PhD, École Polytechnique Fédérale de Lausanne
Robustness of Image Classifiers
In this talk, we will present tools to assess the robustness of image classifiers to a diverse set of perturbations, ranging from adversarial to random noise. In particular, we will propose a semi-random noise regime that generalizes both the random and adversarial noise regimes. We provide theoretical bounds on the robustness of classifiers in this general regime, which depends on the curvature of the classifier's decision boundary. In a final part of the talk, we will show how it can explain the surprising existence of universal perturbation images that cause most natural images to be misclassified by state-of-the-art deep neural network classifiers.
Seyed Mohsen Moosavi is a PhD student in the school of Communication and Computer Sciences at EPFL. Prior to that, he received his B.Sc. in Electrical Engineering from Amirkabir University of Technology (Tehran Polytechnic) and his M.Sc. in Communication Systems from EPFL in 2012 and 2014 respectively. During his master’s studies, he worked on random tomography as a research assistant under the supervision of Prof. Martin Vetterli. He then stayed at ABB Corporate Research Center for six months. He started his PhD in September 2014 and he is currently working on the robustness of deep neural networks at Signal Processing Laboratory 4 under the supervision of Prof. Pascal Frossard.
Watch video from here.