Presenter Information

Anna BuhmanFollow

Advisor Information

Dr. Raj Dasgupta

Location

Dr. C.C. and Mabel L. Criss Library

Presentation Type

Poster

Start Date

2-3-2018 2:15 PM

End Date

2-3-2018 3:30 PM

Abstract

A well-trained neural network is very accurate when classifying data into different categories. However, a malicious adversary can fool a neural network through tiny changes to the data, called perturbations, that would not even be detectable to a human. This makes neural networks vulnerable to influence by an attacker. Generative Adversarial Networks (GANs) have been developed as one possible solution to this problem [1]. A GAN consists of two neural networks, a generator and a discriminator. The discriminator tries to learn how to classify data into categories. The generator stands in for the attacker and tries to discover the best way to cause the discriminator to make wrong classifications through perturbing the input. Our work improves on this method through the application of Bayesian games to model multiple generators and discriminators rather than one of each. Through training against multiple types of input perturbation, the discriminators will improve their classification of adversarial samples.

COinS
 
Mar 2nd, 2:15 PM Mar 2nd, 3:30 PM

Towards Robust Classification in Adversarial Learning using Bayesian Games

Dr. C.C. and Mabel L. Criss Library

A well-trained neural network is very accurate when classifying data into different categories. However, a malicious adversary can fool a neural network through tiny changes to the data, called perturbations, that would not even be detectable to a human. This makes neural networks vulnerable to influence by an attacker. Generative Adversarial Networks (GANs) have been developed as one possible solution to this problem [1]. A GAN consists of two neural networks, a generator and a discriminator. The discriminator tries to learn how to classify data into categories. The generator stands in for the attacker and tries to discover the best way to cause the discriminator to make wrong classifications through perturbing the input. Our work improves on this method through the application of Bayesian games to model multiple generators and discriminators rather than one of each. Through training against multiple types of input perturbation, the discriminators will improve their classification of adversarial samples.