SMARTMed: semi-fragile dual watermarking with deep tamper detection and adversarial DNN fingerprinting for secure...

Cloud-based electronic health records (EHRs) are vulnerable to modification during transmission and use, compromising security and privacy. To reduce this risk, we implemented a semi-fragile watermarking method using Slant Transform and chaotic sequences. This method creates a clinically sensitive tamper dataset. Transfer learning classifies images as original or tampered. Further, a black-box semi-fragile neural network watermarking is proposed, which can determine the model availability by distinguishing accidental modifications and malicious tampering using the generated fingerprints with the pretrained models such as ResNet, EfficientNet, and Vision Transformer architectures. If there is a tampering on the image, the semi-fragile watermarking supports tamper localization and self-recovery. Our approach resists clinically sensitive attacks and can self-recover tampered images, achieving high perceptual quality of ( PSNR of 44.91 dB and SSIM of 0.962), tamper detection (94.9%), and tamper localization (97.9%). The trained networks are watermarked to verify model integrity, and the model detects tampered images with 96.8% accuracy.