${A^{3}D}$A3D: A Platform of Searching for Robust Neural Architectures and Efficient Adversarial Attacks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) require robust evaluation against adversarial attacks.
• Existing platforms lack capabilities for optimizing DNN architectures and adversarial attack configurations.
• There is a need for integrated solutions to enhance DNN robustness and attack effectiveness.

Purpose of the Study:

• To propose a novel platform, A3D (auto-adversarial attack and defense), for enhancing DNN robustness.
• To enable automated searching for robust neural network architectures and efficient adversarial attacks.
• To integrate auto-adversarial attack and defense into a unified framework.

Main Methods:

• A3D integrates multiple neural architecture search methods to identify robust architectures.
• The platform employs various optimization algorithms to discover efficient adversarial attacks.
• A unified framework combines automated attack and defense strategies.

Main Results:

• Experiments on CIFAR10, CIFAR100, and ImageNet datasets validate the platform's effectiveness.
• The proposed A3D platform successfully searches for robust architectures and efficient attacks.
• The integrated framework enhances DNN robustness through novel evaluation and threat modeling.

Conclusions:

• A3D provides a feasible and effective solution for improving DNN robustness.
• The platform advances the field by automating the optimization of adversarial attacks and defenses.
• A3D offers a unified approach to enhance both model security and adversarial attack capabilities.