ELFT HÍRADÓ

[mcsete AT physx.u-szeged.hu]

Seminar

Speaker: Prof. Márk Jelasity
Department of Algorithms and AI
University of Szeged

Date: 07, February (Wednesday) 2024, 3:00 pm
Location: USZ FSI IP, Budó auditorium & online

Title:
Adversarial Robustness of Deep Neural Networks

Deep neural networks (DNNs) are the foundations of AI systems, yet  
they are known to have several vulnerabilities that are unsolved to  
this day. A particularly interesting vulnerability is that it is  
possible to construct inputs to these networks that are extremely  
close to natural inputs (eg. invisible perturbations to images) yet  
result in completely unexpected behaviour. This adversarial behaviour  
can even be designed in advance. In the talk I will give several  
examples of this problem, and I will also present some of our own  
results in the area. I will briefly touch on the formal verification  
of neural networks from the point of view of adversarial robustness,  
and I will also discuss some applications like attacking semantic  
image segmentation networks and attacking entire ensembles of models.

Everyone is welcome to attend.

Dr. Mária Csete


----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.

Attachment: png0YkHuN4B7o.png
Description: PNG image