The next NLP Seminar, entitled “Watermarking AI-generated content” will be given by Pierre Fernandez (Meta FAIR Paris) and will take place on 10th March, in Loria’s room A008 Jean Legras, at 10:00.

Abstract:
Invisible watermarking involves embedding a hidden signal into digital media like text, images, and audio. It has evolved from a tool for copyright protection to a critical solution for AI safety and security, for example to enhance the detection of AI-generated content. For Generative AI, watermarking can be implemented at various stages: post-hoc or at generation-time. We will explore this across modalities, relying on the following works done in the lab:

• Proactive Detection of Voice Cloning with Localized Watermarking https://arxiv.org/abs/2401.17264
• Three Bricks to Consolidate Watermarks for Large Language Models https://arxiv.org/abs/2308.00113
• The Stable Signature: Rooting Watermarks in Latent Diffusion Models https://arxiv.org/abs/2303.15435
• Latent Watermarking of Audio Generative Models https://arxiv.org/abs/2409.02915

The presentation will conclude with a discussion on the limitations and future prospects of invisible watermarking.

Bio:
Pierre Fernandez is a research scientist at Meta FAIR Paris (Meta Fundamental AI Research lab). His research focuses on watermarking, particularly for generative models to ensure AI-generated content is traceable and identifiable. During his PhD at FAIR and Inria Rennes, he notably developed the audio watermarking algorithm used in FAIR’s public research demos like Audiobox and Seamless, as well as the image watermarking algorithm for Meta AI image generation assistant (Imagine).