Title
Classification models at Image Recognition Tasks: Assessing the Risks

Abstract
Nowadays, Convolutional Neural Networks (CNNs) and Transformers are probably the most largely used systems performing automatic classification of images. They are used in a very large series of applications and contexts: self-driving vehicles, medical diagnoses, monitoring of critical infrastructures, anti-terrorism, etc. However, these systems are exposed to errors. Misclassifications may occur for benign reasons; they may also be the result of specifically designed attacks. On the one hand, such misclassifications may have dramatic consequences for obvious reasons. On the other hand, they may also be used to protect the private sphere of individuals on social media. The talk will give an overview of the attack scenarios, the topography of the attacks, and will present some of the most efficient (in terms of speed, visual quality and hostility) attacks as of today, mentioning in particular some very recent black-box attacks developed by our research group, one of them achieving success rates and speed competitive against white-box attacks even.

Biodata
Born in Cherbourg in 1965, Franck Leprévost is a French mathematician and computer scientist. He is professor at the University of Luxembourg since 2003, and was its Vice-President from 2005 to 2015. Before joining the University of Luxembourg, he was a researcher at the CNRS (Paris, France) and a professor at the University Joseph Fourier (Grenoble, France). He was researcher or visiting professor at many universities (Max Planck-Institut für Mathematik Bonn, Technische Universität Berlin, Peter the Great Saint Petersburg Polytechnic University, South Ural State University, Warsaw University of Technology, Cardinal Wyszynski University, Kiev Polytechnic Institute, Shanghai Normal University, etc). He was member of the board of directors or member of the scientific council of private and public entities (ATTF, CEPS, FNR, Luxtrust S.A., UNICA, etc.).

His scientific interests include algorithmic number theory, mathematics of cryptology, convolutional neural networks, deep learning, artificial intelligence and evolutionary algorithms on the one hand, and management of higher education and research organizations, international rankings and the civilizational role of universities on the other hand. He is the author of numerous articles in these domains, as well as tutorials for students. His reports for the European Parliament, in particular his contribution to the report on the Echelon network, have had a substantial technical and legal impact in most European countries.

Professor Leprévost's expertise in the field of higher education and research as well as his managerial skills are frequently solicited by organizations, governments and companies worldwide. Aside from his scientific and managerial activities, Professor Leprévost is an art collector, novelist, playwright and theatrical producer. His articles, books, position papers are available at www.franckleprevost.com