神经架构设计和鲁棒性评估数据库

神经架构设计和鲁棒性评估数据库是由马克斯·普朗克信息学研究所创建，包含6466个非同构网络设计，用于评估深度学习模型在对抗性攻击和数据损坏情况下的鲁棒性。数据集基于广泛使用的NAS-Bench-201搜索空间，通过对所有网络在多种攻击和损坏类型下的评估，提供了关于网络架构设计对鲁棒性影响的重要见解。此数据集不仅支持研究者评估和比较不同网络的鲁棒性，还展示了通过精心设计网络拓扑可以显著提升模型的鲁棒性，从而推动了神经架构搜索在鲁棒模型设计中的应用。

The Database for Neural Architecture Design and Robustness Evaluation was developed by the Max Planck Institute for Informatics. It contains 6,466 non-isomorphic network architectures, and is intended for evaluating the robustness of deep learning models against adversarial attacks and data corruptions. Built upon the widely adopted NAS-Bench-201 search space, this dataset provides critical insights into the impact of neural architecture design on model robustness via comprehensive evaluations of all included networks across diverse attack and corruption types. This dataset not only empowers researchers to evaluate and compare the robustness of distinct network architectures, but also demonstrates that meticulously designed network topologies can substantially boost model robustness, thereby advancing the application of neural architecture search (NAS) in robust model design.