Adversarial Separation Network for Speaker Recognition

Zhang, Hanyi; Wang, Longbiao; Zhang, Yunchun; Li, Meng; Lee, Kong Aik; Wei, Jianguo

doi:10.21437/interspeech.2020-1966

Cited by 21 publications

(21 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, due to limited efforts [26][27][28][29][30][31][32] in adversarial defense on ASV, effective strategies remain an open question. Wang et al [26] adopt adversarial training to mitigate adversarial attacks for ASV by injecting adversarial data into the training set.…”

Section: Githubmentioning

confidence: 99%

“…Li et al [27] propose a detection model for adversarial samples by training it on a mixture of adversarial samples and genuine samples. Zhang et al [28] harness an independent DNN filter trained with adversarial samples and apply it to purify the adversarial samples. However, the above methods [26][27][28] require the knowledge of the attack algorithms used by attackers.…”

Section: Githubmentioning

confidence: 99%

“…Zhang et al [28] harness an independent DNN filter trained with adversarial samples and apply it to purify the adversarial samples. However, the above methods [26][27][28] require the knowledge of the attack algorithms used by attackers. It is impractical to assume that the ASV system designers know in advance which attack algorithms will be implemented by attackers in-the-wild, not to mention such methods [26][27][28] may overfit to a specific adversarial attack algorithm.…”

Section: Githubmentioning

confidence: 99%

“…However, the above methods [26][27][28] require the knowledge of the attack algorithms used by attackers. It is impractical to assume that the ASV system designers know in advance which attack algorithms will be implemented by attackers in-the-wild, not to mention such methods [26][27][28] may overfit to a specific adversarial attack algorithm. Self-supervised learning models have also been proposed [29,30] as a filter to purify the adversarial noise.…”

Section: Githubmentioning

confidence: 99%

“…We propose the contrary to harness the vocoders to defend ASV systems. In contrast the approaches which need to know the adversarial attack method for defense [26][27][28], the proposed approach doesn't require such information. Compared to approaches that defend ASV in the frequency domain [29,30], the proposed approach directly detects adversarial samples in the time domain.…”

Section: Githubmentioning

confidence: 99%

See 4 more Smart Citations

Adversarial Sample Detection for Speaker Verification by Neural Vocoders

Wu¹,

Hsu²,

Zhang³

et al. 2021

Preprint

View full text Add to dashboard Cite

Automatic speaker verification (ASV), one of the most important technology for biometric identification, has been widely adopted in security-critical applications, including transaction authentication and access control. However, previous work has shown that ASV is seriously vulnerable to recently emerged adversarial attacks, yet effective countermeasures against them are limited. In this paper, we adopt neural vocoders to spot adversarial samples for ASV. We use the neural vocoder to re-synthesize audio and find that the difference between the ASV scores for the original and re-synthesized audio is a good indicator for discrimination between genuine and adversarial samples. This effort is, to the best of our knowledge, among the first to pursue such a technical direction for detecting adversarial samples for ASV, and hence there is a lack of established baselines for comparison. Consequently, we implement the Griffin-Lim algorithm as the detection baseline. The proposed approach achieves effective detection performance that outperforms all the baselines in all the settings. We also show that the neural vocoder adopted in the detection framework is dataset-independent. Our codes will be made open-source for future works to do fair comparison 1 .

show abstract