario	Authentication	Spoofing Detection	Accuracy	F1 Score
1K	Best (weak attacker)	0.312	0.912	0.615	0.415
1K	Worst (strong attacker)	0.347	0.972	0.685	0.495
5K	Best	0.957	0.706	0.761	0.825
5K	Worst	0.987	0.756	0.861	0.866
10K	Best	0.938	0.937	0.887	0.891
10K	Worst	0.978	1.000	0.989	0.993

Computational Complexity

Metric	ePASS	PAST-AI (Baseline)
Computational Cost	~8.21 TFLOPs	~1.80 GFLOPs
Input Data Size	500 IQ samples (~5 packets)	10,000 IQ samples (~87 packets)
Data Acquisition Latency	Low (instantaneous)	High (depends on inter-packet delay)
Feasible Hardware	NVIDIA Jetson Orin Nano	Generic IoT CPU/MCU

---

Dataset

This work uses the publicly available Iridium satellite IQ sample dataset:

G. Oligeri, S. Sciancalepore, and R. Di Pietro, "Physical-layer data of IRIDIUM satellites broadcast messages," Data in Brief, vol. 46, p. 108905, 2023.

Dataset Characteristics:

• 66 Iridium LEO satellites
• Over 102 million IQ samples total
• ~1.5 million samples per satellite
• IRA (Iridium Ring Alert) channel messages

---

Citation

@article{dogbey2026epass,
  author    = {Dogbey, Ivy Selorm and Lee, Yongjae and Moon, Jihwan and Kim, Taehoon and Bang, Inkyu},
  title     = {Lightweight Physical-Layer Authentication via IQ Sample Super-Resolution in LEO Satellite Networks},
  journal   = {IEEE Communications Letters},
  year      = {2026},
  doi       = {10.1109/LCOMM.2026.3659912},
  note      = {Early Access}
}

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Acknowledgements

• EDSR: EDSR-PyTorch - Super-resolution model architecture
• ResNet: Deep Residual Learning for Image Recognition - Classification backbone
• Iridium Dataset: PAST-AI - Public satellite IQ sample dataset
• AMC Framework: T. K. Oikonomou et al., "CNN-Based Automatic Modulation Classification Under Phase Imperfections," IEEE WCL, 2024 - Polar coordinate transformation baseline

---

Contact

For questions or collaboration inquiries, please open an issue or contact the corresponding author listed in the paper.