BEARD
收藏资源简介:
BEARD数据集是由北京航空航天大学和利物浦大学联合创建的,旨在评估数据集蒸馏方法的对抗鲁棒性。该数据集包含CIFAR10/100和TinyImageNet等多样化数据集,并整合了多种对抗攻击方法,如FGSM、PGD和C&W。通过对抗游戏框架,BEARD引入了三个关键指标:鲁棒性比率(RR)、攻击效率比率(AE)和综合鲁棒性-效率指数(CREI),以全面评估模型在蒸馏数据集上的对抗鲁棒性。该数据集的应用领域主要集中在深度学习模型的安全性和鲁棒性评估,旨在解决模型在面对对抗攻击时的脆弱性问题。
The BEARD dataset was jointly developed by Beihang University and the University of Liverpool, aiming to evaluate the adversarial robustness of dataset distillation methods. This dataset encompasses diverse benchmark datasets including CIFAR10/100 and TinyImageNet, and integrates multiple mainstream adversarial attack methods such as FGSM, PGD, and C&W. Through the adversarial game framework, BEARD introduces three core metrics: Robustness Ratio (RR), Attack Efficiency Ratio (AE), and Comprehensive Robustness-Efficiency Index (CREI), to comprehensively evaluate the adversarial robustness of models trained on distilled datasets. The application scenarios of this dataset mainly focus on the safety and robustness evaluation of deep learning models, aiming to address the vulnerability of models when facing adversarial attacks.




