用於醫學影像的雙重浮水印機制：擁有者識別與篡改定位

Abstract

隨著醫療系統的數位化，醫療設備曝露於網際網路上已成為嚴重的資安風險。再加上生成式模型的快速進展，攻擊者如今能夠進行高度擬真的醫療影像竄改，甚至連放射科醫師也難以察覺，嚴重威脅病人安全與臨床信任。為因應此問題，本研究聚焦於醫療影像真實性的兩大面向：來源識別（ownership identification）與內容保護中的竄改定位（tampering localization）。過往的浮水印技術多採取全圖嵌入雙重浮水印來同時達成這兩項目標，但識別浮水印與定位浮水印之間會互相干擾，導致影像品質明顯下降，尤其影響診斷區域。為解決此問題，本研究利用醫療影像中感興趣區域（ROI）與非感興趣區域（RONI）的結構差異，提出區域可控的浮水印嵌入機制：將具高破壞性的識別浮水印限制嵌入於RONI，並將高品質的脆弱定位浮水印覆蓋整張影像。實驗結果顯示，即使面對先進的深度學習型竄改，本方法仍能有效維持ROI品質，同時達成高準確度的病灶級竄改定位。在理論上，本研究提出的區域感知浮水印嵌入機制提升了醫療浮水印的控制性；在實務上，則可於不影響診斷可用性的前提下，強化遠距醫療與醫院影像系統的資訊完整性，促進如遠距照護與病患影像安全分享等應用的實現。

With the digitization of healthcare systems, the exposure of medical devices on the internet has become a severe cybersecurity risk. Coupled with recent advances in generative models, attackers can now perform highly realistic manipulations of medical images that even radiologists fail to detect, threatening patient safety and clinical trust. This study addresses the dual dimensions of medical image authenticity: ownership identification and tampering localization for content protection. Prior watermarking approaches embed dual watermarks globally to tackle this dual problem. However, the interference between identification and localization watermarks leads to significant image quality degradation, especially in diagnostic regions. To overcome this, we leverage the structural distinction between Regions of Interest (ROI) and Regions of Non-Interest (RONI) in medical images. By proposing a region-controllable mechanism, we successfully control the robust identification watermarks to be embedded only in RONI, while the high-quality fragile localization watermarks cover the full image. Experiments show our method preserves ROI quality while achieving superior lesion-level tamper localization, even under advanced deep learning-based manipulations. Theoretically, our region-aware embedding mechanism improves controllability in medical watermarking. Practically, it enhances the integrity of telemedicine applications and hospital imaging systems without sacrificing diagnostic usability, enabling real-world applications such as remote healthcare and secure patient image sharing.