Subspace-Guided Unlearning and Recovery: Poisoning Defense for Hierarchical Federated Learning in LEO Satellite Networks

Abstract

Low Earth Orbit (LEO) satellite networks are increasingly adopting Federated Learning (FL) for privacy-preserving collaborative model training. However, this paradigm remains vulnerable to backdoor attacks where compromised satellites inject hidden triggers into their local model updates. Existing defenses are not suited for the LEO environment: they require additional clean validation data on the server with huge transmission overhead, involve lengthy iterative procedures incompatible with limited energy budgets, or exhibit degraded performance in distributed settings. We propose a three-stage defense framework that integrates seamlessly into the FL pipeline without these requirements. In the detection stage, Cluster Heads identify suspicious participants via DBSCAN-based anomaly detection, and the suspicious gradients collected in this process are reused to construct a malicious gradient subspace via Singular Value Decomposition (SVD). In the unlearning stage, this subspace enables targeted backdoor removal through gradient reversal that provably increases backdoor loss, while preserving benign gradient components. In the recovery stage, neuron-level learnable masks are trained to suppress backdoor-related activations while model parameters remain frozen to prevent backdoor relearning. We provide theoretical analysis showing that gradient reversal provably increases backdoor loss and that subspace filtering bounds distortion to benign gradients. Extensive evaluations on MNIST, Fashion-MNIST, CIFAR-10, and SAT-6 datasets demonstrate that our method achieves a defense success rate exceeding 96% while maintaining competitive clean accuracy (95.9% on MNIST, 85.4% on Fashion-MNIST, 77.5% on CIFAR-10, 93.3% on SAT-6), with <inline-formula><tex-math notation="LaTeX">$2-4\times$</tex-math><alternatives><mml:math><mml:mrow><mml:mn>2</mml:mn><mml:mo>-</mml:mo><mml:mn>4</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="shen-ieq1-3716206.gif"/></alternatives></inline-formula> faster recovery time compared to existing defense methods.