SecureLEO: Secure Scheduling for LEO Satellite Networks

Official implementation of "Deep Learning-Based Secure Scheduling and Cooperative Artificial Noise Generation in LEO Satellite Networks"

Authors: Yongjae Lee (Student Member, IEEE), Taehoon Kim (Member, IEEE), Inkyu Bang (Member, IEEE), Erdal Panayirci (Life Fellow, IEEE), and H. Vincent Poor (Life Fellow, IEEE)

Submitted to: IEEE Wireless Communications Letters (under review)

---

Overview

SecureLEO is a deep learning-based physical-layer security (PLS) framework for LEO satellite communication networks. It jointly optimizes satellite scheduling and cooperative artificial noise (AN) generation using Set Transformer architecture, operating with only statistical eavesdropper CSI.

Key Features

• Set Transformer-Based Scheduling: Permutation-invariant architecture that captures inter-satellite interactions for optimal subset selection
• Cooperative AN Generation: Non-data satellites transmit AN in the null space of legitimate channels to degrade eavesdropper reception
• Statistical CSI Only: No instantaneous eavesdropper CSI required — uses ergodic secrecy rate maximization via Monte Carlo sampling
• Multi-Architecture Support:
  • Set Transformer (default): Attention-based, permutation-invariant
  • Deep Sets: Lightweight baseline

Performance Highlights

• 93–96% of genie-aided secrecy rate (Set Transformer, 1M-trial GPU simulation)
• Deep Sets baseline: 64–68% of genie-aided — attention mechanism is essential
• 2.6–3.7x improvement over random scheduling baseline
• Sub-millisecond inference vs. seconds for brute-force search
• Robust across varying system parameters ($N$, $K_d$, $M_e$, $K_{\text{AN}}$)

---

Project Structure

SecureLEO/
├── main.py                 # CLI entry point (train / evaluate / version)
├── pyproject.toml
├── requirements.txt
├── secureleo/               # Main package (flat, ~1500 lines)
│   ├── __init__.py         # Version & public API
│   ├── config.py           # YAML-based experiment configuration
│   ├── channel.py          # Shadowed-Rician fading channel generation
│   ├── signal.py           # Precoder + AN + SINR (ZF/MMSE) + Secrecy Rate
│   ├── scheduling.py       # Brute-force optimal + Random baseline
│   ├── models.py           # Set Transformer + Deep Sets
│   ├── dataset.py          # Environment + Dataset + Data generation
│   ├── trainer.py          # Training loop + Loss functions
│   └── benchmark.py        # Multi-baseline benchmarking + Plotting
├── experiments/
│   └── default.yaml        # Default experiment configuration
├── figures/                # Paper figures
└── supplementary/          # Additional experimental results & technical details

---

Installation

Requirements

# Clone the repository
git clone https://github.com/Yongjae-ICIS/SecureLEO.git
cd SecureLEO

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/macOS

# Install dependencies
pip install -r requirements.txt

# Or install as package
pip install -e .

Dependencies

• Python >= 3.9
• PyTorch >= 1.12.0 (CUDA / MPS / CPU)
• NumPy, SciPy, Matplotlib
• typer, tqdm, PyYAML

---

Usage

Quick Start

# Train with default parameters
python main.py train --num-samples 40000 --num-epochs 12

# Train with YAML config
python main.py train --config experiments/default.yaml

# Evaluate a trained model
python main.py evaluate checkpoints/model.pt --num-trials 1000

# Check version & device
python main.py version

Training

python main.py train \
    --num-samples 100000 \
    --num-epochs 12 \
    --batch-size 256 \
    --arch set_transformer \
    --embed-dim 128 \
    --num-heads 4 \
    --num-layers 2 \
    --lr 1e-3 \
    --mc-samples 100 \
    --device auto

Command Line Arguments

Argument	Description	Default
--config	Path to YAML config file	None
--num-samples	Training dataset size	40000
--num-epochs	Number of training epochs	12
--batch-size	Mini-batch size	256
--arch	Model: set_transformer or deep_sets	set_transformer
--embed-dim	Embedding dimension	128
--num-heads	Attention heads	4
--num-layers	Number of SAB layers	2
--lr	Learning rate	1e-3
--mc-samples	MC samples for ergodic rate	100
--num-data-sats	Data satellites ($K_d$)	2
--device	Device: auto, cpu, cuda, mps	auto

System Parameters

Parameter	Symbol	Default	Description
Visible satellites	$N$	15	LEO satellites in view
Scheduled satellites	$K$	10	Selected for transmission
Data satellites	$K_d$	2	Transmitting data
AN satellites	$K_{\text{AN}}$	$K - K_d$	Transmitting artificial noise
GBS antennas	$M_b$	2	Zero-forcing receiver
Eve antennas	$M_e$	2	MMSE receiver
Rician K-factor	$K_{\text{SR}}$	3	Shadowed-Rician parameter
Nakagami m	$m$	5	Shadowed-Rician parameter

---

Results

Baseline Comparison (1M-Trial GPU Simulation, SNR = 10 dB)

Config	$N$	$K_d$	$M_e$	Proposed/Genie-Aided (%)	DeepSets/Genie-Aided (%)	Random/Genie-Aided (%)
N12_Kd2_Me2	12	2	2	93.6	64.2	40.8
N15_Kd2_Me2	15	2	2	93.5	67.7	40.8
N18_Kd2_Me2	18	2	2	93.5	65.5	40.8
N15_Kd2_Me4	15	2	4	96.2	66.6	43.3

Secrecy sum-rate comparison across N = {12, 15, 18}

Secrecy outage probability across N = {12, 15, 18}

Secrecy Rate Performance (Model / Genie-Aided)

Configuration	$N$	$K_d$	$M_e$	Model/Genie-Aided (%)
N12_Kd2_Me2	12	2	2	81.8
N12_Kd2_Me4	12	2	4	75.5
N15_Kd2_Me2	15	2	2	82.0
N15_Kd2_Me4	15	2	4	75.3
N15_Kd3_Me2	15	3	2	84.4
N18_Kd2_Me2	18	2	2	82.3
N18_Kd3_Me2	18	3	2	84.2

CSI Gap Analysis

SNR (dB)	Genie-Aided	Statistical	Model	Random	Statistical/Genie-Aided (%)
0	0.63	0.44	0.10	0.01	69.8
10	5.07	4.80	4.00	1.52	94.7
14	7.53	7.27	6.46	3.32	96.5
20	11.41	11.16	10.34	6.57	97.8

Set Transformer Ablation Study

Parameter	Value	Model/Genie-Aided (%)
$d=64$	$h=2, d=64, L=2$	81.7
$d=128$ (default)	$h=4, d=128, L=2$	82.1
$d=256$	$h=4, d=256, L=2$	82.1
$L=1$	$h=4, d=128, L=1$	81.0
$L=2$ (default)	$h=4, d=128, L=2$	81.9
$L=3$	$h=4, d=128, L=3$	82.2

---

Supplementary Materials

Due to the page limitations of IEEE WCL, additional experimental results and technical details are provided in the supplementary/ directory:

Document	Description
Parameter Sensitivity Analysis	System parameter robustness across 12 configurations ($N$, $K_d$, $M_e$)
Performance Gap Analysis	SNR-wise gap decomposition: CSI gap vs. algorithm gap
Hyperparameter Ablation Study	Set Transformer architecture ablation ($d$, $h$, $L$)
AN Satellite Configuration	Effect of AN satellite count ($K_{\text{AN}}$) on secrecy rate
Convergence Analysis	Training convergence over 50 epochs with 5 random seeds
AN Generation Protocol	Null-space computation and distributed AN protocol details

---

Citation

Citation information will be added upon publication.

---

License

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

---

Acknowledgements

• Set Transformer: Lee et al., ICML 2019 — Attention-based permutation-invariant architecture
• Deep Sets: Zaheer et al., NeurIPS 2017 — Permutation-invariant baseline (without attention)
• Physical-Layer Security: Wyner's wiretap channel model and extensions

---

Contact

For questions or issues, please open an issue or contact:

• Yongjae Lee: yjlee@edu.hanbat.ac.kr
• Inkyu Bang: ikbang@hanbat.ac.kr