Towards Deep Learning Models Resistant to Adversarial Attacks
Presented by
- Yongqi Dong
- Aden Grant
- Andrew McMurry
- Jameson Ngo
- Baizhi Song
- Yu Hao Wang
- Amy Xu
Introduction
What are adversarial attacks?
Any classifier can be tricked into giving the incorrect result. When an input is specifically designed to do this, it is called an adversarial attack. This is a prominent challenge for classifiers that are used for image processing and security systems: small changes to the input values that are imperceptible to the human eye can also fool high-level neural networks. As such, resistance to adversarial attacks has become an increasingly important aspect for classifiers to have and we need to develop methods that make models robust to adversarial inputs.
While there have been many approaches to defend against adversarial attacks, we can never be certain that these defenses will be able to robust against broad types of adversaries. Furthermore, these defenses can be easily evaded by stronger and adaptive adversaries.
This paper will use robust optimization to explore adversarial robustness of neural networks. In particular, the authors makes the following contributions:
- They conduct an experimental study of the saddle-point formulation. The saddle-point formulation is used for adversarially training. The authors propose that we can reliably solve the saddle-point optimization problem using first-order methods, particularly project gradient descent (PGD) and stochastic gradient descent (SGD), despite the problem's non-convexity and non-concavity.
- They explore how network architecture affects adversarial robustness. They conclude that networks need larger capacity in order to be resistant to strong adversaries.
- The authors train adversarially robust networks on MNIST and CIFAR10 using the saddle point formulation.