April 30, 2026 at 11 AM
Abstract:
The rapid advancements in new foundation models have unlocked transformative opportunities across various domains. By “new foundation models,” we refer to large-scale architectures trained on massive datasets, providing a versatile backbone for a wide range of specialized applications. These models include Large Language Models (LLMs) like ChatGPT and LLaMA, as well as diffusion models such as DALL-E 2, Imagen, and Stable Diffusion. These new foundation models excel at generating human-like text, producing realistic images, and solving complex problems, driving significant progress in AI. However, their adoption in sensitive areas such as healthcare, finance, and education raises critical concerns about trustworthiness, particularly regarding privacy, security, and safety. New foundation models pose privacy risks due to their training on vast datasets that may contain sensitive information, potentially leading to data leakage. They also face safety challenges, as harmful or inappropriate outputs could cause harm. Additionally, their complexity makes them vulnerable to security threats, including model poisoning and prompt injection, which can compromise functionality and trust. The misuse of these models for phishing, impersonation, or disinformation exacerbates these risks. In this presentation, I will discuss the key risks associated with new foundation models and present the four main axes that structure my research agenda for the coming years, focusing on privacy, safety, security, and collaboration in the era of new foundation models, as well as limiting the proliferation of deepfakes and misinformation. I will also highlight the interactions between these directions and hope to stimulate new internal collaborations.
It will also be possible to follow the talk online using the usual Renater Room:
https://rendez-vous.renater.fr/Magnet-Seminar_842d12-8ac1fb-0a26f8
Inria, bâtiment B21
français
English