Attack Agnostic Dataset
收藏资源简介:
Attack Agnostic Dataset是由弗罗茨瓦夫理工大学创建的一个音频深度伪造检测数据集,结合了两个深度伪造音频数据集FakeAVCeleb和WaveFake以及一个反欺骗数据集ASVspoof 2019 LA子集,总计包含31083个真实样本和222035个伪造样本,覆盖27种不同的伪造方法。该数据集旨在通过不重复使用攻击方法来提高检测方法的泛化能力和稳定性。数据集创建过程中,确保训练、测试和评估子集中的伪造方法不重叠,以验证模型的泛化性能。该数据集主要应用于音频深度伪造检测领域,旨在解决现有检测模型泛化能力不足和稳定性差的问题。
Attack Agnostic Dataset is an audio deepfake detection dataset developed by Wrocław University of Science and Technology. It combines two deepfake audio datasets, FakeAVCeleb and WaveFake, along with a subset of the anti-spoofing dataset ASVspoof 2019 LA. In total, the dataset contains 31,083 genuine samples and 222,035 forged samples, covering 27 distinct forgery methods. The core goal of this dataset is to enhance the generalization capability and stability of detection models by avoiding reuse of attack methods. During the dataset construction process, strict measures were implemented to ensure that the forgery methods in the training, test and evaluation subsets do not overlap, so as to validate the generalization performance of the models. This dataset is primarily applied in the field of audio deepfake detection, aiming to address the shortcomings of insufficient generalization ability and poor stability of existing detection models.

- 1Attack Agnostic Dataset: Towards Generalization and Stabilization of Audio DeepFake Detection弗罗茨瓦夫理工大学 · 2022年



