Generative Adversarial Networks in the Context of Cyber-Security

Generative adversarial networks (GANs) introduced by Goodfellow et al. in 2014 found many interesting applications across various domains. GANs enable to improve the performance of neural network-based classifiers as well as to enrich sample set of hard-to-obtain datasets. Moreover, a combination of two GANs (a.k.a., dual-GANs) enables to perform the unsupervised mapping between two different dataset domains. Such an extension can be further applied as a filter of noise in the datasets.

The goal of this thesis is to investigate existing application domains and suitable scenarios of GANs, while focusing on security aspects. For example, the application of GANs for extracting privacy-sensitive data might be analyzed. Next, the thesis should explore new approaches to attacks utilizing GANs and evaluate their success. Finally, the thesis should propose novel defense techniques and discuss their assumptions and limitations.