Invisible medical image watermarking: a biometric–deep learning framework for security and verification

Abstract

The integrity and authenticity of digital medical images are paramount for diagnostic accuracy and patient safety. Existing security measures often fail to create a persistent link between the identity of a patient and their medical scans, leaving decrypted data vulnerable. This paper introduces a novel, integrated security frame- work that forges robust and persistent cryptographic link by embedding a patient’s encrypted fingerprint data directly into their medical images. Our end-to-end system synergistically integrates robust biometric processing with a deep learning-based watermarking model. First, a specialized pipeline preprocesses a patient’s fingerprint using binarization, Zhang-Suen thinning, and minutiae extraction. These biometric features are then secured using AES-GCM (Advanced Encryption Standard with Galois/Counter Mode) and a Fuzzy Vault scheme, generating a compact, encrypted payload. Concurrently, the medical image (DICOM) undergoes modality-specific preprocessing to prepare it for embedding. Crucially, the centerpiece of our framework is a jointly trained embedder-extractor neural network that hides this payload with high imperceptibility, ensuring the watermark remains imperceptible to human eyes and designed to be undetectable by diagnostic algorithms, thereby pre- serving the clinical reliability of the scan. Extensive experiments confirm the method’s efficacy, achieving an outstanding mean Peak Signal-to-Noise Ratio (PSNR) of 41.37 dB, a Structural Similarity Index (SSIM) of 0.9804, and a near-perfect Bit Error Rate (BER) of 0.0020 in a no-attack scenario. Furthermore, the framework demonstrates notable resilience against common signal processing attacks, including noise addition and JPEG compression, proving its potential for deployment in secure clinical environments without compromising medical image quality.