CN114566170A - Lightweight voice spoofing detection algorithm based on class-one classification - Google Patents

Abstract

The invention discloses a light-weight speech deception detection algorithm based on one-class classification. A new loss function DOC-Softmax is designed according to the characteristics of real speech and deceitful speech, that is, the deception speech space of the class-classification loss function OC-Softmax is designed. A decentralized loss function is introduced to alleviate the problem of mismatch of feature distribution between training data and test data, thereby improving the accuracy and generalization ability of the speech deception detection model. At the same time, the speech deception detection algorithm is designed as a lightweight speech deception detection algorithm by using the knowledge distillation framework, which reduces the number of parameters of the model and makes it easy to deploy into mobile terminals or embedded devices. This model generalizes better than the model obtained using the exact same model structure, training data, and using only the hard label training method.

Description

A Lightweight Speech Spoofing Detection Algorithm Based on One-Class Classification

technical field

The invention relates to the technical field of voiceprint recognition, in particular to a lightweight voice deception detection algorithm based on one-class classification.

Background technique

Voiceprint recognition refers to a biometric identification technology that recognizes the identity of the speaker based on the speaker information in the voice signal. In social security, government and enterprise, Internet of Things and other scenarios.

However, there are many uncertainties in the practical application environment, especially artificial malicious spoofing attacks, which make the performance of the existing voiceprint recognition system drop sharply. Voice spoofing refers to "modifying and counterfeiting" an illegal voice that has not been authenticated by the automatic speaker verification system by means of recording, speech synthesis, voice conversion, etc. to pass the detection of the automatic speaker verification system. Currently, there are mainly three kinds of spoofing attacks: (1) deliberate imitation of other speakers; (2) lifelike speech obtained by speech synthesis or speech conversion technology; (3) recording playback or recording splicing of high-fidelity recording equipment. Among the above three spoofing attacks, deliberately imitating this spoofing attack method can be identified by mainstream voiceprint recognition systems.

However, the improvement of the quality of recording equipment and the rapid development of speech processing technologies such as speech synthesis and speech conversion have brought more and more severe challenges to the security of speech deception detection and voiceprint recognition systems. Speech deception detection refers to the use of deep learning or machine learning methods to learn by inputting handcrafted features or raw speech into deep learning or machine learning models, and ultimately achieve the purpose of speech forgery. Next, we introduce residual network, knowledge distillation, Softmax loss function and AM-Softmax loss function.

(1) Residual network

In 2015, He Yuming et al. proposed Residual Network (ResNet) to alleviate the gradient disappearance problem caused by increasing network depth in deep neural network, and it is widely used in image classification, target detection, speech recognition and other fields. The residual network solves the degradation problem by introducing a deep residual learning framework. The main idea is to remove the same main part, highlight the small changes, and use some stacked nonlinear layers to fit a residual map F(x):= H(x)-x instead of fitting an underlying map H(x) directly, so the original map becomes F(x)+x, and optimizing the residual map is easier than optimizing the original map. This residual learning structure can be implemented by a forward neural network + shortcut connection. As shown in Figure 1, the shortcut connection is equivalent to performing the same mapping, without generating additional parameters and increasing computational complexity, and the entire network It can still be trained with end-to-end backpropagation.

(2) Knowledge distillation

Deep learning has achieved incredible performance in a wide range of fields such as computer vision, speech recognition, natural language processing, and more. However, most deep learning models are too computationally expensive to run on mobile or embedded devices. Therefore, it is necessary to compress the model, and knowledge distillation is one of the important techniques in model compression. Knowledge distillation was first proposed by Hinton et al. It mainly includes three distillation settings. One is model compression, which distills knowledge from complex models to small models; the other is cross-modal transfer knowledge, which transfers knowledge from one modality to another. One modality; the third is ensemble distillation, which distills knowledge from multiple models onto a single model. The invention utilizes the knowledge distillation framework to compress the model of a class of classification speech deception detection models. The knowledge distillation framework usually includes a complex model (called the Teacher model) and a small model (called the Student model). The complex model is generally a single complex network or a collection of several networks, which has good performance and generalization ability. Due to the small size of the network, the model has limited expressive ability. The knowledge distillation framework is to use the knowledge learned by the large model to guide the training of the small model, that is, the Soft-target predicted by the Teacher model is used to assist the Hard-target training, so that the Student model has the same performance as the Teacher model, which greatly reduces the model. parameters, so as to achieve model compression and reduce model inference delay.

(3) Softmax loss function and AM-Softmax loss function

The formula of the original Softmax loss function for binary classification is as follows:

为嵌入向量，y_i∈{0,1}为第i个样本的标签，y_i＝0表示该样本为目标类，y_i＝1表示该样本为非目标类，

是权重向量，N为一个批次中的样本数量。in,

is the embedding vector, y _i ∈{0,1} is the label of the ith sample, y _i =0 indicates that the sample is the target class, y _i =1 indicates that the sample is the non-target class,

is the weight vector, and N is the number of samples in a batch.

AM-Softmax improves the Softmax function by introducing an angular interval to make the embedding distribution of the two classes more compact. The specific formula is as follows:

和

分别为w₀,w₁和x的标准化表示。where α is a scaling factor, m is the cosine additive interval,

and

are the normalized representations of w ₀ , w ₁ and x, respectively.

For these two loss functions, the embedding vectors of the target and non-target classes tend to converge in two opposite directions, w ₀ -w ₁ and w ₁ -w ₀ , respectively. For the AM-softmax function, the target and non-target classes have the same compact edge, and the larger the value of m, the more compact the embedding.

In speech spoofing detection, it is reasonable to train a tight embedding space for real speech, but training a tight space for spoofing attacks may overfit known attacks. To sum up, although voice deception detection algorithms have achieved a lot of results, these algorithms generally have the problem of insufficient generalization of unknown deception attacks, because most methods express the problem of speech deception as two types of real speech and deception speech. Meta-classification problem, which essentially assumes the same or similar distribution of features between training and test data. While this assumption is reasonable for real-like speech, it is not so true for spoofed speech. Due to the development of speech deception techniques such as speech conversion, speech synthesis, etc., deception attacks in the training set may never catch up with the expansion of the distribution of actual deception attacks. In addition, most of the existing voice spoofing detection algorithms have the problems of complex network structure, large computational load and slow speed, and it is difficult to transplant into mobile terminals or embedded devices.

SUMMARY OF THE INVENTION

Aiming at the problem of mismatch of feature distribution between training data and test data, the invention proposes a class-based classification-based lightweight class for the purpose of improving the accuracy and generalization ability of the speech deception detection algorithm and reducing the model inference delay. Voice Spoofing Detection Algorithm.

In order to achieve the above object, the present invention provides the following technical solutions:

A lightweight speech deception detection algorithm based on one-class classification, using the knowledge distillation framework to learn a feature space through a one-class classification loss function DOC-Softmax based on dispersion loss, in this feature space, the real speech embedding has a Tight boundaries, there is a certain distance between the deceitful speech and the real speech, and at the same time, a dispersion loss is introduced in the deceitful speech feature space to maximize the distance of each deceitful speech sample to its center, so that the deceitful speech covers the entire deception speech. space.

Further, the total loss function L _DOCS of the one-class classification loss function DOC-Softmax based on dispersion loss is the weighted sum of the one-class classification loss L _OCS and the dispersion loss L _D , and the weight is λ. The specific formula is as follows:

是权重向量，

为w₀的标准化表示，是指真实语音的优化方向，α是一个缩放因子，两个间隔m₀和m₁分别被引入来限定真实语音、欺骗语音与真实语音权重向量之间的角度，即

与

间的角度θ_i，m₀,m₁∈[-1,1],m₀＞m₁；in,

is the weight vector,

is the normalized representation of w ₀ , which refers to the optimization direction of the real speech, α is a scaling factor, and two intervals m ₀ and m ₁ are respectively introduced to define the angle between the real speech, the spoofed speech and the real speech weight vector, namely

and

The angle between θ _i , m ₀ , m ₁ ∈[-1,1], m ₀ >m ₁ ;

The formula of the one-class classification loss function OC-Softmax is as follows:

与

间的角度θ_i，当y_i＝0，m₀用于使θ_i小于arccosm₀；当y_i＝1，m₁用于使θ_i大于arccosm₁，一个小的arccosm₀使目标类聚集在权重向量w₀，一个相对大的arccosm₁使非目标类远离权重向量w₀；Two intervals m ₀ and m ₁ , where m ₀ , m ₁ ∈ [-1,1], m ₀ >m ₁ , are introduced to define the real and spoofed samples, respectively

and

When y _i = 0, m ₀ is used to make θ _i smaller than arccosm ₀ ; when y _{i =} 1, m ₁ is used to make θ _i greater than arccosm ₁ , a small arccosm ₀ makes the target class cluster in weight vector w ₀ , a relatively large arccosm ₁ keeps non-target classes away from the weight vector w ₀ ;

The formula for introducing dispersion loss is as follows:

为嵌入向量，

为x的标准化表示，y_i∈{0,1}为第i个样本的标签，y_i＝0表示该样本为真实语音，y_i＝1表示该样本为欺骗语音，N为一个批次中的样本数量，M为一个批次中欺骗样本的数量，ε为一个很小的常数，用来避免出现分母为0的情况，

为每个批次中欺骗样本的中心，分散损失L_D是为了最大化欺骗语音样本

与他们中心μ的距离，使欺骗语音尽可能的覆盖整个欺骗区域。in,

is the embedding vector,

is the standardized representation of x, y _i ∈{0,1} is the label of the ith sample, y _i =0 indicates that the sample is real speech, y _i =1 indicates that the sample is deceptive speech, and N is a batch of , M is the number of cheat samples in a batch, ε is a small constant to avoid the case where the denominator is 0,

For the center of the spoofed samples in each batch, the dispersion loss _LD is to maximize the spoofed speech samples

The distance from their center μ makes the deception speech cover the entire deception area as much as possible.

Further, α=20, m ₀ =0.9, m ₁ =0.2.

Further, when the speech sample is real speech, that is, when y _i =0, m ₀ is used to make θ _i smaller than arccosm ₀ , and a small arccosm ₀ makes the real speech gather near the weight vector w ₀ ; when the speech sample is deception When speech, ie when y _i = 1, m ₁ is used to make θ _i larger than arccosm ₁ , a relatively large arccosm ₁ keeps the spoofing speech away from the weight vector w ₀ .

Further, the teacher model adopts the network structure based on the deep residual network ResNet-18, and uses attention pooling instead of global average pooling.

Further, the teacher model takes the extracted LFCC features as input, and takes the output of the fully connected layer as the embedding of the input speech.

此置信度分数代表输入语音属于真实语音还是欺骗语音的概率，，进而得到输入语音的分类结果。Further, the embedding is sent to the DOC-Softmax loss function to calculate the confidence score

This confidence score represents the probability that the input speech belongs to real speech or deceptive speech, and then the classification result of the input speech is obtained.

与教师模型输出的置信度分数

的均方误差损失L_MSE，具体公式如下：Further, the model architecture of the student model is basically the same as that of the teacher model, except that three residual modules are removed, and the soft label predicted by the teacher model is used to assist the hard label training of the student model. The loss function of the student model consists of two parts, one is One-class classification loss _LDOCS based on dispersion loss, one is the confidence score output by the student model

Confidence score with teacher model output

The mean square error loss L _MSE of , the specific formula is as follows:

The total loss function of the student model is to minimize the weighted sum of L _DOCS and L _MSE , and the weight is β. The specific formula is as follows:

Compared with the prior art, the beneficial effects of the present invention are:

The light-weight speech deception detection algorithm based on one-class classification of the present invention designs a new loss function DOC-Softmax according to the characteristics of real speech and deceitful speech, that is, introduces dispersion into the deception speech space of a class of classification loss function OC-Softmax The loss function is used to alleviate the problem of mismatch of feature distribution between training data and test data, thereby improving the accuracy and generalization ability of the speech deception detection model. At the same time, the speech deception detection algorithm is designed as a lightweight speech deception detection algorithm by using the knowledge distillation framework, which reduces the number of parameters of the model and makes it easy to deploy into mobile terminals or embedded devices. This model generalizes better than the model obtained using the exact same model structure, training data, and using only the hard label training method.

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only described in the present invention. For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings.

Figure 1 is a schematic diagram of the residual structure.

FIG. 2 is a model architecture diagram provided by an embodiment of the present invention.

FIG. 3 is a schematic diagram of a residual module provided by an embodiment of the present invention.

Figure 4 is a comparison of the four loss functions of Softmax, AM-Softmax, OC-Softmax, and DOC-Softmax. Figure 4(a) is Softmax, Figure 4(b) is AM-Softmax, and Figure 4(c) is OC -Softmax, Figure 4(d) is DOC-Softmax.

Detailed ways

The invention designs a light-weight speech deception detection algorithm based on one-class classification, which can well alleviate the problem of mismatch of feature distribution between training data and test data. In one-class classification methods, this distribution mismatch problem does not exist for target classes, while for non-target classes, the samples in the training set are either non-existent or statistically unrepresentative. The key idea of one-class classification methods is to capture the target class distribution and set a strict classification boundary around it so that all non-target data is kept outside the boundary.

The one-class classification loss function OC-Softmax is defined as follows:

它为目标类嵌入的优化方向。此公式中的

和

进行了标准化。两个间隔m₀和m₁(m₀,m₁∈[-1,1],m₀＞m₁)分别被引入来限定真实类和欺骗类样本

与

间的角度θ_i。当y_i＝0，m₀用于使θ_i小于arccosm₀；当y_i＝1，m₁用于使θ_i大于arccosm₁，一个小的arccosm₀可以使目标类聚集在权重向量w₀，一个相对大的arccosm₁可以使非目标类远离权重向量w₀。This loss function has only one weight vector

It is the optimization direction of the target class embedding. in this formula

and

standardized. Two intervals m ₀ and m ₁ (m ₀ , m ₁ ∈ [-1,1], m ₀ >m ₁ ) are introduced to define the real and spoofed samples, respectively

and

The angle θ _i between them. When _yi = 0, m ₀ is used to make θ _i smaller than arccosm ₀ ; when y _i =1, m ₁ is used to make θ _i greater than arccosm ₁ , a small arccosm ₀ can make the target class cluster in the weight vector w ₀ , A relatively large arccosm ₁ can keep non-target classes away from the weight vector w ₀ .

However, it can be seen from formula (3) that when the target class and the non-target class are closely around the weight vector w ₀ and the opposite direction w ₀ ′ of the weight vector respectively, the one-class classification loss function L _OCS reaches the minimum, so , the final direction of non-target class optimization is still closely around the opposite direction w ₀ ′ of the weight vector. When the OC-Softmax function is optimized well enough, this loss function is essentially no different from the AM-Softmax function.

In order to improve the accuracy and generalization ability of the voice deception detection algorithm, and reduce the model inference time delay, the invention proposes a lightweight voice deception detection algorithm based on one-class classification. A new loss function DOC-Softmax is introduced, that is, a decentralized loss function is introduced into the deception speech space of the above class of classification loss function OC-Softmax to alleviate the problem of mismatching feature distributions between training data and test data, thereby improving speech deception detection. Model accuracy and generalization ability. At the same time, the above-mentioned speech deception detection algorithm is designed as a lightweight speech deception detection algorithm by using the knowledge distillation framework, which reduces the amount of parameters of the model and makes it easy to deploy in mobile terminals or embedded devices.

The invention proposes a light-weight speech deception detection algorithm based on one-class classification, which is a single-feature, single-system speech deception detection algorithm. The algorithm defines the speech deception problem as a class of feature learning problems to improve the model's performance Generalization.

The method of the present invention mainly includes two aspects: one is the design of a class of classification loss functions based on dispersion loss, and the other is the design of a lightweight speech deception detection model based on knowledge distillation. By using a one-class classification loss function based on dispersion loss, a feature space is learned in which the real speech embeddings have a tight boundary and the spoofed speech is some distance from the real speech. At the same time, the dispersion loss is used to maximize the distance of each spoofing speech sample to its center, so that the spoofing speech covers the entire spoofing speech space. This method improves the generalization ability of the model by increasing the probability of unknown attacks entering the spoofing space.

In the embodiment of the present invention, the 60-dimensional linear frequency cepstral coefficient LFCC feature is extracted from the utterance, the network structure based on the deep residual network ResNet-18 is adopted, and the global average pooling is replaced by the attention pooling. The network structure takes the extracted LFCC features as input, and the output confidence score represents the classification result, which improves the performance of the speech deception detection algorithm without using any data augmentation, feature fusion and model ensemble methods. In addition, based on the knowledge distillation framework, we designed a lightweight speech deception detection model StudentNet for TeacherNet with the above network, which greatly reduces the model parameters (about 30 times the amount of parameters), which is convenient for the model to be deployed on mobile terminals or embedded. in the device.

1. Design of a class of classification loss function based on dispersion loss

The comparison of the four loss functions of Softmax, AM-Softmax, OC-Softmax, and DOC-Softmax is shown in Figure 4. Figure 4(a) is Softmax, Figure 4(b) is AM-Softmax, and Figure 4(c) is OC-Softmax, and Figure 4(d) is DOC-Softmax.

For the two loss functions, Softmax and AM-Softmax, the embedding vectors of the target and non-target classes tend to converge to two opposite directions, which are w ₀ -w ₁ and w ₁ -w ₀ , as Figure 4(ab) shows. The target and non-target classes in the AM-softmax loss function have the same compact edge, and the larger the value of m, the more compact the embedding. In speech spoofing detection, it is reasonable to train a tight embedding space for real speech, but training a tight space for spoofing attacks may overfit known attacks. One-class classification loss function OC-Softmax can solve this problem well by introducing two different intervals to better compact real speech and isolate deceptive speech, but it can be seen from formula (3) that when the target class, non-target class When the class is closely around the weight vector w ₀ and the opposite direction w ₀ ′ of the weight vector, the one-class classification loss function OC-Softmax reaches the minimum. Therefore, the final direction of non-target class optimization is still closely around the opposite direction w ₀ ' of the weight vector, as shown in Figure 4(c), when the OC-Softmax function is optimized well enough, this loss function essentially There is no difference with the AM-Softmax function. Therefore, we introduce a dispersion loss for non-target class samples with the following formula:

为嵌入向量，

为x的标准化表示，y_i∈{0,1}为第i个样本的标签，y_i＝0表示该样本为真实语音，y_i＝1表示该样本为欺骗语音，N为一个批次中的样本数量，M为一个批次中欺骗样本的数量。ε为一个很小的常数，用来避免出现分母为0的情况，

为每个批次中欺骗样本的中心。分散损失L_D是为了最大化欺骗语音样本

与他们中心μ的距离，使欺骗语音尽可能的覆盖整个欺骗区域，也就是使欺骗语音尽可能的覆盖θ_i大于arccosm₁的整个区域，如图4(d)所示，从而提高欺骗语音进入欺骗区域的概率，提高模型的泛化性能。因此，总的损失函数L_DOCS为一类分类损失L_OCS和分散损失L_D的加权和，权重为λ，具体公式如下：in,

is the embedding vector,

is the standardized representation of x, y _i ∈{0,1} is the label of the ith sample, y _i =0 indicates that the sample is real speech, y _i =1 indicates that the sample is deceptive speech, and N is a batch of The number of samples, M is the number of cheat samples in a batch. ε is a small constant to avoid the situation where the denominator is 0,

is the center of the spoofed samples in each batch. The dispersion loss _LD is to maximize the deception speech samples

The distance from their center μ makes the cheating voice cover the entire cheating area as much as possible, that is, the cheating voice covers the entire area where θ _i is greater than arccosm ₁ as much as possible, as shown in Fig. 4(d), thereby improving the entry of the cheating voice The probability of deceiving regions improves the generalization performance of the model. Therefore, the total loss function L _DOCS is the weighted sum of the one-class classification loss L _OCS and the dispersion loss L _D , and the weight is λ. The specific formula is as follows:

是权重向量，

为w₀的标准化表示，它为真实语音的优化方向。α是一个缩放因子，本发明中α＝20。两个间隔m₀和m₁(m₀,m₁∈[-1,1],m₀＞m₁)分别被引入来限定真实语音、欺骗语音与真实语音权重向量之间的角度，即

与

间的角度θ_i，本发明中取m₀＝0.9，m₁＝0.2。当语音样本为真实语音时，即y_i＝0时，m₀用于使θ_i小于arccosm₀，一个小的arccosm₀可以使真实语音聚集在权重向量w₀附近；当语音样本为欺骗语音时，即y_i＝1时，m₁用于使θ_i大于arccosm₁，一个相对大的arccosm₁可以使欺骗语音远离权重向量w₀。in,

is the weight vector,

is the normalized representation of w ₀ , which is the optimized direction for real speech. α is a scaling factor, α=20 in the present invention. Two intervals m ₀ and m ₁ (m ₀ , m ₁ ∈ [-1,1], m ₀ >m ₁ ) are respectively introduced to define the angle between the real speech, the spoofed speech and the real speech weight vector, i.e.

and

The angle θ _i between them is taken as m ₀ =0.9 and m ₁ =0.2 in the present invention. When the speech sample is real speech, that is, y _i =0, m ₀ is used to make θ _i smaller than arccosm ₀ , and a small arccosm ₀ can make the real speech gather near the weight vector w ₀ ; when the speech sample is deceitful speech , that is, when y _i =1, m ₁ is used to make θ _i larger than arccosm ₁ , and a relatively large arccosm ₁ can make the deceitful speech far away from the weight vector w ₀ .

2. Design of a lightweight speech deception detection model based on knowledge distillation

此置信度分数代表输入语音属于真实语音还是欺骗语音的概率，进而得到输入语音的分类结果。The overall architecture diagram of the present invention is shown in FIG. 2 , and the residual module therein is shown in FIG. 3 . The teacher model TeacherNet is designed based on the deep residual network ResNet-18, in which the global average pooling layer is replaced with an attention temporal pooling layer. The input of the teacher model is the extracted 60-dimensional linear frequency cepstral coefficient LFCC feature, and the output of the fully connected layer is used as the embedding of the input speech, with a dimension of 256. This embedding is fed into the DOC-Softmax loss function to calculate the confidence score

This confidence score represents the probability that the input speech belongs to the real speech or the deceitful speech, and then the classification result of the input speech is obtained.

来辅助硬标签训练学生模型StudentNet，此软标签带有TeacherNet归纳推理的大量信息。学生模型的损失函数包括两部分，一个是基于分散损失的一类分类损失L_DOCS，一个是学生模型输出的置信度分数

与教师模型输出的置信度分数

的均方误差损失L_MSE，具体公式如下：The model architecture of the student model StudentNet is basically the same as that of the teacher model, except that three residual modules are removed. Soft labels predicted with the teacher model TeacherNet

To assist in training the student model StudentNet with hard labels, this soft label carries a wealth of information for inductive reasoning from TeacherNet. The loss function of the student model consists of two parts, one is a one-class classification loss L _DOCS based on the dispersion loss, and the other is the confidence score output by the student model

Confidence score with teacher model output

The mean square error loss L _MSE of , the specific formula is as follows:

The total loss function of the student model is to minimize the weighted sum of L _DOCS and L _MSE , and the weight is β. The specific formula is as follows:

The StudentNet trained based on the knowledge distillation framework has better generalization ability than the model obtained by using the exact same model structure, training data and only using the hard label training method. The model parameters of this StudentNet are about 402K, which is about 30 times less than that of TeacherNet. The model size of StudentNet is 1590KB, which is convenient for deployment to mobile or embedded devices. The voice deception detection algorithm proposed by the present invention is a single-system, single-feature algorithm, which improves the accuracy of the algorithm and greatly reduces the inference time of the model without using any data enhancement, feature fusion and model integration methods. extension.

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The recorded technical solutions are modified, or some technical features thereof are equivalently replaced, but these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A lightweight speech deception detection algorithm based on one-class classification, characterized in that, using a knowledge distillation framework, a feature space is learned through a class of classification loss function DOC-Softmax based on dispersion loss, in this feature space , the real speech embedding has a tight boundary, there is a certain distance between the deceitful speech and the real speech, and at the same time, a dispersion loss is introduced in the deception speech feature space to maximize the distance of each deception speech sample to its center, so that the deception speech The speech covers the entire space of the cheat speech. 2. the light-weight speech deception detection algorithm based on one-class classification according to claim 1, is characterized in that, the total loss function L _DOCS of one-class classification loss function DOC-Softmax based on dispersion loss is one-class classification loss L The weighted sum of _OCS and dispersion loss _LD , the weight is λ, and the specific formula is as follows:

in,

is the weight vector,

is the normalized representation of w ₀ , the optimization direction of real speech, α is a scaling factor, and two intervals m ₀ and m ₁ are respectively introduced to define the angle between real speech, spoofed speech and real speech weight vector, i.e.

and

The angle between θ _i , m ₀ , m ₁ ∈[-1,1], m ₀ >m ₁ ; The formula of the one-class classification loss function OC-Softmax is as follows:

Two intervals m ₀ and m ₁ , where m ₀ , m ₁ ∈ [-1,1], m ₀ >m ₁ , are introduced to define the real and spoofed samples, respectively

and

When y _i = 0, m ₀ is used to make θ _i smaller than arccosm ₀ ; when y _{i =} 1, m ₁ is used to make θ _i greater than arccosm ₁ , a small arccosm ₀ makes the target class cluster in weight vector w ₀ , a relatively large arccosm ₁ keeps non-target classes away from the weight vector w ₀ ; The formula for introducing dispersion loss is as follows:

in,

is the embedding vector,

is the standardized representation of x, y _i ∈{0,1} is the label of the ith sample, y _i =0 indicates that the sample is real speech, y _i =1 indicates that the sample is deceptive speech, and N is a batch of , M is the number of cheat samples in a batch, ε is a small constant to avoid the case where the denominator is 0,

For the center of the spoofed samples in each batch, the dispersion loss _LD is to maximize the spoofed speech samples

The distance from their center μ makes the deception speech cover the entire deception area as much as possible. 3 . The lightweight voice spoofing detection algorithm based on one-class classification according to claim 2 , wherein α=20, m ₀ =0.9, m ₁ =0.2. 4 . 4. The lightweight voice deception detection algorithm based on one-class classification according to claim 2, wherein when the voice sample is real voice, that is, when y _i =0, m ₀ is used to make θ _i less than arccosm ₀ , a small arccosm ₀ makes the real speech cluster around the weight vector w ₀ ; when the speech sample is deceitful speech, that is, when y _i =1, m ₁ is used to make θ _i larger than arccosm ₁ , a relatively large arccosm ₁ Move the deceptive speech away from the weight vector w ₀ . 5. The lightweight voice deception detection algorithm based on one-class classification according to claim 1, wherein the teacher model adopts the network structure based on the deep residual network ResNet-18, and uses attention pooling to replace the global average pooling. 6 . The light-weight speech deception detection algorithm based on one-class classification according to claim 1 , wherein the teacher model takes the extracted LFCC feature as an input, and takes the result of the fully connected layer output as the embedding of the input speech. 7 . 7. The lightweight voice deception detection algorithm based on one-class classification according to claim 1, wherein the embedding is sent into the DOC-Softmax loss function to calculate the confidence score

This confidence score represents the probability that the input speech belongs to the real speech or the deceitful speech, and then the classification result of the input speech is obtained. 8. the light-weight speech deception detection algorithm based on one-class classification according to claim 1, is characterized in that, the model architecture of student model is basically consistent with teacher model, just removed 3 residual error modules, with teacher model The predicted soft labels are used to assist the hard labels to train the student model. The loss function of the student model consists of two parts, one is the one-class classification loss L _DOCS based on the dispersion loss, and the other is the confidence score output by the student model.

Confidence score with teacher model output

The mean square error loss L _MSE of , the specific formula is as follows:

The total loss function of the student model is to minimize the weighted sum of L _DOCS and L _MSE , and the weight is β. The specific formula is as follows:

Images