CN118427755A - Training method, deep counterfeiting identification method and electronic equipment - Google Patents

Abstract

The present application provides a training method, a deep fake identification method and an electronic device, which obtains picture sample data, audio sample data and video sample data when obtaining a sample data set, and trains based on the picture sample data, audio sample data and video sample data to obtain a target deep fake identification model. The target deep fake identification model obtained can perform authenticity identification on picture data, audio data and video data, and realizes the cross-modal comprehensive identification function through one identification model.

Description

A training method, a deep fake identification method and an electronic device

Technical Field

The present application relates to the field of digital content security technology, and in particular to a training method, a deep fake identification method and an electronic device.

Background technique

With the development of artificial intelligence technology, deep fake technology can generate realistic images, audio and video content, which poses a threat to social order and personal privacy security to a certain extent. Therefore, studying an effective deep fake identification method is of great significance to ensure the security of digital content. However, current deep fake identification technology can usually only identify content of a single modality, for example, it can only identify image content or audio content.

Summary of the invention

The purpose of the embodiments of the present application is to provide a training method, a deep fake identification method and an electronic device to solve the above-mentioned technical problems.

On the one hand, a method for training a deep fake identification model is provided, the method comprising:

Obtain a sample data set; the sample data set includes a plurality of sample data, each of the sample data is marked with label information for indicating whether the sample data is deep fake data, and the sample data set includes picture sample data, audio sample data, and video sample data;

Preprocessing each of the sample data to obtain preprocessed sample data;

Performing feature extraction on each of the preprocessed sample data to obtain sample data features corresponding to each of the sample data;

Training is performed based on the sample data features and a preset initial deep fake identification model to obtain a target deep fake identification model.

In one embodiment, the preprocessing of each sample data to obtain preprocessed sample data includes:

Perform image normalization processing on each of the picture sample data to obtain preprocessed picture sample data, perform audio noise reduction processing on each of the audio sample data to obtain preprocessed audio sample data, and perform key frame extraction on each of the video sample data to obtain preprocessed video sample data.

In one embodiment, the extracting features of each of the preprocessed sample data to obtain sample data features corresponding to each of the sample data includes:

Feature extraction is performed on the preprocessed image sample data to obtain corresponding sample visual features, and feature extraction is performed on the preprocessed audio sample data to obtain corresponding sample auditory features, and feature extraction is performed on the preprocessed video sample data to obtain corresponding features combined by the sample visual features and the sample auditory features; the sample visual features include at least one of color features, texture features and shape features, and the sample auditory features include at least one of pitch features, sound rhythm features and sound intensity features.

In one embodiment, the loss function of the initial deep fake identification model is: in, represents the model prediction value of the ith sample data, _yi represents the true label value of the ith sample data, N represents the total number of sample data, and L represents the loss value.

On the other hand, a deep fake identification method is provided, comprising:

Obtaining data to be identified;

Preprocessing the data to be identified to obtain preprocessed data to be identified;

Extracting features from the pre-processed data to be identified to obtain features of the data to be identified;

Inputting the features of the data to be identified into any of the above-mentioned target deep fake identification models to obtain a deep fake score of the data to be identified;

The deep fake score is compared with a preset deep fake score threshold, and an identification result of the data to be identified is obtained based on the comparison result.

In one embodiment, the data to be identified includes video data to be identified, and the preprocessing of the data to be identified to obtain preprocessed data to be identified includes:

Extracting key frames from the video data to be identified to obtain video key frame data;

The image normalization processing and the audio noise reduction processing are performed on each of the video key frame data to obtain the pre-processed data to be identified.

In one embodiment, comparing the deep fake score with a preset deep fake score threshold, and obtaining an identification result of the data to be identified based on the comparison result, includes:

When the deep fake score is less than a preset first deep fake score threshold, determining that the data to be authenticated is real data;

When the deep fake score is greater than a preset second deep fake score threshold, the data to be identified is determined to be deep fake data.

In one embodiment, the second deep fake score threshold is greater than the first deep fake score threshold, and comparing the deep fake score with a preset deep fake score threshold, and obtaining an identification result of the data to be identified according to the comparison result, includes:

When the deep fake score is greater than the first deep fake score threshold and less than the second deep fake score threshold, the data to be identified is determined to be data whose authenticity is pending.

In one embodiment, before comparing the deep fake score with a preset deep fake score threshold, the method includes:

Determining a source scenario of the data to be authenticated;

According to the preset correspondence between the source scene and the deep fake score threshold, the deep fake score threshold corresponding to the source scene of the data to be identified is determined.

On the other hand, an electronic device is provided, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above methods.

Through the training method, deep fake identification method and electronic device provided by the present application, image sample data, audio sample data and video sample data are obtained when obtaining a sample data set, and a target deep fake identification model is obtained through training based on the image sample data, audio sample data and video sample data. Therefore, the obtained target deep fake identification model can perform authenticity identification on image data, can also perform authenticity identification on audio data, and can also perform authenticity identification on video data, thereby realizing a cross-modal comprehensive identification function through one identification model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a process of training a deep fake identification model provided in Example 1 of the present application;

FIG2 is a schematic diagram of a process flow of a deep fake identification method provided in Example 1 of the present application;

FIG3 is a schematic diagram of the structure of an electronic device provided in Embodiment 2 of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

Embodiment 1:

The present application embodiment provides a method for training a deep fake identification model, as shown in FIG1 , comprising the following steps:

S11: Obtain a sample data set; the sample data set includes multiple sample data, each sample data is marked with label information for indicating whether the sample data is deep fake data, and the sample data set includes picture sample data, audio sample data and video sample data.

S12: Preprocess each sample data to obtain preprocessed sample data.

S13: Extract features from each preprocessed sample data to obtain sample data features corresponding to each sample data.

S14: Training is performed based on the characteristics of each sample data and a preset initial deep fake identification model to obtain a target deep fake identification model.

The following is a detailed introduction to the contents of each of the above steps.

In step S11, a large amount of real sample data and deep fake sample data can be collected from a variety of different source scenes. These sample data can cover different characters, scenes, expressions and actions to ensure diversity and representativeness. In addition, for deep fake sample data, it should include fake content of different qualities, from low resolution to high resolution, from simple editing to complex synthesis, to simulate various situations in the real world.

Exemplarily, the sample data set in the embodiment of the present application may include N real sample data and M deep fake sample data. Specifically, the sample data may be collected through the following ways:

Public datasets: Leverage existing deep fake and real content datasets such as CelebA, DeepFakeDetection Challenge, etc.

Social Media: Obtain publicly shared real images, real audio, real videos, and content marked as deepfakes from social media platforms.

Partners: Collaborate with film production companies, news organizations, etc. to obtain high-quality real and fake sample data.

To ensure the diversity of the collected sample data sets, the collected sample data should cover different:

Scenes: indoor, outdoor, nature, city, etc.

Expressions: joy, anger, sorrow, happiness, etc.

Lighting conditions: strong light, weak light, shadow, etc.

In step S12, different preprocessing strategies should be used for preprocessing different types of sample data, including:

Perform image normalization processing on each image sample data to obtain preprocessed image sample data, perform audio noise reduction processing on each audio sample data to obtain preprocessed audio sample data, and perform key frame extraction on each video sample data to obtain preprocessed video sample data.

It should be noted that audio sample data usually contains both image content and audio content. Therefore, after extracting key frames from video sample data, image normalization processing and audio noise reduction processing can be performed on the image content and audio content corresponding to the key frames, respectively, to obtain preprocessed video sample data. Video frame extraction can be achieved through a key frame detection algorithm, which determines key frames based on motion vectors V and color changes C.

In the embodiment of the present application, the same image normalization processing strategy is used to perform image normalization processing on the image content in each picture sample data and video sample data, including but not limited to adjusting the brightness, contrast, saturation, size and color of the image, so that the image content in each picture sample data and video sample data meets the input standard of the initial deep fake identification model.

Exemplarily, in an embodiment of the present application, image brightness normalization processing can be performed based on the formula Inorm=(I-μI)/σI, wherein I represents the original image, that is, the original brightness value of each pixel in the image content in the picture sample data or the video sample data, μI represents the average brightness of the original image, and σI represents the standard deviation of the brightness of the original image. Based on this formula, the brightness of each image can be adjusted. It can be understood that the contrast and saturation of each image can be adjusted with reference to this formula.

When resizing images, each image can be adjusted to a uniform resolution, such as 224x224 pixels. When adjusting the color of an image, a standardization method can be used to make the RGB values of the image distributed within a range with a mean of 0 and a standard deviation of 1.

In the embodiment of the present application, when performing audio noise reduction processing on the audio content corresponding to the audio sample data or the video sample data, a signal processing technology can be used to remove background noise and improve the clarity of the audio. The audio noise reduction can use an adaptive filter, and its output y[n] can be expressed as: x[n] represents the original audio signal, which is a sampling point in the time series, where n represents the sampling sequence number, and represents the sampling point in the time series. h[m] represents the impulse response of the filter, which is a weight coefficient vector, where m represents the sequence number in the impulse response, and is used to combine with x[nm] to produce the filtered output. y[n] represents the output signal after filtering, and M represents the length of the impulse response, that is, the order or size of the filter.

It should be noted that performing audio noise reduction processing on each audio sample data also includes: after removing the background noise, converting the time domain audio signal into a frequency domain audio signal, for example, using FFT (Fourier Transform) or MFCC (Mel Spectrum Coefficient) to convert the time domain audio signal into a frequency domain audio signal.

Step S13 in the embodiment of the present application may include:

Feature extraction is performed on the preprocessed image sample data to obtain corresponding sample visual features, feature extraction is performed on the preprocessed audio sample data to obtain corresponding sample auditory features, and feature extraction is performed on the preprocessed video sample data to obtain corresponding features combined by the sample visual features and the sample auditory features; the sample visual features include at least one of color features, texture features and shape features, and the sample auditory features include at least one of pitch features, sound rhythm features and sound intensity features.

It should be noted that in step S13, a deep learning model can be used to extract features. Specifically, a CNN (convolutional neural network) can be used to extract features from preprocessed image sample data or preprocessed video sample data to obtain corresponding visual features. RNN (recurrent neural network) or LSTM (long short-term memory network) can be used to extract features from preprocessed audio sample data to obtain corresponding auditory features.

The loss function of the initial deep fake identification model in the embodiment of the present application is: in, represents the model prediction value of the i-th sample data, _yi represents the true label value of the i-th sample data, N represents the total number of sample data, and L represents the loss value. The goal of model training is to minimize the loss function.

It should be noted that during the model training phase, the following deep learning framework can be used:

VAE (Variational Autoencoder) is used to learn a low-dimensional representation of real content and use it to reconstruct real content as a benchmark for identifying forged content.

The GAN generative adversarial network (GAN) is used to generate samples of forged content to train the discriminator and improve its ability to identify forged content.

The Transformer model can be used to handle long-range dependencies in video and audio data, such as the correlation between consecutive frames in video or long-term spectral features in audio.

VAE is a generative model that compresses input data into a low-dimensional latent representation through an encoder and then reconstructs the input data through a decoder. The key feature of VAE is that it uses variational inference to optimize the distribution of the latent space to make it close to the prior distribution. The objective function of the VAE model can be expressed as: min _q,p E _q(z|x) [logp(x|z)]-KL(q(z|x)||p(z)), where q is the encoder distribution, p is the decoder distribution, and KL is the KL divergence. The architecture of VAE can be expressed as:

1) Encoder _qφ (z|x)=N( _μφ (x), _σφ (x) ^2I ): The input data x is mapped to a distribution in the latent space z, where _μφ and _σφ are parameterized functions of the mean and log-variance of the latent representation, respectively.

2) Decoder p _θ (x|z) = N(μ _θ (z), σ _θ (z) ² I: The latent representation z is mapped back to the data space, where μ _θ and σ _θ are parameterized functions of the mean and log-variance of the reconstructed data.

3) The loss function consists of two parts: reconstruction loss and KL divergence. The former measures the difference between the reconstructed data and the original data, and the latter measures the difference between the distribution of the potential representation and the prior distribution.

GAN consists of two parts: the generator and the discriminator. The training process of the GAN model includes the adversarial process of the generator G and the discriminator D. They compete with each other during the training process, and their objective function is:

1) Generator G: The goal is to generate data G(z) that is as realistic as possible, where z is the noise sampled from the prior distribution pz.

2) Discriminator D: The goal is to distinguish between real data x and fake data G(z) generated by the generator.

3) Adversarial process: The generator and discriminator are continuously updated through the adversarial process. The generator learns to generate more and more realistic data, while the discriminator learns to classify true and false data more accurately.

4) Loss Function: The loss function of the discriminator is the classification accuracy of real and generated data, while the loss function of the generator is its ability to deceive the discriminator.

The Transformer model is a model based on the self-attention mechanism, which can effectively handle long-distance dependency problems in sequence data:

1) Self-attention mechanism: allows the model to directly compute representations between different positions in the sequence without having to process them step by step like a recurrent neural network.

2) Multi-head attention: Multiple self-attention mechanisms are executed in parallel, with each head focusing on a different part of the sequence.

3) Positional encoding: Since the Transformer lacks a recurrent or convolutional structure, it uses positional encoding to provide the position information of the words in the sequence.

4) Output representation: Generate the final output through linear layers and non-linear activation functions.

In the embodiment of the present application, the initial deep fake identification model is trained by a sample data set annotated with label information, so that it can learn the subtle differences between real content and deep fake content. The following training strategies can be used during the training process:

Data enhancement: Increase the diversity of samples by rotating, scaling, cropping, etc.

Loss function: Use binary cross entropy loss function to distinguish between real and fake content;

Optimizer: Use the Adam optimizer, whose adaptive learning rate helps fast convergence.

During the model validation and optimization phase, the following methods can be used to divide the data set:

Stratified sampling: Ensure that the distribution of samples in the training set, validation set, and test set is consistent.

Based on the evaluation results, the model parameters and structure can be adjusted to optimize performance, including:

Vary the depth and width of the network;

Adjust learning rate and optimizer;

Use a different initialization method.

During the hyperparameter tuning phase, the following methods can be used to optimize hyperparameters:

Grid search: traverse different hyperparameter combinations to find the optimal solution.

Random Search: Randomly select hyperparameter combinations to avoid getting stuck in local optima.

In the embodiment of the present application, after obtaining the target deep fake identification model, deep fake identification can be performed based on the target deep fake identification model. Therefore, the embodiment of the present application also provides a deep fake identification method, as shown in FIG2, including:

S21: Obtain the data to be authenticated.

S22: Preprocessing the data to be identified to obtain preprocessed data to be identified.

S23: Extracting features from the pre-processed data to be identified to obtain features of the data to be identified.

S24: Input the features of the data to be identified into the target deep fake identification model to obtain the deep fake score of the data to be identified.

S25: Compare the deep fake score with a preset deep fake score threshold, and obtain an identification result of the data to be identified based on the comparison result.

It can be understood that the data to be identified in step S21 can be any one of video data to be identified, picture data to be identified, and audio data to be identified.

When the data to be identified is video data to be identified, step S22 includes:

Extract key frames from the video data to be identified to obtain video key frame data;

The image normalization processing and the audio noise reduction processing are performed on the key frame data of each video to obtain the preprocessed data to be identified.

The process of performing image normalization and audio noise reduction on the video key frame data here is the same as the process of preprocessing the video sample data mentioned above, and will not be repeated here. Similarly, the process of extracting features from the preprocessed data to be identified to obtain features of the data to be identified here is the same as the process of extracting features from the preprocessed sample data to obtain the corresponding sample data features, and will not be repeated here.

The target deep fake identification model in step S24 is a target deep fake identification model trained using the above-mentioned deep fake identification model training method.

Step S25 may specifically include the following contents:

When the deep fake score is less than a preset first deep fake score threshold, the data to be identified is determined to be real data; when the deep fake score is greater than a preset second deep fake score threshold, the data to be identified is determined to be deep fake data.

The deep fake score output by the target deep fake identification model represents the probability that the data to be identified is deep fake data. The first deep fake score threshold and the second deep fake score threshold can be flexibly set by the developer. Specifically, the optimal threshold can be selected by analyzing the ROC curve.

In an optional implementation, the first deep fake score threshold is equal to the second deep fake score threshold. For example, its value can be 0.5. When the deep fake score output by the target deep fake identification model is less than 0.5, the data to be identified is determined to be real data. When the deep fake score is greater than 0.5, the data to be identified is determined to be deep fake data. When the deep fake score output by the target deep fake identification model is equal to 0.5, the data to be identified can be determined to be real data, or it can also be determined to be deep fake data.

In another optional embodiment, the first deep fake score threshold is less than the second deep fake score threshold, and when the deep fake output by the target deep fake identification model is greater than or equal to the first deep fake score threshold, and less than or equal to the second deep fake score threshold, the data to be identified is determined to be data whose authenticity is pending. When the data to be identified is data whose authenticity is pending, the data to be identified can be further identified manually. Exemplarily, the first deep fake score threshold can be T-δ, and the second deep fake score threshold can be T+δ. When the monitored deep fake score is between [T-δ, T+δ], the data to be identified is determined to be data whose authenticity is pending.

In the embodiment of the present application, after obtaining the identification result of the data to be identified, the identification result can be displayed to the user. For example, different colors can be used to prompt the user. Specifically, the real data can be presented in green, the deep fake data can be presented in red, and the authenticity of the data to be determined can be presented in yellow. It is understandable that in some other embodiments, the user can also be prompted in the form of other annotation labels.

It should be noted that the identification requirements for identification data may be different in different application scenarios. For example, in some application scenarios, it may be necessary to ignore the identification accuracy and identify data suspected of deep forgery as much as possible. In other application scenarios, it may be necessary to identify real data as much as possible. Therefore, in view of the different identification requirements in different application scenarios, the following steps may be included before step S25 in some embodiments:

Determine the source scenario of the data to be identified;

According to the preset correspondence between the source scene and the deep fake score threshold, the deep fake score threshold corresponding to the source scene of the data to be identified is determined.

It is understandable that the source scenario here is essentially the application scenario mentioned in the above content. In the embodiment of the present application, it is necessary to set the corresponding deep fake score threshold for different source scenarios in advance. The deep fake score threshold here includes a first deep fake score threshold and a second deep fake score threshold.

When setting the corresponding deep fake score threshold for a certain source scenario, it can be determined according to the identification requirements of the data to be identified in the source scenario. For example, when it is necessary to identify data suspected of deep fakes as much as possible, the second deep fake score threshold can be set lower. When it is necessary to identify data that may be real as much as possible, the first deep fake score threshold can be set higher. However, no matter how it is set, it is necessary to ensure that the first deep fake score threshold is less than or equal to the second deep fake score threshold.

During the authentication process, when the deep fake score of the data to be authenticated is less than the first deep fake score threshold corresponding to the source scene of the data to be authenticated, the data to be authenticated is determined to be real data. When the deep fake score of the data to be authenticated is greater than the second deep fake score threshold corresponding to the source scene of the data to be authenticated, the data to be authenticated is determined to be deep fake data.

The source scenarios in the embodiments of the present application include but are not limited to security monitoring system scenarios, social media platform scenarios, etc.

It is understandable that the security monitoring system or social media platform can implement the present method by calling the identification service through an API interface, or the electronic device executing the present method can obtain the data to be identified from the security monitoring system or social media platform.

In order to facilitate external systems or platforms to call the identification service, plug-ins can be developed to enable other external systems or platforms to use the identification function in the form of plug-ins.

The deep fake identification model training method and deep fake identification method provided in the embodiments of the present application enhance the generalization ability and robustness of the model through multimodal feature extraction and comprehensive analysis, and realize cross-modal comprehensive identification function through one identification model.

It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

Embodiment 2:

Based on the same inventive concept, an embodiment of the present application provides an electronic device, as shown in FIG3 , including a processor 301 and a memory 302. The memory 302 stores a computer program, the processor 301 and the memory 302 communicate via a communication bus, and the processor 301 executes the computer program to implement the steps of the method described above, which will not be described in detail here.

It should be noted that the electronic device in the embodiment of the present application can be a terminal or a server. Among them, the terminal can be a smart phone, a tablet computer, a laptop, a personal computer (PC), a smart home, a wearable electronic device, a VR/AR device, a car computer, etc. The server can be an intercommunication server or a background server between multiple heterogeneous systems, or an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and basic cloud computing services such as big data and artificial intelligence platforms, etc.

The processor 301 may be an integrated circuit chip with signal processing capabilities. The processor 301 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. It may implement or execute various methods, steps and logic block diagrams disclosed in the embodiments of the present application. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

The memory 302 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable read-only memory (EPROM), electrically erasable read-only memory (EEPROM), etc.

This embodiment also provides a computer-readable storage medium, such as a floppy disk, a CD, a hard disk, a flash memory, a USB flash disk, an SD card, an MMC card, etc., in which one or more programs for implementing the above steps are stored. These one or more programs can be executed by one or more processors to implement the steps of the method in the above embodiment 1, which will not be repeated here.

It should be noted that the diagram provided in the present embodiment only illustrates the basic concept of the present invention in a schematic manner, so the diagram only shows the components related to the present invention rather than drawing according to the number, shape and size of the components during actual implementation. The type, quantity and ratio of each component during actual implementation can be a random change, and the component layout type may also be more complicated. The structure, ratio, size, etc. illustrated in the drawings of the present specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions that the present invention can implement, so they have no technical substantive significance. Any modification of the structure, change of the proportional relationship or adjustment of the size should still fall within the scope of the technical content disclosed by the present invention without affecting the effect that the present invention can produce and the purpose that can be achieved. At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" quoted in this specification are only for the convenience of narration, and are not used to limit the scope of the present invention. The change or adjustment of its relative relationship should also be regarded as the scope of the present invention without substantial change of the technical content.

The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims (10)

1. A training method for a deep forgery identification model, comprising:

acquiring a sample data set; the sample data set comprises a plurality of sample data, each sample data is marked with label information for indicating whether the sample data is depth counterfeit data, and the sample data set comprises picture sample data, audio sample data and video sample data;

preprocessing each sample data to obtain preprocessed sample data;

extracting characteristics of each piece of preprocessed sample data to obtain sample data characteristics corresponding to each piece of sample data;

training based on the characteristics of the sample data and a preset initial depth counterfeit identification model to obtain a target depth counterfeit identification model.

2. The method of training a deep forgery identification model of claim 1, wherein preprocessing each of the sample data to obtain preprocessed sample data includes:

Performing image normalization processing on each piece of picture sample data to obtain preprocessed picture sample data, performing audio noise reduction processing on each piece of audio sample data to obtain preprocessed audio sample data, and performing key frame extraction on each piece of video sample data to obtain preprocessed video sample data.

3. The method for training a deep forgery identification model of claim 2, wherein the performing feature extraction on each of the preprocessed sample data to obtain a sample data feature corresponding to each of the sample data includes:

Extracting features from the preprocessed picture sample data to obtain corresponding sample visual features, extracting features from the preprocessed audio sample data to obtain corresponding sample auditory features, and extracting features from the preprocessed video sample data to obtain corresponding features combined by the sample visual features and the sample auditory features; the sample visual features include at least one of color features, texture features, and shape features, and the sample auditory features include at least one of pitch features, sound tempo features, and sound intensity features.

4. The training method of a depth counterfeit authentication model of claim 1, wherein said initial depth counterfeit authentication model has a loss function of: wherein, Representing a model predicted value of the ith sample data, y _i representing a true tag value of the ith sample data, N representing a total number of the sample data, and L representing a loss value.

5. A method of depth counterfeit authentication, comprising:

Acquiring data to be authenticated;

Preprocessing the data to be authenticated to obtain preprocessed data to be authenticated;

Extracting the characteristics of the data to be authenticated by preprocessing to obtain the characteristics of the data to be authenticated;

Inputting the data to be authenticated characteristics into the target depth forgery authentication model according to any one of claims 1-4 to obtain a depth forgery score of the data to be authenticated;

And comparing the depth falsification score with a preset depth falsification score threshold value, and obtaining an identification result of the data to be identified according to the comparison result.

6. The depth forgery identification method of claim 5, wherein the data to be identified includes video data to be identified, and the preprocessing of the data to be identified results in preprocessing the data to be identified, including:

extracting key frames of the video data to be identified to obtain video key frame data;

And carrying out image normalization processing and audio noise reduction processing on each video key frame data to obtain the preprocessing data to be authenticated of the data to be authenticated.

7. The method for identifying deep forgery as claimed in claim 5, wherein comparing the deep forgery score with a preset deep forgery score threshold value, and obtaining the identification result of the data to be identified according to the comparison result, comprises:

When the depth falsification score is smaller than a preset first depth falsification score threshold value, judging the data to be authenticated as real data;

And when the depth falsification score is larger than a preset second depth falsification score threshold value, judging that the data to be authenticated is the depth falsification data.

8. The depth-based forgery identification method of claim 7, wherein the second depth-based forgery score threshold is greater than the first depth-based forgery score threshold, the comparing the depth-based forgery score with a preset depth-based forgery score threshold, and obtaining the identification result of the data to be identified according to the comparison result, includes:

and when the depth forgery is larger than the first depth forgery score threshold value and smaller than the second depth forgery score threshold value, judging that the data to be authenticated is true and false data to be authenticated.

9. The depth-forgery-identification method of claim 5, wherein prior to said comparing the depth-forgery score to a preset depth-forgery-score threshold, the method comprises:

determining a source scene of the data to be authenticated;

and determining a depth falsification score threshold corresponding to the source scene of the data to be identified according to the corresponding relation between the preset source scene and the depth falsification score threshold.

10. An electronic device comprising a processor and a memory, the memory having stored therein a computer program, the processor executing the computer program to implement the method of any of claims 1-9.