What question did this study set out to answer?

June 30, 20260 citationsOpen Access

Ensemble Federated Learning for Attack Detection in FSO Satellite Networks

INIqbal Narub AMAli Arshad Muhammad UMUmair Muhammad

Key Points

This research aims to develop a machine learning framework for real-time detection and classification of attacks in FSO satellite communication systems.
Developed a hybrid multi-model machine learning framework integrating Random Forest, 1D-CNN, and LSTM networks.
Generated a dataset of 15,000 samples using realistic atmospheric turbulence models and multiple jamming attack types.
Implemented a Federated Learning architecture for distributed attack detection across satellite nodes.
Random Forest achieved 100% accuracy on the synthetic test set.
LSTM achieved 99.90% accuracy, and 1D-CNN achieved 99.83% accuracy on the test dataset.
Federated Learning converged to 98.97% accuracy after 15 communication rounds, maintaining data privacy across satellite nodes.

Abstract

Free Space Optical (FSO) communication has emerged as a fundamental technology for next-generation satellite internet constellations, offering ultra-high bandwidth, low latency, and inherent directional security for inter-satellite and satellite-to-ground links. However, FSO systems remain vulnerable to deliberate jamming attacks and covert interception attempts that can severely compromise communication integrity and data confidentiality. This paper presents a comprehensive machine learning framework for real-time detection and classification of security threats in FSO satellite communication systems. We propose a hybrid multi-model approach integrating Random Forest, one- dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM) network, and a Federated Learning architecture specifically designed for distributed satellite constellations. A physically-grounded dataset of 15,000 samples is generated incorporating realistic atmospheric turbulence models based on Rytov theory, four distinct jamming attack modalities (laser jamming, broadband interference, sweep jamming, and pulsed jamming), and interception attacks modeled through beam splitting and energy tapping. Eight discriminative signal features including signal-to-noise ratio (SNR), bit error rate (BER), scintillation index, and Rytov variance are extracted for classification. Experimental results demonstrate exceptional detection performance: Random Forest achieves 100% accuracy on the synthetic test set, though this upper bound may reflect dataset regularity and requires validation on real FSO captures, 1D- CNN achieves 99.83%, LSTM achieves 99.90%, and the Federated Learning model converges to 98.97% accuracy after 15 communication rounds, only 1.03% below the centralized RF baseline while preserving data privacy across satellite nodes. Feature importance analysis reveals SNR and scintillation index as the most discriminative features, and comprehensive confusion matrix analysis confirms robust per- class performance. These results establish the viability of ML- based security monitoring for FSO satellite systems and provide the first systematic evaluation of federated learning for distributed attack detection in satellite constellations.

问 AI

Bookmark

View Full Paper