11 février 2026 à 14h30
The presentation attempts to define a reading grid in order to organize the security problems related to ML. It emphasizes the difference between robustness and security, adverse examples, backdoors, membership inference attacks. The reading grid is based on the type of AI (decision-making or generative), the access to the AI (white or black box), the security issues (intrinsic vulnerabilities of Machine Learning or malicious use of AI), and the ultimate goals (patch vulnerabilities, audit, or dissuasion of misuse). A definition of ML security relies on the basic definition of Machine Learning: learning a model from training data and applying it to some test data. The confidence in the results implies protecting the training data, the model, and the test data. Protecting resorts to the definition of security in IT: it means defending the cardinal values of confidentiality, privacy, and integrity.
Master Data Science Seminar
Inscription :
https://moodle.univ-lille.fr/course/view.php?id=11305
Bâtiment ESPRIT, Salle Agora
français
English