Speaker
Andras Horvath
(Pázmány Péter Catholic University - Faculty of Information Technology and Bionics)
Description
Modern machine learning enabled the solution of complex problems, where a vast amount of data is available.
Unfortunately in many cases the developed solutions are not robust and even the slightest modifications can drastically change the networks output. The most interesting deliberate modifications are the so called adversarial attacks, which can be considered as optical illusions for neural networks.
In this talk I would like to introduce some theoretical results which can help in the development of robust solutions against adversarial attacks.
Primary author
Andras Horvath
(Pázmány Péter Catholic University - Faculty of Information Technology and Bionics)
