SoundnessBench
收藏资源简介:
SoundnessBench数据集由加州大学洛杉矶分校、杜克大学和伊利诺伊大学厄巴纳-香槟分校的研究团队创建,旨在评估神经网络验证器的可靠性。该数据集包含206个隐藏的反例,这些反例被设计为难以通过常规对抗攻击发现。数据集的创建过程涉及训练神经网络,使其在大多数输入上表现正确和鲁棒,但在预定义的反例上表现错误。该数据集主要用于测试神经网络验证器的声誉性,旨在解决现有验证器在面对隐藏反例时可能出现的误判问题。
The SoundnessBench dataset was developed by research teams from the University of California, Los Angeles, Duke University, and the University of Illinois Urbana-Champaign, with the goal of evaluating the reliability of neural network verifiers. It contains 206 hidden adversarial examples that are designed to be difficult to detect via conventional adversarial attacks. The dataset creation process involves training neural networks to perform correctly and robustly on most inputs, while behaving erroneously on predefined adversarial examples. This dataset is primarily used to test the soundness of neural network verifiers, aiming to address potential misjudgment issues of existing verifiers when faced with these hidden adversarial examples.




