A Taxonomy of Watermarking Methods for AI-Generated Content

Pierre Fernandez, Hady Elsahar, Sylvestre-Alvise Rebuffi, Tomas Soucek, Valeriu Lacatusu, Tuan Tran, Alexandre Mourachko

Track: long paper (up to 9 pages)

Keywords: Watermarking, Taxonomy, Generative AI

TL;DR: A taxonomy of AI-generated content watermarking into post-hoc, out-of-model, and in-model methods, each with distinct trade-offs in flexibility, level of protection, and implementation complexity.

Abstract: As AI-generated content features more prominently in our lives, it becomes important to develop methods for tracing their origin. Watermarking is a promising approach, but a clear categorization of existing techniques is lacking. We propose a simple taxonomy of watermarking methods for generative AI based on where they are applied in the deployment of the models: (1) *post-hoc watermarking*, adding watermarks after content generation; (2) *out-of-model watermarking*, embedding watermarks during generation without modifying the model; (3) *in-model watermarking*, integrating watermarks directly into the model's parameters. By providing a structured overview of existing techniques across image, audio, and text domains, this taxonomy aims to help researchers, policymakers, and regulators make informed decisions about which approach best fits their needs, acknowledging that no single method is universally superior and that different approaches may be suited to specific use cases and requirements.

Presenter: ~Hady_Elsahar2

Format: Yes, the presenting author will definitely attend in person because they are attending ICLR for other complementary reasons.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 49