A Systematic Review of Federated Deep Learning Models for Intrusion Detection in Distributed Satellite Data Centres.

Abstract:

This systematic review evaluates existing federated and deep learning intrusion detection approaches with a focus on their suitability for distributed satellite and critical infrastructure environments. Following PRISMA procedures, 13,213 records were identified, 11,265 duplicates removed, 1,948 screened, and 10 studies met full eligibility criteria. Across these studies, federated learning models achieved competitive detection outcomes, with reported accuracy ranging from eighty-eight percent to ninety-seven percent and F1 scores between eighty-five percent and ninety-six percent, often differing from centralised models by less than three percent. Communication efficiency improved substantially, with several studies demonstrating reductions of thirty to sixty percent in update bandwidth due to parameter rather than data transmission. Privacy preservation scored consistently high across all federated implementations, while centralised systems showed significant exposure risk. Weaknesses emerged in handling non-independent data, where performance dropped by up to ten percent in some studies, and in susceptibility to gradient poisoning, which degraded accuracy by seven percent in controlled experiments. Mathematical formulations remained underdeveloped, with limited convergence proofs and insufficient modelling of secure aggregation. The findings indicate that federated deep learning is methodologically superior for satellite data centres but requires realistic satellite traffic datasets, hybrid optimisation such as blockchain-assisted aggregation, and improved mathematical modelling to ensure robustness.