4- Adversarial Training: Attacks on Deep Networks and Generative Adversarial Models

Title: Adversarial Training: Attacks on Deep Networks and Generative Adversarial Models

Organizers:
Aykut Erdem, Hacettepe University, aykut{^a_*}cs.hacettepe.edu.tr
Erkut Erdem, Hacettepe University, erkut{^a_*}cs.hacettepe.edu.tr
Levent Karacan, Hacettepe University, karacan{^a_*}cs.hacettepe.edu.tr

Description:
A machine learning technique known as deep learning has made considerable progress in solving difficult problems in computer vision such as image classification and object detection. These models, which achieve prediction quality comparable to human performance, have begun to be deployed commercially in a large variety of areas from surveillance systems to autonomous vehicles. An approach that has emerged in the last few years in this rapidly developing field is the concept of adversarial training. This concept, which finds its way into a growing number of areas from the attacks to deep models to generative models, will form the main theme of our seminar.

Despite all the successes of deep neural network models, they are not indeed completely secure and robust. In this tutorial, we will first show that the convolutional neural networks which are commonly used in computer vision can be deceived by realistic or synthetic samples referred to as adversarial examples. In the second part of the tutorial, an introduction will be made to the recently proposed generative adversarial networks and how natural visual realistic images can be generated using this network model or its variants will be examined. Finally, we will provide information about the image editing applications we have developed based on generative adversarial networks and present our results.

Topics:

A quick review of convolutional neural networks
Approaches for generating adversarial examples
Defensive mechanisms against adversarial attacks
An introduction to generative adversarial networks
Image generation using generative adversarial networks
Applications based on generative adversarial networks

Tutorial Objectives:

To give the audience a well-founded understanding of adversarial examples, adversarial training and generative adversarial networks
To provide a comprehensive overview of the current applications of adversarial training and generative adversarial networks
To bring participants into a position where they can contribute to this research area

Target Audience:
The target audience of this tutorial is those who want to familiarize with the limitations of convolutional neural networks, and to learn about adversarial training and generative adversarial networks. The tutorial content is mainly focused on those who do not have knowledge on these subjects, but there will also be some advanced topics for the audience who has a basic knowledge.

Speaker Biographies:
Aykut Erdem received his bachelor's and master's degrees in Computer Engineering from Middle East Technical University (METU) in 2001 and 2003. During his doctoral studies at the same institution, he was a guest researcher at Virginia Tech (Blacksburg, USA) in the summer of 2004, and a visiting scholar at MIT (Cambridge, USA) in the fall of 2007. After completing his doctorate in 2008, he worked as a post-doctoral researcher in the Department of Computer Science at the Ca 'Foscari University in Venice. He is currently working as Assistant Professor in the Department of Computer Engineering at Hacettepe University, where he joined in 2010, and is one of the founders of Hacettepe University Computer Vision Laboratory (HUCVL). The main purpose of his research is to find effective ways to understand, interpret and manipulate visual data, and recently he has been working intensively on image editing, visual data mining and integrating language and visual art. For more information: http://web.cs.hacettepe.edu.tr/~aykut/

Erkut Erdem received his bachelor's and master's degrees from Computer Engineering Department of Middle East Technical University in 2001 and 2003. In 2004 and 2007 she worked as a visiting researcher at Virginia Tech and the University of California, Los Angeles. After completing his doctoral studies at the Middle East Technical University in 2008, he pursued his post-doctoral research in 2009-2010 at Télécom ParisTech, Ecole Nationale Supérieure des Télécommunications. He started working as a lecturer in Computer Engineering Department of Hacettepe University in 2010 and since 2014 he has been working as Assistant Professor in the same department. He is one of the founders of Hacettepe University Computer Vision Laboratory. His research interests are computer vision and machine learning in general, and he is particularly involved in image editing and matting, visual saliency estimation, and integrated language and visual applications. For more information: http://web.cs.hacettepe.edu.tr/~erkut/

Levent Karacan received his undergraduate degree from Erciyes University Computer Engineering Department in 2011. He got his master's degree from Hacettepe University Department of Computer Engineering with his thesis on image smoothing based on texture and structure decomposition in 2014. He is currently a PhD candidate in the same department, pursuing his doctoral studies under the supervision of Aykut Erdem and Erkut Erdem. His research interests include image editing and smoothing, and more recently, deep generative models. He has presented his studies at prestigious conferences such as SIGGRAPH Asia, IEEE International Conference on Computer Vision (ICCV). For more information: http://web.cs.hacettepe.edu.tr/~karacan/