Institutional Repository of Vilnius University
Giliųjų neuroninių tinklų priešiškas apmokymas /
Abstract
dc:descriptionDeep Neural Network Adversarial Training Deep learning algorithms have a critical role in an increasing number of domains, yet studying their reliability and safety has not received much attention up until 2013 when the existence of adversarial examples - inputs that are carefully crafted to fool neural networks - was discovered. This discovery lead to dozens of new adversarial attacks being introduced as well as new methods to defend against them, but this deep neural network vulnerability problem is still far from being solved. In this work one of the most successful defense strategies - adversarial training - was chosen in order to improve a convolutional neural network model. Two types of adversarial training were used: PGD and ”free” adversarial training. Both of these defenses were evaluated against a few of the strongest whitebox attacks. It has been concluded that both of these methods are effective in making the model more robust. Increase in robustness was also linked to a significant decrease in model accuracy on clean samples. In order to evaluate the effect of training models adversarially using transfer learning, deep neural network was trained using ”free” adversarial training with two different techniques - finetuning and feature extraction. Of these methods only finetuning model was successful in learning robust features.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pyrantaitė, Greta,
- Contributors dc:contributor
-
- Valaitis, Vytautas
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD81705247&prefLang=en_US