Embedded Files
Benedetta Tondi
University of Siena, Italy
Watermarking of Generative AI: From Copyright Protection to the Authentication of Media Contents
Abstract. As generative AIcontinues to evolve, concerns around transparency, authenticity, and intellectual property are becoming more pressing. Tools like GPT-4, DALL·E, and Midjourney, powered by large-scale foundation models, are now widely used to generate realistic images, text, and other media. However, their ability to produce synthetic content indistinguishable from human-generated material raises critical issues—from copyright infringement to the spread of misinformation and deepfakes.
In response to these challenges, model watermarking has emerged as a promising approach to protect the intellectual property of generative AI models and to promote trustworthy AI. By modifying the AI models in such a way that imperceptible signals are embedded into AI-generated outputs, watermarking enables downstream identification, authentication, and traceability of synthetic media.
After a general introduction on the topic of Deep Neural Network (DNN) watermarking, the talk delves into technical strategies for embedding robust watermarks into GANs and Diffusion Models—two of the most powerful families of generative models—in such a way that each model output contains an embedded watermark that can be reliably recovered during the identification phase. These watermarks enable copyright protection and provide a means to trace generated content back to the source model that produced it, supporting model attribution and content authentication. Methods embedding watermarks post-training are also discussed. The keynote aims to offer both a conceptual and practical understanding of watermarking as a cornerstone for building trust in generative AI, while also outlining open challenges and limitations.
Bio. Benedetta Tondi received her Master’s degree (cum laude) in Electronics and Communications Engineering from the University of Siena in 2012 and her PhD in Information Engineering and Mathematical Sciences in 2016, focusing on adversarial detection and multimedia forensics. She is currently an Assistant Professor in the Department of Information Engineering and Mathematics at the University of Siena and a member of the Visual Information Processing and Protection (VIPP) group led by Prof. Mauro Barni. She is also part of CNIT, IEEE Young Professionals, and the IEEE Signal Processing Society, and has served on the IEEE Information Forensics and Security Technical Committee since 2019. Her research spans multimedia security, adversarial signal processing, and information-theoretic and game-theoretic methods for forensics and counter-forensics, with recent work centered on machine learning security and deep learning–based digital forensics. She has contributed to major projects, including DARPA’s SemaFor program (as Co-PI) and the Italian PRIN Premier project. Dr. Tondi has served as Area Chair for ICASSP and ICIP, Technical Program Chair of ACM IH&MMSec 2022, and reviewer or editor for leading journals. She is a recipient of multiple Best Paper Awards at IEEE WIFS and the 2017 GTTI PhD Award.
Antitza Dantcheva
Inria, France
Generation and Detection of Deepfakes
Abstract. I will talk about our work related to design of generative models, which allow for realistic generation of talking heads. We have placed emphasis on disentangling motion from appearance and have learned motion representations directly from RGB, without structural representations such as facial landmarks or 3D meshes. We have aimed at constructing motion as linear displacement of codes in the latent space. Based on this, our model LIA (Latent Image Animator) and LIA-X are able to animate images via navigation in the latent space, allowing for control over generation.
While highly intriguing, video generation has thrusted upon us the imminent danger of deepfakes, which can offer unprecedented levels of increasingly realistic manipulated videos. Deepfakes pose an imminent security threat to us all, and to date, deepfakes are able to mislead face recognition systems, as well as humans. Hence, we design generation and detection methods in parallel.
In the second part of my talk, I will discuss our associated work.
Bio. Antitza Dantcheva is a Directrice de Recherche with the STARS team at the Inria Center of Université Côte d’Azur in Sophia Antipolis, France. She was previously a Marie Curie Fellow at Inria and a Postdoctoral Fellow at Michigan State University and West Virginia University, USA. She received her Ph.D. in image processing and biometrics from Telecom ParisTech/Eurecom, France in 2011 and obtained her Habilitation from the Université Côte d'Azur, France in 2021. Her research focuses on computer vision, particularly on developing algorithms for interpreting and generating human faces, with applications in security and healthcare. More recently, she has focused on generative models for realistic video generation and the disentanglement of appearance and motion in latent spaces. She is the recipient of several distinctions, including the ANR JCJC grant, the ECCV 2022 New Technology Show Award, the IEEE FG 2019 Best Poster Award, and the BEFA Challenge Award at ECCV 2018. Furthermore, she serves as Associate Editor for multiple journals, has been Workshop Chair at CVPR 2024, Area Chair at ECCV 2024 and ACM Multimedia 2022, a member of IEEE Biometrics Council and ELLIS network, and a reviewer for the French ANR and other research funding agencies.
Page updated
Google Sites
Report abuse