Adversarial robustness and attacks in Deep Learning Керівни
Loading...
Date
2022
Authors
Кузьменко, Дмитро
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The theoretical underpinnings for this field involve the notions of robustness and astuteness, local Lipschitzness, r-separability of datasets, robustness-accuracy tradeoff, and L-inf distance. This work will cover all the preliminaries, explain the choice of CIFAR-10 with L-inf metric space and eps=8/255 as a main dataset for the task, make use of already well-known attacks and defenses, introduce new ones, and try different ensembles on the 3 most robust models available on the benchmark – Adversarial Weight Perturbation, Augmentations and weight averaging, and Self-COnsistent Robust Error (SCORE-based model).
Description
Keywords
Local Lipschitzness, Carlini-Wagner, Self-COnsistent Robust Error, Scale and std hyperparameters in Jitter, магістерська робота