Building an Invisible Shield to Enable Traceable Privacy Protection for Medical Images in Telemedicine

As telemedicine continues to expand, the frequent transmission and sharing of medical images over networks has become central to remote diagnostics. However, these images often contain sensitive lesion information. Once they are illicitly obtained during transmission or storage, they can be misused to train malicious segmentation models, resulting in serious patient privacy violations. While encryption and steganography provide basic protection, encrypted content may draw adversarial attention, and some stego-images may be exposed due to abnormal formats or semantics. More critically, most existing modes lack traceability, making it difficult to identify the source once a privacy breach occurs. To address these challenges, we present a traceable robust adversarial watermarking model that acts as an invisible shield to protect medical image privacy in telemedicine scenarios. This model seamlessly integrates invisibility, privacy protection, and traceability into a unified watermark embedding framework, enabling proactive defense against segmentation-based attacks while maintaining diagnostic quality. Specifically, receiver identity information is embedded into medical images through adversarial perturbations. These perturbations suppress lesion extraction by attackers while allowing reliable identity decoding in case of data leakage. Experimental results show that the model achieves strong privacy protection on various polyp and ISIC datasets. Moreover, the model maintains reliable ID extraction under different noise perturbations, validating its robustness and traceability. Visual quality assessments further confirm the invisibility of the embedded watermark, Ensuring that diagnostic usability remains unaffected. This provides a promising direction for safeguarding medical image privacy in telemedicine environments.