ResMax: Detecting Voice Spoofing Attacks with Residual Network and Max Feature Map

Kwak, Il-Youp; Kwag, Sungsu; Lee, Junhee; Huh, Jun Ho; Lee, Choong-Hoon; Jeon, Youngbae; Hwang, Jeonghwan; Yoon, Ji Won

doi:10.1109/icpr48806.2021.9412165

Cited by 11 publications

(13 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In addition to Light CNN, various ResNet-based deep learning models have appeared. In particular, with ResMax (for 'Residual network with Max Feature Map') [18], the model showed high levels of performance in the form of equal error rates (EERs) of 2.19% for the LA attack set and 0.37% for the PA attack set, respectively, in the ASVspoof 2019 competition. [18] used a model that considers the MFM activation function in the residual block.…”

Section: A Voice Spoofing Attackmentioning

confidence: 99%

See 1 more Smart Citation

Experimental Case Study of Self-Supervised Learning for Voice Spoofing Detection

et al. 2023

Self Cite

View full text Add to dashboard Cite

This study aims to improve the performance of voice spoofing attack detection through self-supervised pretraining. Supervised learning needs appropriate input variables and corresponding labels for constructing the machine learning models that are to be applied. It is necessary to secure a large number of labeled datasets to improve the performance of supervised learning processes. However, labeling requires substantial inputs of time and effort. One of the methods for managing this requirement is self-supervised learning, which uses pseudo-labeling without the necessity for substantial human input. This study experimented with contrastive learning, a well-performing self-supervised learning approach, to construct a voice spoofing detection model. We applied MoCo's dynamic dictionary, SimCLR's symmetric loss, and COLA's bilinear similarity in our contrastive learning framework. Our model was trained using VoxCeleb data and voice data extracted from YouTube videos. Our self-supervised model improved the performance of the baseline model from 6.93% to 5.26% for a logical access (LA) scenario and improved the performance of the baseline model from 0.60% to 0.40% for a physical access (PA) scenario.

show abstract

Section: A Voice Spoofing Attackmentioning

confidence: 99%

“…For training, we used two features, STFT and CQT. We used ResMax models that had displayed good performance levels in the ASVspoof challenge as a base model [18]. The proposed method applies the data augmentation technique and has the following characteristics:…”

Section: E Experimental Case Study Of Contrastive Learning For Voice ...mentioning

confidence: 99%

Experimental Case Study of Self-Supervised Learning for Voice Spoofing Detection

et al. 2023

Self Cite

View full text Add to dashboard Cite

show abstract

“…In this competition, PCG signal data was converted into 2D features like Log-mel spectrogram, STFT, and CQT, and LCNN and ResMax models already been proven in many audio competitions ASVspoof 2017, 2019, and 2021 [8,[12][13][14] were applied.…”

Section: Modelsmentioning

confidence: 99%

“…ResMax is a model that showed excellent performance in the ASVspoof 2019 competition dataset [8]. The Res-Max model consists of four parameters.…”

Section: Resmaxmentioning

confidence: 99%

See 1 more Smart Citation

Deep Learning Based Heart Murmur Detection using Frequency-time Domain Features of Heartbeat Sounds

Lee¹,

Kang²,

Kim³

et al. 2022

Computing in Cardiology Conference (CinC)

View full text Add to dashboard Cite

The goal of the George B. Moody PhysioNet Challenge 2022 was to use heart sound recordings gathered from various auscultation locations to identify murmurs and clinical outcomes. Our team, CAU UMN, proposes a deep learning-based model that automatically identifies heart murmurs from a phonocardiogram (PCG). We converted the heartbeat sound into 2D features in the frequencytime domain through feature extraction techniques such as log-mel spectrogram, Short Time Fourier Transform (STFT), and Constant Q Transform (CQT). The frequencytemporal 2D features were modeled using voice classification models such as Convolutional neural networks (CNN) and Light CNN (LCNN). The model using log-mel spectrogram and LCNN was ranked 5th for murmur detection with a weighted accuracy of 0.767 and 5th for clinical outcome detection with a cost of 11933 in the test dataset of the George B. Moody PhysioNet Challenge. We believe that our deep learning based heart murmur detection system will be a promising system for automatic heart murmur detection from PCG.

show abstract

Voice Spoofing Detection Through Residual Network, Max Feature Map, and Depthwise Separable Convolution

et al. 2023

Self Cite

View full text Add to dashboard Cite

The goal of the ''2019 Automatic Speaker Verification Spoofing and Countermeasures Challenge'' (ASVspoof) was to make it easier to create systems that could identify voice spoofing attacks with high levels of accuracy. However, model complexity and latency requirements were not emphasized in the competition, despite the fact that they are stringent requirements for implementation in the real world. The majority of the top-performing solutions from the competition used an ensemble technique that merged numerous sophisticated deep learning models to maximize detection accuracy. Those approaches struggle with real-world deployment restrictions for voice assistants which would have restricted resources. We merged skip connection (from ResNet) and max feature map (from Light CNN) to create a compact system, and we tested its performance using the ASVspoof 2019 dataset. Our single model achieved a replay attack detection equal error rate (EER) of 0.30% on the evaluation set using an optimized constant Q transform (CQT) feature, outperforming the top ensemble system in the competition, which scored an EER of 0.39%. We experimented using depthwise separable convolutions (from MobileNet) to reduce model sizes; this resulted in an 84.3 percent reduction in parameter count (from 286K to 45K), while maintaining similar performance (EER of 0.36%). Additionally, we used Grad-CAM to clarify which spectrogram regions significantly contribute to the detection of fake data.

show abstract

ResMax: Detecting Voice Spoofing Attacks with Residual Network and Max Feature Map

Cited by 11 publications

References 16 publications

Experimental Case Study of Self-Supervised Learning for Voice Spoofing Detection

Experimental Case Study of Self-Supervised Learning for Voice Spoofing Detection

Deep Learning Based Heart Murmur Detection using Frequency-time Domain Features of Heartbeat Sounds

Voice Spoofing Detection Through Residual Network, Max Feature Map, and Depthwise Separable Convolution

Contact Info

Product

Resources

About