This academic paper discusses the vulnerability of Automatic Speaker Verification (ASV) systems built on deep neural networks to adversarial attacks. The researchers have developed an open-source adversarial attack dataset for speaker verification research, with a focus on over-the-air attacks. The dataset simulates over-the-air attack settings and can be extended to include more types of adversarial attacks. It is constructed using the Voxceleb1 Verification test set as its foundation and will be released to the public under the CC-BY license. The paper also provides a detection baseline for reproducible research.
Publication date: 10 Oct 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2310.05369
