Enhanced image integrity through self-referencing neural network watermarking

Abstract

With the rise of sophisticated image manipulation tools, ensuring the authenticity of digital images has become critical to combating misinformation and preserving trust in visual content. This paper introduces an innovative method leveraging deep learning principles to verify image integrity through watermarking techniques. We propose a novel CNN-based watermarking approach that generates self-referencing watermarks from image sub-blocks using SHA-256 hashing, eliminating the need for external keys and capable of both embedding and extracting watermarks, enabling the detection of tampering attempts. Rigorous evaluation against state-of-the-art techniques demonstrates improved performance in terms of accuracy, robustness, and efficiency, as measured by metrics such as PSNR (up to 45.72 dB), SSIM (up to 0.987), and tamper detection accuracy (98%). Our empirical findings confirm the method’s effectiveness in preserving image quality while exhibiting resilience to various image manipulations. Validation on established benchmark datasets, such as CIFAR-10 and Celeb-HQ256 and DIV2K and RAISE corroborates these results. By significantly enhancing the security and trustworthiness of digital images, our proposed approach offers a robust solution for ensuring image integrity.