A multi-layer active intellectual property protection framework for hyperspectral image classification models

Cloud-hosted hyperspectral image (HSI) classification models are valuable yet vulnerable to theft, tampering, and misuse. Existing deep model watermarking techniques have mainly been validated on natural-image classifiers, and their direct use in HSI settings can be affected by spectral dimensionality reduction, label-consistency artifacts in pixel-wise classification maps, and the delayed nature of passive ownership verification. To address these issues, we propose a multi-layer active spectral-domain framework that integrates band-aware steganography for per-user fingerprinting, a class-hiding backdoor watermark for authorization and misuse tracing, a lightweight Lambda-layer access controller for multi-user identification and revocation, and an encoder-decoder that imprints owner fingerprints into classification outputs for remote verification. The framework is designed to preserve model fidelity while enabling ownership verification and access control. On standard HSI benchmarks (Indian Pines, Pavia University, Salinas) across three spectral-spatial convolutional neural network (CNN) architectures (HybridSN, 3D-CNN, 3D-DL), it achieves watermark detection accuracy above 97.8%, authorized user traceability exceeding 99.7%, and no statistically significant loss in overall accuracy under 10 repeated runs. Robustness tests against pruning, fine-tuning, and knowledge distillation further indicate that the embedded ownership evidence remains detectable after common model-purification attacks. These results support spectral-domain active IP protection as a practical option for HSI classification models deployed in cloud MLaaS environments.