Abstract & Overview
With the thriving of generative artificial intelligence, an unprecedented crisis of content authenticity isbecoming real because any type of text, image, audio, and video can no longer be perceived as thework of humans. This is a systematic article on generative watermarking as an active paradigm ofauthentication that answers four research questions on the technical strategies, robustness, adoptionand the future. Key discoveries include the fact that image watermarking has reached a high degree ofmaturity, with the most advanced systems recording 93% true positive rates and not degrading tounder 38 dB compression at imperceptibility (PSNR). Nevertheless, text watermarking is susceptible,falling below 70% accuracy with paraphrasing attacks, and regeneration attacks lower all existingtechniques to close to random detection. Regardless of the regulatory requirements on theimplementation of watermarking by the European Union and China, only 38 percent of platformsapply verified watermarking. It concludes in the review that watermarking is not a sufficient tool andthat multi-layered architectures between watermarking and provenance tracking, cryptographicsignatures, and harmonized international standards are necessary to achieve effective contentauthentication.