A3D: Joint Search for Robust DNNs and Adversarial Attacks via AutoML

Abstract

Deep neural networks (DNNs) are widely used today. These methods work well in many tasks but show weakness against adversarial attacks. Attacks use small changes in input images to fool DNNs. Because of this, many platforms exist to test the strength or weakness of a DNN. However, most current platforms face two problems. First, improving the structure of neural networks is not possible. Second, creating stronger attacks is not supported. As a result, these platforms do not help make DNNs more secure or smarter. To address this, a new system was proposed named A3D. It stands for Auto-Adversarial Attack and Defense. This platform does two main jobs. It finds strong DNN structures. It finds smart attack methods. This is done by using automatic search techniques. A3D uses different search tools. These tools help to build better DNNs that are harder to fool. A3D finds better ways to fool the DNNs. The system works in both directions. Stronger the attacks, better the defense models it creates. The stronger the models, the smarter the attacks it designs. A3D is tested on popular datasets. These include CIFAR10, CIFAR100 and ImageNet. The results show that A3D works well. It creates DNNs that are more secure. It creates attacks that are more effective. A3D is a powerful tool. It helps researchers test and improve DNNs in a smart and automatic way. It solves the limits of older systems. And it brings better security and performance to deep learning.