The digitization of healthcare and the rise of AI-based diagnostics have made the secure handling of medical data a fundamental requirement; not just for compliance, but for preserving patient safety, trust, and the integrity of diagnostic outputs. Conventional encryption methods protect data in transit and at rest, but once data is decrypted for AI model training or visualization, its exposure risk sharply increases.

A promising solution to this challenge is deep learning-based zero watermarking, a method that embeds authentication and traceability into the feature space of medical data without altering the original image or record. This approach enables robust, reversible, and imperceptible security for clinical images and electronic health records (EHRs), even under adversarial conditions.

Technical Approach

This zero watermarking framework introduces a hybrid design that combines:

• AlexNet: A convolutional neural network for feature extraction from visibly marked carrier images.
• NSST (Non-subsampled Shearlet Transform): For multi-resolution and multi-directional decomposition.
• SVD (Singular Value Decomposition): To isolate high-energy components for embedding.
• SSFC (Step Space-Filling Curve): A scrambling method to enhance watermark confidentiality and resistance to reverse engineering.

Workflow:

1. A visible watermark (e.g., institutional logo) is embedded in the image to signify origin.
2. A scrambled version of the secure watermark (e.g., patient ID, EHR hash) is generated using SSFC.
3. Features from the visibly marked image are extracted using AlexNet.
4. The scrambled watermark is embedded into selected NSST-SVD coefficients, creating a zero watermark that can be recovered without modifying the image.

Why Zero Watermarking?

Unlike conventional watermarking that alters image pixels or frequency components, zero watermarking retains the original image untouched. The watermark is embedded in derivative feature spaces, which preserves image quality while enabling traceability and authentication. This is particularly valuable for medical imaging, where even minor perturbations can affect diagnostic accuracy or invalidate AI model inputs.

Use Cases

• Clinical data traceability: Establish content origin and prevent tampering during image exchange.
• AI pipeline integrity: Embed training data lineage into image features used in model development.
• Federated learning audits: Verify the provenance of shared datasets without compromising privacy.
• Post-decryption assurance: Ensure EHR or WSI authenticity even after encryption is removed for processing.

Outlook

Deep learning-based zero watermarking represents a major step forward in medical data security. It enables secure, reversible, and verifiable sharing of health records and imaging data without sacrificing clinical quality or machine learning utility. As healthcare infrastructure grows more interconnected and AI-driven, embedding security directly into the data’s digital signature (not just its containers) will become a foundational design principle.

References

Anand A, Bedi J, Aggarwal A, Khan MA, Rida I. Authenticating and securing healthcare records: A deep learning-based zero watermarking approach. Image Vis Comput. 2024;145:104975. https://doi.org/10.1016/j.imavis.2024.104975

Mousavi SM, Naghsh A, Abu-Bakar SAR. Watermarking techniques used in medical images: A survey. J Digit Imaging. 2014;27(6):714–729. https://doi.org/10.1007/s10278-014-9700-5

Kaur G, Gupta B, Anand A. Deep learning based zero watermarking for authentication of medical records. Comput Sci Inf Technol (CS & IT). 2023;132:129–41. doi:10.5121/csit.2023.132411.