Federated Learning for Anomaly Detection in Satellite Constellations Using an Attention-Based LSTM-Autoencoder

Abstract

As satellites become more complex and satellite constellations grow, there is a need for privacy-preserving and autonomous anomaly detection systems that can function under harsh communication and computational limitations. In this paper, federated learning is proposed for anomaly detection in satellite telemetry data, using an attention based LSTM autoencoder. To overcome these key privacy and bandwidth constraints in space-based systems, our system allows multiple satellites to jointly train a shared anomaly detection model without sharing any raw telemetry data. We propose an architecture that integrates temporal attention with LSTM autoencoders, allowing for the capture of temporal events over long horizons and subtle anomalous event patterns in multivariate telemetry streams. Our framework is evaluated using a synthetic satellite telemetry dataset, under centralized, local only and federated learning approaches in different configurations. Our primary contribution is the architectural integration of federated learning with attention-based LSTM autoencoders, rather than achieving state-of-the-art anomaly discrimination. Our federated attention-LSTM autoencoder can obtain a validation loss of 0.7221 and anomaly detection rate of 5.00% with total communication overhead of 1.01 MB in five satellite clients. The attention mechanism demonstrates greater stability in reconstruction error with a reduction of 95.3% in the standard deviation compared to standard LSTM autoencoders. The AUC score of 0.5105 indicates that the current model performs near-random in terms of anomaly discrimination; we therefore focus on the framework's architectural and communication advantages. Our results confirm that FL is a viable, privacy-respecting, communication-efficient paradigm for onboard satellite anomaly detection in a resource-constrained constellation environment, while honestly acknowledging performance limitations and positioning this work as a foundational architectural framework.