Privacy-Preserving Federated Remote Sensing via Spatially Modulated Feature Perturbation and Ground Alignment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Space-to-ground collaborative remote sensing has become a promising mode for global-scale earth observation. However, the open satellite-ground communication link faces severe security threats, especially when unauthorized eavesdroppers deploy gradient or feature inversion attacks based on deep learning to reconstruct sensitive geographic targets. Traditional differential privacy methods apply uniform noise across the entire image space, which inevitably damages the geometric structure of the non-sensitive areas and the performance of downstream tasks. This paper proposes Fed-SMDP, an end-to-end Federated Spatially Modulated Differential Privacy framework for secure satellite-ground collaborative sensing. Specifically, an on-board self-supervised masked autoencoder is deployed to extract robust latent feature representations. To maintain geographical sovereignty, a lightweight spatial modulation mechanism is designed to identify sensitive regions dynamically, where non-uniform noise is injected directly into the corresponding neural activations. Therefore, the transmitted feature representations become visually ambiguous to potential eavesdroppers while preserving essential semantics for downstream applications. At the ground station, a weak-supervised alignment protocol is established to fine-tune task-specific heads using historical labeled datasets, periodically feedback the lightweight model parameter increments to the satellites, and complete the closed-loop federated evolution. The theoretical analysis establishes a rate-distortion-based lower bound on the expected reconstruction distortion of sensitive regions under the considered threat model. Extensive experiments on public datasets demonstrate that Fed-SMDP substantially degrades reconstruction quality under the evaluated generative inversion attacks while maintaining negligible accuracy degradation for global land-cover monitoring tasks.