Seminar on Adversarial Deep Learning
Deep learning systems achieve remarkable accuracy on many tasks such as image classification. However, they can be fooled by adversarial attacks, i.e. modification of data in a way that is unnoticeable to a human, but that completely changes the predictions made by the system. In this course, we aim to resolve the apparent paradox between the accuracy and vulnerability of deep learning systems. We will also study various adversarial attacks techniques, attempts to defend against them, and ways of using adversarial examples to improve training.
The seminar is aimed for students who want to understand better the capabilities and limitations of deep learning in theory, or who want to learn to defend against adversarial attacks in practice. We assume you have previous experience with implementing and studying deep learning systems.
The first meeting will be on Tuesday at 14-16 in room R001/U259 of Otakaari 1. Enroll at https://oodi.aalto.fi/a/opettaptied.jsp?OpetTap=1134770810