Method and system for discriminating deeply-forged video

Abstract

The invention discloses a method and a system for discriminating a deeply-forged video, and mainly relates to the technical field of deeply-forged video detection. Comprising the following steps: acquiring a real image, and applying spatial disturbance to the real image to generate a corresponding forged image; based on the real image and the forged image, adopting an instance re-grading strategy to generate composite images with different forgery degrees; performing low-frequency enhancement processing on the synthesized image to obtain a frequency enhanced image; respectively extracting spatial domain features and frequency domain features of the frequency enhanced image; fusing the extracted spatial domain features with the frequency domain features to obtain fused discrimination features; and based on the discrimination features, carrying out authenticity classification and counterfeit intensity grade determination on the input video frame. The beneficial effects of the invention are that the method can effectively alleviate the overfitting problem caused by a binary classification normal form, and constructs a space-frequency domain feature fusion method on the basis, thereby improving the adaptability of a model to diversified counterfeit types and the robustness of detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep fake video detection, and particularly relates to a deep fake video discrimination method and system based on fake synthetic additional feature maps. BACKGROUND

[0002] With the rapid development of generative artificial intelligence technologies such as generative adversarial networks and diffusion models, deep fake technology can synthesize images and videos with extremely realistic visual effects and extremely strong deception. Although this technology has potential application value, its misuse has seriously threatened public trust, personal privacy and social security, making high-robustness deep fake detection technology an urgent need.

[0003] The current mainstream deep fake detection methods generally face two major challenges: first, the problem of insufficient generalization ability. The fake technology itself is rapidly evolving, and the constantly emerging new fake means make the detection model significantly decline in performance when facing “unknown” or “unseen” fake types, making it difficult to maintain robust discrimination ability. Second, most existing methods rely on a simple binary (true / false) classification learning paradigm. This paradigm has inherent defects: it forcibly discretizes the continuous and gradual fake strength (such as the degree of fusion and the degree of obviousness of editing traces) into a black-or-white label, leading to “information collapse”. Specifically, the model loses information about the fine-grained differences in fake degree during the learning process, which not only limits its discrimination accuracy, but also easily leads to overfitting of the model on training data, further damaging its generalization performance in real complex scenarios.

[0004] To improve generalization, recent research has proposed a detection idea based on data synthesis, which constructs training data by actively simulating edge artifacts, texture inconsistencies and other traces generated during the fake process. However, this method still has obvious limitations: first, its core paradigm still does not break out of the binary classification framework, and it does not effectively utilize the continuous properties of fake strength; second, its feature extraction is excessively focused on the spatial domain (such as image pixel-level texture and color), and it generally ignores the analysis of frequency domain features. Studies have shown that fake operations in the frequency domain (especially the low-frequency band) leave more consistent “fingerprints” or traces across methods. The lack of modeling of frequency domain features makes it difficult for the model to capture the common rules behind different fake methods, thereby limiting the further improvement of generalization ability.

[0005] Therefore, there is an urgent need for a deep fake video discrimination method and system based on fake synthetic additional feature maps to solve the above problems. SUMMARY

[0006] The application aims to provide a deep fake video discrimination method and system based on fake synthetic additional feature maps, which can effectively alleviate the overfitting problem caused by the binary classification paradigm, and on this basis, construct a space-frequency feature fusion method, thereby improving the adaptability of the model to diversified fake types and the robustness of detection.

[0007] To achieve the above-mentioned purpose, the application is implemented through the following technical solutions: In one aspect, the application provides a deep fake video discrimination method, comprising the following steps: Step S1: obtaining a real image and applying spatial disturbance to the real image to generate a corresponding fake image; Step S2: based on the real image and the fake image, generating synthetic images with different fake degrees by using an instance re-hierarchical strategy, wherein the instance re-hierarchical strategy controls the mixing ratio Quantifying the fake degree as a continuous supervision signal; Step S3: performing low-frequency enhancement processing on the synthetic images to enhance the low-frequency structure components in the images, obtaining frequency-enhanced images; Step S4: extracting the spatial domain features and frequency domain features of the frequency-enhanced images respectively; Step S5: fusing the extracted spatial domain features and frequency domain features to obtain fused discrimination features; Step S6: based on the discrimination features, classifying the input video frames as real or fake and determining the fake intensity level.

[0008] Preferably, in step S2, the instance re-hierarchical strategy for generating synthetic images with different fake degrees is specifically: The real image is mixed with the fake image generated by spatial disturbance according to a preset mixing ratio to generate a synthetic image , which is represented as: ; wherein, represents pixel-by-pixel multiplication, and the binary mask defines the mixed transition area; the mixing ratio The greater the value, the higher the fake signal intensity in the synthetic image ; A joint loss function is constructed to supervise the model, and the joint loss function includes a binary cross-entropy loss for distinguishing real and fake samples and a classification cross-entropy loss for learning the fine-grained differences between different fake degrees, which is defined as: ; wherein, and are loss weights, and satisfy .

[0009] Preferably, in step S3, the low-frequency enhancement processing on the synthetic image is specifically: wavelet decomposition is performed on the synthetic image to decouple out frequency components including low-frequency component , horizontal high-frequency component , vertical high-frequency component , and diagonal high-frequency component ; wherein, the wavelet decomposition is implemented through four directional convolution filters , each filter is used to capture the frequency characteristics in a specific direction, and the decomposition process is represented as: ; wherein, represents convolution operation; using the low-frequency component , its spatial resolution is restored through transposed convolution operation , and pixel-level average operation is performed with the original synthetic image to generate a frequency-enhanced image , represented as: ; wherein, represents pixel-level average operation.

[0010] Preferably, in step S4, the spatial domain feature of the frequency-enhanced image is specifically extracted as: a pre-trained convolutional neural network is used as a spatial backbone network to encode the frequency-enhanced image , and extract high-level spatial feature representation, i.e. spatial domain feature , the process is represented as: ; wherein, represents the encoder of the spatial backbone network.

[0011] Preferably, in step S4, the frequency domain feature of the frequency-enhanced image is specifically extracted as: two-level wavelet transform decomposition is performed on the spatial domain feature to obtain a set of directional and scale frequency domain components; for each channel of the spatial domain feature , a set of wave bands after two-level wavelet decomposition is defined as: ; The band set of all channels of the spatial domain feature is spliced to obtain an initial frequency domain feature , denoted as: ; Wherein, is the total number of channels of the spatial domain feature ; The initial frequency domain feature is input into a frequency domain encoder composed of multiple layers of convolution to obtain a discriminative frequency domain feature ; The frequency domain encoder includes at least three layers of structure, each of which sequentially performs convolution, batch normalization, and activation function processing, and contains a squeeze-and-excitation module to model the dependency relationship between channels.

[0012] Preferably, the processing process of the frequency domain encoder is specifically: For the input , sequentially pass through the first encoding stage, the second encoding stage and the third encoding stage to obtain the discriminative frequency domain feature , the process is represented as: ; ; ; Wherein, , indicates a squeeze-and-excitation module, indicates a GELU activation function, , , indicates a batch normalization operation, indicates a convolution operation with a convolution kernel size of , indicates a convolution operation with a convolution kernel size of and a step size of 2.

[0013] Preferably, in step S5, the extracted spatial domain feature and the frequency domain feature are fused, specifically: The spatial domain feature and the frequency domain feature are respectively mapped to a unified embedding dimension through a projection layer; The mapped features are spliced into a token sequence ; The token sequence is input into a multi-head self-attention module based on Transformer for cross-domain information interaction; extracting the enhanced tokens corresponding to the spatial feature part from the sequence output by the multi-head self-attention module, and passing them through an output projection layer with adjustable weights with the original spatial domain features performing residual connection to obtain fused discriminative features, and the process is represented as: ; wherein, is the fused discriminative feature, is the output projection layer, represents the first enhanced token extracted from the token sequence.

[0014] Preferably, in step S6, the output of the model includes a binary true-false classification result and a quantitative value representing the level of forgery strength, and the quantitative value is associated with the mixing ratio used in the instance re-ranking strategy .

[0015] In another aspect, the present application provides a discriminative system for deepfake videos, which is used to implement the discriminative method for deepfake videos as described above, and comprises: an image synthesis module, which is used to obtain real images, apply spatial disturbance to generate fake images, and generate synthetic images with different degrees of forgery based on an instance re-ranking strategy; a low-frequency enhancement module, which is used to perform wavelet decomposition on the synthetic images, extract and enhance the low-frequency structural components thereof, and generate frequency-enhanced images; a feature extraction module, which is used to extract spatial domain features and frequency domain features of the frequency-enhanced images, respectively; a feature fusion module, which is used to fuse the spatial domain features and the frequency domain features to obtain discriminative features; a classification and determination module, which is used to output a true-false classification result and a level of forgery strength based on the discriminative features.

[0016] Compared with the prior art, the present application has the following advantages: 1. By means of the innovative instance re-ranking strategy, the degree of forgery is quantified as a continuous supervision signal, so that the model can learn the fine-grained differences in the level of forgery strength, rather than simply performing true-false classification. This effectively avoids the problems of degradation of feature representation ability and overfitting caused by information collapse, so that the model can still maintain stable and accurate detection performance when facing unknown or new forgery techniques.

[0017] 2. By introducing a low-frequency enhancement module, the low-frequency components representing the global structure information in the image are actively extracted and strengthened, effectively making up for the shortcomings of existing methods that rely too much on spatial domain texture details. Combined with the subsequent spatial-frequency domain feature fusion network, the local artifact sensitivity in the spatial domain and the cross-method stability in the frequency domain can be utilized simultaneously to achieve a more comprehensive and robust representation of the forgery traces.

[0018] 3. Unlike traditional methods that only output binary results of authenticity, the method of the present application can further output quantitative grades or risk assessment results reflecting the strength of forgery. This not only provides more rich decision-making information, which is helpful for content classification management, but also makes the decision-making process of the model more interpretable.

[0019] 4. The method proposed in the present application, through efficient network architecture design (such as using lightweight backbone network, carefully designed fusion module), on the basis of realizing high-precision detection and strong generalization ability, maintains a low computational complexity and model parameter amount, so as to support real-time or efficient batch processing detection on video stream, and has good actual deployment prospect. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a method flowchart of an embodiment of the present application; Figure 2 is a system structure schematic diagram of an embodiment of the present application; Figure 3 is a schematic diagram of online video platform authenticity identification results and forgery trace detection results of an embodiment of the present application. DETAILED DESCRIPTION

[0021] The present application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. In addition, it should be understood that after reading the content taught by the present application, those skilled in the art can make various modifications or changes to the present application, and these equivalent forms also fall within the scope defined by the present application.

[0022] In the present application, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicate the orientation or positional relationship shown in the drawings, which is only a relationship word determined for the convenience of describing the structural relationship of the components or elements of the present application, and cannot be understood as a limitation on the present application.

[0023] Embodiment: As shown in Figure 1 , the present embodiment provides a method for discriminating deepfake videos, comprising the following steps: Step S1: Obtain a real image, and apply spatial disturbance to the real image to generate a corresponding fake image; Step S2: Based on the real image and the fake image, generate a synthetic image with different degrees of forgery using an instance re-ranking strategy, wherein the instance re-ranking strategy quantifies the degree of forgery as a continuous supervision signal through a controllable mixing ratio Step S3: Perform low-frequency enhancement processing on the synthetic image to enhance the low-frequency structure component in the image, to obtain a frequency-enhanced image; Step S4: Extract spatial domain features and frequency domain features of the frequency-enhanced image, respectively; Step S5: Fuse the extracted spatial domain features and frequency domain features to obtain fused discriminative features; Step S6: Based on the discriminative features, perform authenticity classification and forgery intensity level determination on the input video frame.

[0024] Specifically, The embodiment proposes an instance re-ranking method for fake samples to capture the diversity of the degree of forgery. The re-ranking strategy explicitly incorporates the forgery level as a supervision signal into the model, which helps to reduce information loss and promote the learning of fine-grained discriminative representations. Before the re-ranking mixing operation, the embodiment first applies a series of spatial disturbances to the original real image, aiming to enhance the sensitivity of the model to local structure and texture changes. For input real samples , the corresponding fake image obtained through spatial domain operation can be represented as: ; wherein represents the th spatial transformation.

[0025] IRGS (Instance Re-Ranking Strategy) aims to build a re-ranking framework for quantifying the degree of forgery and promoting the model's learning of multi-level forgery features, thereby improving the generalization ability to different forgery methods. The core innovation lies in converting the degree of forgery into a calculable quantitative index. The mixing ratio can be used to quantify the contribution weight of the original image and the re-ranking mixed image in the synthesis process, and its mathematical expression is: ; wherein represents pixel-wise element multiplication, is a binary mask that defines and transition boundaries. As can be seen from the above formula, the larger the mixing ratio , the more the low-frequency structure component in the output image​ The higher the proportion, the more significant the counterfeit signal strength. This mechanism makes a controllable and quantifiable counterfeit strength index, which can be used to evaluate the detection behavior of the model under the condition of gradually increasing counterfeit strength.

[0026] IRGS is essentially an extension of the binary classification learning paradigm, aiming to introduce more refined counterfeit degree modeling to improve discrimination. To this end, the embodiment designs a joint optimization mechanism, combining the basic binary cross-entropy loss used to distinguish real samples from fake samples with the classification cross-entropy loss which guides the model to learn the fine hierarchical differences between fake samples, defined as follows: where is the real label, is the predicted value. represents the real label of the th sample in the class , and represents the predicted probability that the th sample belongs to class . Therefore, the total loss is defined as: ; where and are the loss weights of and respectively. IRGS converts the counterfeit strength into a supervised signal, guiding the model to learn multi-level counterfeit features and thus improve its discrimination ability.

[0027] Existing counterfeit detection methods usually focus on spatial domain operations, resulting in limited ability to describe the structural features of generated images. To this end, the embodiment proposes a low-frequency enhancement module, aiming to enhance the low-frequency structural components in the counterfeit images generated by the re-ranking mixing operation. Specifically, the LFEM (Low-Frequency Enhancement Module) first applies four directional convolution filters to perform wavelet decomposition, whose expression is as follows: ; where denotes the filter set, . Each captures the frequency features in a specific direction. The symbol denotes convolution operation. Among the decoupled components, the low-frequency component encodes the structural information of the image and is the main domain representing the counterfeit traces. Therefore, the LFEM generates a frequency-enhanced image using , which is calculated as follows: ; where denotes the transpose convolution operation, which recovers the spatial resolution to match the input, while denotes the pixel-wise average operation.

[0028] This embodiment proposes a spatial-frequency feature fusion network that integrates spatial and frequency domain information to enhance the model's sensitivity to fake artifacts.

[0029] To extract rich spatial information, this embodiment uses EfficientNet-B5

[10] as the spatial backbone network. Specifically, for the input image , the spatial feature extractor encodes it into a high-level representation: ; where denotes the encoder, represents the extracted spatial features.

[0030] To complement the spatial features with frequency domain representations, the SFFF (spatial-frequency feature fusion network) designs a dedicated frequency domain stream. The input feature is first decomposed by two levels of wavelet transform to generate multi-directional frequency domain components. For each channel , the waveband set can be defined as: ; Then the output feature is the concatenation of all channels, which can be defined as: ; To encode into discriminative frequency representations, this embodiment designs a convolutional encoder consisting of three stages. Each stage contains convolutional layers, batch normalization, and GELU activation functions, and is enhanced by a squeeze-activation module (SE) to model inter-channel dependencies. Specifically, for the input , the encoding process is defined as: ; where denotes the GELU activation function, denotes the convolution operation with a kernel size of and a stride of 2.

[0031] To simultaneously obtain complementary information from the spatial and temporal domains, this embodiment designs a fusion module based on Transformer. Given the spatial feature and the frequency feature , first project them into a unified embedding space and concatenate them into a token sequence: ; After multi-head self-attention encoding, the enhanced spatial representation is obtained from the first token and fused with the original features in the following way: ; where is a tunable fusion weight hyperparameter. This fusion mechanism enables the spatial representation to be adaptively enriched by frequency-aware cues through cross-domain attention modeling.

[0032] As shown in Figure 2 , the embodiment also provides a deep fake video discrimination system, comprising: An image synthesis module is configured to obtain real images, apply spatial disturbances to generate fake images, and generate synthetic images with different degrees of forgery based on an instance re-ranking strategy; A low-frequency enhancement module is configured to perform wavelet decomposition on the synthetic images, extract and enhance the low-frequency structural components thereof, and generate frequency-enhanced images; A feature extraction module is configured to extract spatial domain features and frequency domain features of the frequency-enhanced images, respectively; A feature fusion module is configured to fuse the spatial domain features and the frequency domain features to obtain discrimination features; A classification and determination module is configured to output a true or false classification result and a fake intensity level based on the discrimination features The embodiment is deployed in the video review background of an online content platform, and is used for deep fake detection and risk assessment of real-time video streams uploaded by users. The FAIR (Frequency Domain Augmented Instance Re-Ranking) method proposed in the present application is integrated into the real-time review process, and the operation process of the system is as follows: 1. Instance re-ranking generation and learning strategy: Collect videos in the online video system that are definitely real as real samples, apply spatial disturbances to generate fake images . The weight of the original image and the fake image is controlled by the mixing ratio , and the degree of forgery is converted into a calculable quantitative index. Then a multi-level supervised model based on the online system is designed to guide the platform's model to learn the fine differences between the fake samples and improve the generalization ability of the model.

[0033] 2. Low-frequency enhancement and feature supplement: Existing detection methods mainly focus on the spatial domain, and are prone to ignore frequency domain forgery traces. The system first performs wavelet decomposition on the fake images Performing wavelet decomposition, decoupling out the low-frequency component to encode the structural information of the image, and then restoring its spatial resolution by transposed convolution to generate a frequency-enhanced image, so that the recognition model of the online platform can capture the overall fake traces of the video.

[0034] 3. Spatial-frequency domain feature fusion network: Based on the online platform, the spatial-frequency domain feature fusion network is deployed, and EfficientNet-B5 is used as the spatial backbone network to extract spatial features , and the frequency domain features are extracted by multi-level wavelet transform and convolutional encoder . Based on and , the system designs a fusion module based on Transformer, which improves the detection accuracy and stability of the system under diversified fake methods.

[0035] 4. System test effect: The video review system based on the FAIR method in this embodiment can not only realize accurate true-false binary classification in real-time detection of the content platform, but also output fine-grained fake strength levels as shown in Figure 3 , realizing multi-level risk assessment of video content. As shown in Table 1, under four commonly used fake detection test sets, the true-false detection performance reaches 96.2%, 99.3%, 84.8%, and 92.5%, respectively, which is an average of more than 4% higher than other detection methods. In addition, the system breaks through the limitations of the binary classification paradigm through IRGS, and the processing speed reaches 4065 frames per second, and the model parameters only need 14.4MB, and the video processing per second reaches 127. It reaches the level of real-time video deep fake detection with high generalization and high precision.

[0036] Table 1 FAIR and other method performance comparison table

[0037] The above is a specific description of the preferred embodiment of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or replacements without deviating from the spirit of the present application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A method for discriminating a deepfake video, characterized by, The method comprises the following steps: Step S1: obtaining a real image, and applying spatial disturbance to the real image to generate a corresponding fake image; Step S2: generating synthetic images with different degrees of forgery based on the real images and the fake images, wherein the instance re-ranking strategy is adopted by a controllable mixing ratio quantifying the degree of forgery as a continuous supervision signal; Step S3: performing low-frequency enhancement processing on the synthesized image to enhance the low-frequency structure component in the image to obtain a frequency-enhanced image; Step S4: extracting the spatial domain feature and the frequency domain feature of the frequency-enhanced image, respectively; Step S5: fusing the extracted spatial domain feature and the frequency domain feature to obtain fused discriminative features; Step S6: based on the discriminative features, performing authenticity classification and fake intensity level determination on the input video frame. 2.The method of claim 1, wherein, In step S2, the instance re-ranking strategy is used to generate synthesized images with different degrees of forgery, specifically as follows: real image with a fake image generated by spatial disturbance , according to a preset mixing ratio to generate a synthetic image , expressed as: ； wherein, represents a pixel-wise multiplication, is a binary mask defining a blending transition area; blending ratio The greater the value of the The higher the strength of the fake signal in the composite image. Constructing a joint loss function Supervising the model, the joint loss function comprises a binary cross-entropy loss for distinguishing real and fake samples and a categorical cross-entropy loss for learning fine-grained differences among different degrees of forgery which is defined as: ； wherein and is a loss weight, and satisfies .

3. The method of claim 1, wherein the method further comprises: In step S3, the low-frequency enhancement processing on the synthesized image is specifically as follows: to the synthetic image wavelet decomposition, decoupling frequency components including low frequency components , horizontal high frequency components , vertical high frequency components and diagonal high frequency components ​ wherein the wavelet decomposition is by four directional convolution filters implemented, each filter for capturing frequency features in a particular direction, the decomposition process is represented as: ； wherein denotes a convolution operation; using the low frequency components by transposed convolution operation restoring its spatial resolution, and performing a pixel-wise average operation with the original composite image to generate a frequency enhanced image is expressed as: ； wherein, represents a pixel-wise average operation.

4. The method of claim 1, wherein, In step S4, the extraction of the spatial domain feature of the frequency-enhanced image is specifically as follows: using a pre-trained convolutional neural network as a spatial backbone network, encoding the frequency enhanced image to extract high-level spatial feature representations, i.e. spatial domain features is represented by the process: ； wherein, represents an encoder of the spatial backbone network.

5. The method of claim 4, wherein the method further comprises: In step S4, the extraction of the frequency domain feature of the frequency-enhanced image is specifically as follows: On spatial domain features Two-level wavelet transform is performed to obtain a set of directional and scale frequency domain components; For each channel of spatial domain features two-level wavelet decomposition of the channel the set of wavebands is defined as: ； The spatial domain features of all channels are spliced to obtain initial frequency domain features , denoted as: ； wherein, the total number of channels of the spatial domain features the total number of channels of the spatial domain features The initial frequency domain features are obtained by inputting the image into a frequency domain encoder composed of multiple layers of convolution Input a frequency domain encoder composed of multiple layers of convolution to obtain discriminative frequency domain features The frequency domain encoder includes at least three layers, each of which sequentially performs convolution, batch normalization and activation function processing, and contains a squeeze-excitation module to model the dependency between channels.

6. The method of claim 5, wherein the method further comprises: The processing process of the frequency domain encoder is specifically as follows: For input , the first encoding stage, the second encoding stage and the third encoding stage are sequentially passed through to obtain the discriminative frequency domain feature , the process is represented as: ； ； ； wherein, , represents an extrusion-stimulating module, represents a GELU activation function, , , represents a batch normalization operation, represents a convolution operation with a convolution kernel size of , represents a convolution operation with a convolution kernel size of and a step size of 2.

7. The method of claim 1, wherein, In step S5, the fusion of the extracted spatial domain feature and the frequency domain feature is specifically as follows: The spatial domain features and the frequency domain features are mapped to a unified embedding dimension respectively through a projection layer; concatenate the mapped features into a token sequence ; The token sequence is obtained The input is cross-domain information interaction based on a Transformer multi-head self-attention module. extracting the enhanced tokens corresponding to the spatial feature part from the sequence output by the multi-head self-attention module, and passing them through an output projection layer to obtain the spatial feature part with adjustable weights with the original spatial domain features performing residual connection to obtain the fused discriminative features, and the process is represented as: ； wherein, is the fused discriminative feature, is the output projection layer, denotes the first augmented token extracted from the token sequence.

8. The method of claim 1, wherein, In step S6, the output of the model comprises a binary authenticity classification result as well as a quantified value representing a level of forgery strength, said quantified value being associated with the mixing ratio used in the instance re-ranking strategy associated.

9. A deepfake video discrimination system for implementing a deepfake video discrimination method according to any one of claims 1-8, characterized in that, The method comprises the following steps: An image synthesis module is configured to obtain a real image, generate a fake image by applying spatial disturbance, and generate synthesized images with different degrees of forgery based on an instance re-ranking strategy; A low-frequency enhancement module is configured to perform wavelet decomposition on the synthesized image, extract and enhance the low-frequency structure component thereof, and generate a frequency-enhanced image; A feature extraction module is configured to extract the spatial domain feature and the frequency domain feature of the frequency-enhanced image, respectively; A feature fusion module is configured to fuse the spatial domain feature and the frequency domain feature to obtain discriminative features; A classification and determination module is configured to output authenticity classification results and fake intensity levels based on the discriminative features.

Images