A learned perceptual and neural detection framework for spread spectrum watermarking in medical images

Medical image watermarking must ensure data authenticity and integrity without compromising diagnostic quality. Conventional spread spectrum methods suffer from key dependency, geometric vulnerability, hand-crafted perceptual models, and generative AI susceptibility. This paper presents a hybrid DWT-based spread spectrum framework for medical imaging with three innovations: a learned perceptual masking network (JND_net) that replaces analytical JND models, a lightweight CNN (SubBandSelector) that dynamically selects resilient wavelet sub-bands, and a neural detector (DetectorNet) that supplants fixed-threshold detection, all integrated with dual-key security (2 128 key space) and MAC authentication. Experiments on 1,200 medical images with radiologist validation (51.3% detection accuracy, chance level; diagnostic confidence unchanged, p = 0.34) show bit error rates below 0.03 under JPEG, noise, and filtering, outperforming seven baselines (p < 0.001). Under generative AI attacks, BER reaches 0.112 (diffusion) and 0.087 (inpainting). The framework operates in real time (3.7 ms on T4 GPU, 25 ms on CPU) under blind extraction.