Star-ov: satellite-oriented trusted authorization and responsibility with stratified model encryption and ownership verifications

Abstract With the increasing adoption of deep neural networks (DNNs) in remote sensing interpretation and on-orbit intelligent processing, protecting the intellectual property of deployed models faces new challenges, including limited onboard computational resources, constrained satellite–ground communication bandwidth, differentiated multi-user services, and difficulty in ownership attribution after model leakage. Existing model protection approaches mainly rely on either active encryption or passive watermarking mechanisms, which are insufficient to simultaneously satisfy the requirements of access control, service continuity, and post-leakage accountability in satellite-based scenarios. In this paper, we propose STAR-OV, a stratified and controllable model protection framework tailored for satellite deployment. By introducing a two-stage reversible parameter perturbation mechanism, STAR-OV enables progressive degradation of model performance from the original high-accuracy state to an intermediate-accuracy state and finally to a fully unusable state. The perturbations introduced at different stages are organized into distributable hierarchical keys, thereby supporting differentiated service access for users with varying authorization levels. Furthermore, after the model is fully restored to its original accuracy, a latent watermark is embedded. The watermark remains inactive during normal inference without the activation key, while producing stable and verifiable responses on a trigger set once the corresponding key is applied, enabling reliable ownership verification in the event of model leakage. Experimental results on multiple benchmark datasets demonstrate that the proposed method significantly reduces the usability of unauthorized models by modifying an extremely small fraction of parameters, while maintaining strong robustness against fine-tuning, pruning, and weight noise attacks. Owing to its low computational overhead and flexible deployment capability, STAR-OV provides a practical solution for resource-constrained satellite-based onboard intelligent inference scenarios.