Eva Giboulot is an Inria researcher in Rennes, as a member of the Artishau Team.. She received the Ph.D. Degree in Engineering science from the University of Technology of Troyes, France with research activities in information security and signal processing, focusing on steganography and steganalysis. She served as a Post-Doc Researcher at the Artificial Intelligence Center of the Czech Technical University, Prague, Czechia. Her current interests are in the area of AI security, particularly in the design of secure watermarking methods as well as theoretically grounded forensics methods for the detection of generated content. She is a member of the IEEE Information Forensics and Security Technical Committee since 2025.

Watermarking is the secure communication of a mark over an unreliable channel. This tutorial aims to present a general methodology for designing inference-based text watermarking. In particular, the goal is to teach how to construct watermarking schemes which possess explicit theoretical guarantees in terms of probability of false alarm and, when possible, in terms of detectability. We start with the design of a powerful, yet naïve, approach: the KGW algorithm. From this baseline we touch upon the most common design mistakes of watermarking in the literature: the use of z-scores, insecure designs and poorly defined quality metrics. To address these issues, we construct a general watermarking framework based on statistical hypothesis testing and information theory. We illustrate the framework with a plug-and-play zero-bit scheme, WaterMax, as well as a tentative multi-bit one. We conclude with the limitations of inference-based schemes, notably the difficulty of using them in an open-source setting, as well as potential remedies: weight-based schemes and finetuning.

Marija Slavkovik is a Professor with the Faculty for Social Sciences of the University of Bergen. Her background is in computer science and artificial intelligence. She has been doing research in machine ethics since 2012. Machine ethics studies how moral reasoning can or should be automated. Marija works on formalising ethical collective decision-making. She has held held several seminars, tutorials and graduate courses on AI ethics (http://slavkovik.com/teaching.html). Marija is a vice-chair of the t f European Association for Artificial Intelligence, and Society track editor of JAIR. Her publications can also be found here (https://scholar.google.com/citations?user=TVmqPq0AAAAJ&hl=en&oi=ao)

I am a junior professor in the department of mathematics at Sorbonne University. My main line of research is in statistical machine learning with a focus on the privacy and robustness of machine learning algorithms. Most of my work has a theoretical flavor, but I also like to work on more applied projects to get deeper understanding of state-of-the-art methods, or simply to better grasp the gap that can exist between the theoretical analysis and the empirical performance of some machine learning algorithms.

The vast amount of data collected every day, combined with the increasing complexity of machine learning models, has led to the emergence of distributed learning schemes. In the now classical Federated learning architecture, the learning procedure consists of multiple data owners (or clients) collaborating to build a global model with the help of a central entity (the server), typically using a distributed variant of SGD. Nevertheless, this algorithm is vulnerable to « misbehaving » clients that could (either intentionally or inadvertently) sabotage the learning by sending arbitrarily bad gradients to the server. These clients are commonly referred to as Byzantine and can model very versatile behaviors going from crashing machines in a datacenter to colluding bots attempting to biais the outcome of a poll on the internet. The purpose of this talk is to present a small introduction the emerging topic of Byzantine-Robustness. Essentially, the goal is to enhance distributed optimization algorithms, such as distributed SGD, in a way that guarantees convergence despite the presence of some Byzantine clients. We will take the time to present the setting and review some recent results as well as open problems in the community.

Maximilian Kiener is Head of the Institute for Ethics in Technology at Hamburg University of Technology (tuhh.de/ethics) and Associate Research Fellow at the Oxford Uehiro Institute for Practical Ethics. He holds a BPhil and DPhil in Philosophy from the University of Oxford, where he also served as an Extraordinary Junior Research Fellow and Leverhulme Early Career Fellow. His research explores the ethics of artificial intelligence, with a particular emphasis on ethics by design and moral responsibility.

This talk presents deep ethics as an emerging field concerned with translating qualitative ethical reasoning into computational forms, including metrics, models, and decision procedures. Deep ethics promises to embed ethical reasoning directly into AI design. But it also risks excluding what cannot be measured, disguising assumptions as technical facts, and narrowing practical judgement. The central question is whether ethics can be made computable without losing the openness and contestability that make ethical deliberation possible.

This course addresses the verification and certification of vision-based landing systems in aeronautics under the ARP AI standard framework. Participants will explore robust testing via adversarial attacks, robust validation, and formal verification tailored to object detection metrics like IoU. The curriculum covers formal explainability for failure hazard analysis and provides hands-on Colab tutorials using PyTorch, auto_LiRPA solvers, and the LARD dataset to deliver actionable industrial certification assets.