ANALYSIS OF MULTIPLE ADVERSARIAL ATTACKS ON CONVOLUTIONAL NEURAL NETWORKS

The thesis studies different kind of adversarial attacks on Convolutional Neural Network by using electric load data set in order to fool deep neural network. With the improvement of Deep Learning methods, their securities and vulnerabilities have become an important research subject. An adversary who gains access to the model and data sets may add some perturbations to the datasets, which may cause significant damage to the system. By using adversarial attacks, it shows how much these attacks affect the system and shows the attacks' success in this research.