ImageShield: a responsibility-to-person blind watermarking mechanism for image datasets protection

• Published in: Applied intelligence (Dordrecht, Netherlands) Netherlands), 2025-01, Vol.55 (1), p.84
• Format: Article
• Language: English
• Subjects: Computer vision; Copy protection; Datasets; Feature extraction; Generative adversarial networks; Image enhancement; Image quality; Robustness (mathematics); Watermarking
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-024-06093-7

High-quality diverse image datasets, particularly those featuring faces and medical records, are essential for deep learning applications in computer vision. Protecting the copyright and privacy of such datasets remains a challenge, as existing blind watermarking methods fall short of effectively monitoring and restricting unauthorized access. To address this issue, we propose ImageShield, a responsibility-to-person blind watermarking mechanism for image datasets protection. It introduces a hybrid approach, combining traditional transform domain watermarking with an enhanced generative adversarial network (GAN), achieving an optimal balance between watermark imperceptibility and robustness. In the extraction phase, an optimized network architecture integrates the enhanced GAN with a specially designed attack layer, improving the efficiency of watermark feature retrieval, and reducing computational overhead. The method ensures high fidelity in extracting watermark features, even under potential distortions or attacks, by leveraging the robust structure of the GAN and attack layer. Experimental results on the Helen datasets and custom datasets demonstrate ImageShield’s superiority in terms of imperceptibility (PSNR: 37.2963 dB, SSIM: 0.9906), robustness, watermark embedding capacity, and execution efficiency. These contributions offer a novel solution to enhance the security and traceability of protected image datasets.