Abstract:

The rapid advancement of generative AI has created unprecedented challenges for the authenticity and security of digital content. Traditional watermarking methods—based on spatial, frequency, or deep learning—are proving increasingly vulnerable to contemporary attacks, including adversarial perturbations, generative model regeneration, and latent space manipulation. This paper examines the threats posed to watermarking and highlights diffusion models as a promising foundation for robust watermarking. Two representative approaches, ZoDiac and SuperMark, are presented and their characteristics and empirical robustness are compared, both being training-free approaches. Experimental evidence shows that diffusion-based watermarking achieves near-perfect robustness (>99%) against standard distortions and high robustness against adaptive AI-based attacks, while maintaining high fidelity. These findings suggest that diffusion models provide an inherently robust framework that can overcome the limitations of conventional watermarking strategies.