Interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception

By using a frequency-guided hierarchical difference perception algorithm in face forgery detection, combined with multi-layer backbone network and attention module, the problems of incomplete feature capture and poor generalization in the existing technology are solved, and face forgery is realized More comprehensive detection and identification.

CN118968584APending Publication Date: 2024-11-15NANJING TECH UNIV

Patent Information

Application Number
CN202410954003.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing face forgery detection algorithm is not comprehensive enough in capturing forgery features. Excessive focus on significant features leads to poor generalization and difficulty in dealing with data in unknown forgery methods, especially in subtle tampering areas.

Method used

Using a hierarchical difference-aware interactive face forgery detection algorithm based on frequency guidance, multiple levels of differences are captured through diversified semantic controllers and attention modules that weaken specific perceptions, combining multi-layer backbone networks and frequency-aware hierarchical interaction modules. Features are learned interactively in the frequency domain to enhance the detection ability of subtle artifacts.

Benefits of technology

It realizes more comprehensive feature capture and better generalization capabilities, can effectively identify diverse and unpredictable forged information, and improves the accuracy and discernment ability of face forgery detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968584A_ABST
    Figure CN118968584A_ABST
Patent Text Reader

Abstract

The invention provides an interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception. According to the algorithm, starting from a spatial domain and a frequency domain, a diversified controller is designed to generate diversified semantic components for an original face sample, and then the semantic components are input into a multi-layer network for feature perception interaction. Therefore, the effective information proportion is improved, the feature mapping is expanded, and the problem of incomplete counterfeit feature capture is solved. According to the algorithm, a backbone network Xception is improved, an attention module for weakening specific perception is introduced into the backbone network Xception, and related semantic information in each semantic component space is gathered, so that multi-level difference features are obtained. Besides, a frequency-guided hierarchical interaction module is introduced, multi-level difference features are interacted and converted to frequency, clustering features of all layers are interacted, and then fine and unpredictable artifacts are captured. And finally, guiding complementary semantic information to be close to each other, and converging different counterfeit information components to carry out true and false face judgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of computer vision and is an interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception. Background Art

[0002] With the rapid development of artificial intelligence technology, especially in the fields of video and image generation, its application has progressed rapidly. Deep generative adversarial technology has attracted widespread attention from academia and industry in the multimedia field (such as deepfake), especially those cases that use generated processed facial images. The rise of generative learning has enabled models to have the ability to imitate and generate highly realistic fake faces, but it has also provided criminals with a means to spread false information and exacerbate social conflicts. Therefore, in order to prevent this hidden danger, effective detection deserves the attention of researchers.

[0003] Existing face forgery detection algorithms mainly focus on mining forgery patterns, such as texture features and frequency artifacts, to detect abnormal tampered areas. Although the performance has improved, these methods do not fully utilize the diverse information in different fields and have difficulty processing data with unknown forgery methods, resulting in poor cross-domain generalization.

[0004] In face forgery detection, the size, number and location of forged features are unpredictable. Especially when the tampered areas are subtle or even difficult to detect, these areas will be filtered out, further misleading the convergence direction of the content during training. At the same time, there are common local artifacts in the frequency features that are easy to capture. How to combine the frequency domain information is very important to capture comprehensive and various false defects. Summary of the invention

[0005] The purpose of the present invention is to address the problem that face forgery detectors do not fully capture forgery features and over-focus on salient features resulting in poor generalization, and propose an interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception.

[0006] The technical solution of the present invention is:

[0007] An interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception, the method comprises the following steps:

[0008] Step 1: Prepare the forged face video dataset, use DLIB to perform face extraction and alignment operations on all video datasets with a uniform number of frames, and adjust all face image samples to a consistent 256×256 size; then input them into the diversified semantic controller and perform shuffling operations at the same time to obtain diversified semantic information that can be used for hierarchical networks.

[0009] Step 2: Input the diversified semantic information into the multi-layer backbone network Xception. By designing the shallow network output with relatively low-level receptive field, the fine semantics with a high proportion of forged information are perceived, and the network level is gradually deepened to input the coarse semantic information with a low proportion of forged information. At the same time, after the output of the multi-layer backbone network, an attention module that weakens specific perception is introduced to gather relevant semantic information in each semantic space, thereby obtaining multi-level difference features.

[0010] Step 3: A frequency-guided hierarchical interaction module is introduced to interactively perceive the obtained multi-level difference feature maps, interact with the clustering features of all layers under the guidance of frequency prompts, capture subtle artifacts, guide complementary semantic information to approach each other, and finally converge different forged information components to determine the authenticity of face image samples.

[0011] Furthermore, the diversified semantic controller is given an input face image I∈R C×H×W , where H×W is the image spatial resolution, n semantic components are set according to the feature scale s, and the image I is divided into s×s windows along the spatial dimension to increase the proportion of effective information in each semantic window. The window for a specific semantic i is represented as The perturbation operation is used to aggregate different feature distributions and generate a new graph G(I, si) to enhance the learning representation.

[0012] Among them, the partition window size is controlled by s, and the feature scale s i where i∈[1,n],s i The larger the size, the smaller each window will be. n Indicates that the window is the original image size;

[0013] The new graph G has n different semantics, which not only increases the proportion of effective information in each window of the same semantics, but can also be used to hierarchically and interactively integrate multiple effective information in diversified semantics.

[0014] Furthermore, obtaining multi-level difference features in step 2 includes the following steps:

[0015] Step 2.1. Due to the complementary effects between the diverse semantics, we use a multi-layer backbone network Xception. By designing truncated outputs in the Entry Flow, Middle Flow, and Exit Flow of Xception, we obtain multi-level feature maps to explore feature representations of different dimensions. Starting from the semantics with fine-grained and high proportion of forged information, we extract low-dimensional features and gradually shift to extracting high-dimensional feature representations of coarse-grained and low-proportion semantics. For the original semantics, we share encoders at each layer and connect the feature outputs of each layer for difference perception, thereby maximizing the use of diverse semantic information;

[0016] Step 2.2: The attention modules that weaken specific perception are located after the backbone network Xception at each level. First, the feature map F∈R C×H×W Input to each attention layer, where C, H, and W are the number of channels, height, and width of the feature map, respectively, and then the attention calculation is performed on each feature component. This can locally aggregate the pseudo features of the adjacent spaces extracted in each layer and enhance the attention to the common semantics of the adjacent space. Finally, the multi-level difference feature H is obtained. l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

[0017] Furthermore, step 2.2 is specifically as follows:

[0018] Step 2.2.1: Given a feature map F, first project all channels C according to the window size and expand them into a feature representation of a certain area. The sum value represents V∈R C×H×W Then we take out a local window V of size M×M with (i, j) as the center L(i,j) ∈R C×H×W , as all values ​​V in the local space L(i,j) ;

[0019]

[0020] Among them, m and n represent the values ​​that can be obtained by traversing the local window;

[0021] Step 2.2.2: In order to aggregate relevant features in the local space, Reinventing M 2 ×M 2 Local aggregated attention map

[0022]

[0023] in Represents the reshaping operation. Then use softmax to transform the attention map Atta (i,j) The values ​​in are normalized to probability distribution, thus obtaining local aggregation weights, which represent the spatial correlation in the local area;

[0024] Step 2.2.3: Perform matrix multiplication and weighted summation on the local aggregation weights and all values ​​to obtain the local aggregation feature map F′ as the final attention output F′ (i,j) ;

[0025] F′ (i,j) =Mat(Sotfmax(Atta (i,j) , VL(i,j) ))

[0026] Among them, Atta (i,j) is the local aggregate attention map, V L(i,j) represents all values ​​in the local window centered at (i, j), Mat(·) represents matrix multiplication and weighted summation; the obtained attention features F′ are output from each layer (i,j) , after a 3×3 Conv convolution layer and a ReLU activation function, more abstract multi-level difference feature information H is extracted l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

[0027] Furthermore, the frequency-guided hierarchical interaction module in step 3 includes the following steps:

[0028] Step 3.1: Connect all levels of difference features H l , and use the fast Fourier transform FFT to transform the features of the image domain into the frequency domain to obtain all clustered frequency domain features G;

[0029] G = FFT(Concat(H1, H2, ... H L ))

[0030] Among them, FFT(·) represents the fast Fourier transform, which can quickly transform features into the frequency domain, but requires less computing time. For a given input feature H l ∈R C×H×W , C, H, and W represent the dimension, width, and height of the feature map, respectively. The input features through FFT(·) will be projected into frequency features, where each frequency component contains the relationship in a specific frequency of the input component;

[0031] Step 3.2: After transformation, use the learnable global filter K∈R K×L×C To exchange all frequency information, and further use the inverse FFT, that is Convert the features back to the original features X;

[0032]

[0033] in, represents element-by-element multiplication, G represents all clustering features, X∈R K×L×C , L is the number of layers in the hierarchical network, and C is the number of channels; at the same time, we further set up jump connections between input and output to speed up the convergence process of the network;

[0034] Step 3.3: Aggregate all features together through multi-head clustering projection and reduce them to a single feature vector to generate the prediction result Y for face forgery detection.

[0035] Y=σ(w·ψ(X)+b)

[0036] Where ψ(·) represents a multi-head clustering projection operation, and the number of clusters is set to 1, which can focus on reducing redundant features to 1 centralized cluster, achieving the effect of hierarchical aggregation of features from all layers. σ(·) represents an activation function, where w and b represent the weight matrix and bias term of the classification activation function, and Y is the prediction result of 0 for a real face or 1 for a fake face.

[0037] The advantages of the present invention compared with the prior art are:

[0038] The present invention extracts and interacts multiple layers of features from the spatial domain and the frequency domain for the face forgery detection task, and can more comprehensively capture rich and generalizable discriminative features for deep forgery detection.

[0039] The present invention improves the shortcomings of the traditional backbone network Xception that has poor perception of scattered local subtle features, and proposes a diversified semantic controller and an attention module that weakens specific perception. Diversified semantic components are generated according to the proportion of effective information to expand the feature map, while weakening specific difference boundaries, thereby aggregating multiple forged components in different spaces and increasing common features. In addition, a frequency-guided hierarchical interaction module is introduced to transform the multi-level spatial domain features extracted by the backbone network into frequency domain features, and multi-level interactive learning is performed to enhance the learned semantic representation. The extracted information is maximized and all facial details are explored to capture diverse and unpredictable forged information.

[0040] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0042] Figure 1 It is a flow chart of the present invention.

[0043] Figure 2 This is a structural diagram of an interactive face forgery detection network model based on frequency-guided hierarchical difference perception.

[0044] Figure 3 This is a structural diagram of the attention module that weakens specific perception. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0046] like Figure 1 As shown, the present invention proposes an interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception, and the method comprises the following steps:

[0047] Step 1: Prepare the face forgery video dataset, use DLIB to perform face extraction and alignment operations on all video datasets with a uniform number of frames, and adjust all face image samples to a consistent 256×256 size; then input them into the diversified semantic controller and perform shuffling operations at the same time to obtain diversified semantic information that can be used in the hierarchical network. The diversified semantic controller is specifically:

[0048] For a given input face image I∈R C×H×W , where H×W is the image spatial resolution, n semantic components are set according to the feature scale s, and the image I is divided into s×s windows along the spatial dimension to increase the proportion of effective information in each semantic window. The window for a specific semantic i is represented as The perturbation operation is used to aggregate different feature distributions and generate a new graph G(I, si), thereby enhancing the learning representation. The size of the partition window is controlled by s, and the feature scale s i where i∈[1,n],s i The larger the size, the smaller each window will be. n Indicates that the window is the original image size;

[0049] The new graph G has n different semantics, which not only increases the proportion of effective information in each window of the same semantics, but can also be used to hierarchically and interactively integrate multiple effective information in diversified semantics.

[0050] like Figure 2 As shown, the frequency-guided hierarchical difference-aware interactive face forgery detection network model includes:

[0051] Step 2: Input the diversified semantic information into the multi-layer backbone network Xception. By designing the shallow network output with relatively low-level receptive field, the fine semantics with a high proportion of forged information are perceived, and the network level is gradually deepened to input the coarse semantic information with a low proportion of forged information. At the same time, after the multi-layer backbone network output, an attention module that weakens specific perception is introduced to gather relevant semantic information in each semantic space, thereby obtaining multi-layer difference features. Obtaining multi-layer difference features includes the following steps:

[0052] Step 2.1. Due to the complementary effects between the diverse semantics, we use a multi-layer backbone network Xception. By designing truncated outputs in the Entry Flow, Middle Flow, and Exit Flow of Xception, we obtain multi-level feature maps to explore feature representations of different dimensions. Starting from the semantics with fine-grained and high proportion of forged information, we extract low-dimensional features and gradually shift to extracting high-dimensional feature representations of coarse-grained and low-proportion semantics. For the original semantics, we share encoders at each layer and connect the feature outputs of each layer for difference perception, thereby maximizing the use of diverse semantic information;

[0053] Step 2.2: The attention modules that weaken specific perception are located after the backbone network Xception at each level. First, the feature map F∈R C×H×W Input to each attention layer, where C, H, and W are the number of channels, height, and width of the feature map, respectively, and then the attention calculation is performed on each feature component. This can locally aggregate the pseudo features of the adjacent spaces extracted in each layer and enhance the attention to the common semantics of the adjacent space. Finally, the multi-level difference feature H is obtained. l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

[0054] like Figure 3 As shown, the attention module that weakens specific perception includes:

[0055] Step 2.2.1: Given a feature map F, first project all channels C according to the window size and expand them into a feature representation of a certain area. The sum value represents V∈R C×H×W Then we take out a local window V of size M×M with (i, j) as the center L(i,j) ∈R C×H×W , as all values ​​V in the local space L(i,j) ;

[0056]

[0057] Among them, m and n represent the values ​​that can be obtained by traversing the local window;

[0058] Step 2.2.2: In order to aggregate relevant features in the local space, Reinventing M 2 ×M 2 Local aggregated attention map

[0059]

[0060] in Represents the reshaping operation. Then use softmax to transform the attention map Atta (i,j) The values ​​in are normalized to probability distribution, thus obtaining local aggregation weights, which represent the spatial correlation in the local area;

[0061] Step 2.2.3: Perform matrix multiplication and weighted summation on the local aggregation weights and all values ​​to obtain the local aggregation feature map F′ as the final attention output F′ (i,j) ;

[0062] F′ (i,j) =Mat(Sotfmax(Atta (i,j) , V L(i,j) ))

[0063] Among them, Atta (i,j) is the local aggregate attention map, V L(i,j) represents all values ​​in the local window centered at (i, j), Mat(·) represents matrix multiplication and weighted summation; the obtained attention features F′ are output from each layer (i,j) , after a 3×3 Conv convolution layer and a ReLU activation function, more abstract multi-level difference feature information H is extracted l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

[0064] Step 3: The frequency-guided hierarchical interaction module is introduced to interactively perceive the obtained multi-level difference feature maps, interact with the clustering features of all layers under the guidance of frequency prompts, capture subtle artifacts, guide complementary semantic information to approach each other, and finally converge different forged information components to determine the authenticity of face image samples. The frequency-guided hierarchical interaction module includes the following steps:

[0065] Step 3.1: Connect all levels of difference features H l , and use the fast Fourier transform FFT to transform the features of the image domain into the frequency domain to obtain all clustered frequency domain features G;

[0066] G = FFT(Concat(H1, H2, ... H L ))

[0067] Among them, FFT(·) represents the fast Fourier transform, which can quickly transform features into the frequency domain, but requires less computing time. For a given input feature H l ∈R C×H×W , C, H, and W represent the dimension, width, and height of the feature map, respectively. The input features through FFT(·) will be projected into frequency features, where each frequency component contains the relationship in a specific frequency of the input component;

[0068] Step 3.2: After transformation, use the learnable global filter K∈R K×L×C To exchange all frequency information, and further use the inverse FFT, that is Convert the features back to the original features X;

[0069]

[0070] in, represents element-by-element multiplication, G represents all clustering features, X∈R K×L×C , L is the number of layers in the hierarchical network, and C is the number of channels; at the same time, we further set up jump connections between input and output to speed up the convergence process of the network;

[0071] Step 3.3: Aggregate all features together through multi-head clustering projection and reduce them to a single feature vector to generate the prediction result Y for face forgery detection.

[0072] Y=σ(w·ψ(X)+b)

[0073] Where ψ(·) represents a multi-head clustering projection operation, and the number of clusters is set to 1, which can focus on reducing redundant features to 1 centralized cluster, achieving the effect of hierarchical aggregation of features from all layers. σ(·) represents an activation function, where w and b represent the weight matrix and bias term of the classification activation function, and Y is the prediction result of 0 for a real face or 1 for a fake face.

[0074] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. An interactive face forgery detection algorithm based on frequency-guided hierarchical difference perception, characterized in that: The method comprises the following steps: Step 1: Prepare the forged face video dataset, use DLIB to perform face extraction and alignment operations on all video datasets with a uniform number of frames, and adjust all face image samples to a consistent 256×256 size; then input them into the diversified semantic controller and perform shuffling operations at the same time to obtain diversified semantic information that can be used for hierarchical networks. Step 2: Input the diversified semantic information into the multi-layer backbone network Xception. By designing the shallow network output with relatively low-level receptive field, the fine semantics with a high proportion of forged information are perceived, and the network level is gradually deepened to input the coarse semantic information with a low proportion of forged information. At the same time, after the output of the multi-layer backbone network, an attention module that weakens specific perception is introduced to gather relevant semantic information in each semantic space, thereby obtaining multi-level difference features. Step 3: A frequency-guided hierarchical interaction module is introduced to interactively perceive the obtained multi-level difference feature maps, interact with the clustering features of all layers under the guidance of frequency prompts, capture subtle artifacts, guide complementary semantic information to approach each other, and finally converge different forged information components to determine the authenticity of face image samples.

2. The frequency-guided hierarchical difference-aware interactive face forgery detection algorithm according to claim 1, characterized in that: The diverse semantic controllers in step 1 include: For a given input face image I∈R C×H×W , where H×W is the image spatial resolution, n semantic components are set according to the feature scale s, and the image I is divided into s×s windows along the spatial dimension to increase the proportion of effective information in each semantic window. The window for a specific semantic i is represented as The perturbation operation is used to aggregate different feature distributions and generate a new graph G(I, si), thereby enhancing the learned representation. Among them, the partition window size is controlled by s, and the feature scale s i where i∈[1,n],s i The larger the size, the smaller each window will be. n Indicates that the window is the original image size; The new graph G has n different semantics, which not only increases the proportion of effective information in each window of the same semantics, but can also be used to hierarchically and interactively integrate multiple effective information in diversified semantics.

3. The frequency-guided hierarchical difference-aware interactive face forgery detection algorithm according to claim 1, characterized in that: Obtaining multi-level difference features in step 2 includes the following steps: Step 2.

1. Due to the complementary effects between the diverse semantics, we use a multi-layer backbone network Xception. By designing truncated outputs in the Entry Flow, Middle Flow, and Exit Flow of Xception, we obtain multi-level feature maps to explore feature representations of different dimensions. Starting from the semantics with fine-grained and high proportion of forged information, we extract low-dimensional features and gradually shift to extracting high-dimensional feature representations of coarse-grained and low-proportion semantics. For the original semantics, we share encoders at each layer and connect the feature outputs of each layer for difference perception, thereby maximizing the use of diverse semantic information; Step 2.2: The attention modules that weaken specific perception are located after the backbone network Xception at each level. First, the feature map F∈R C×H×W Input to each attention layer, where C, H, and W are the number of channels, height, and width of the feature map, respectively, and then the attention calculation is performed on each feature component. This can locally aggregate the pseudo features of the adjacent spaces extracted in each layer and enhance the attention to the common semantics of the adjacent space. Finally, the multi-level difference feature H is obtained. l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

4. The frequency-guided hierarchical difference-aware interactive face forgery detection algorithm according to claim 1 is characterized in that Step 2.2 is as follows: Step 2.2.1: Given a feature map F, first project all channels C according to the window size and expand them into a feature representation of a certain area. The sum value represents V∈R C×H×W Then we take out a local window V of size M×M with (i, j) as the center L(i,j) ∈R C×H×W , as all values ​​V in the local space L(i,j) ; Among them, m and n represent the values ​​that can be obtained by traversing the local window; Step 2.2.2: In order to aggregate relevant features in the local space, Reinventing M 2 ×M 2 Local aggregated attention map in Represents the reshaping operation. Then use softmax to transform the attention map Atta (i,j) The values ​​in are normalized to probability distribution, thus obtaining local aggregation weights, which represent the spatial correlation in the local area; Step 2.2.3: Perform matrix multiplication and weighted summation on the local aggregation weights and all values ​​to obtain the local aggregation feature map F′ as the final attention output F′ (i,j) ; <h2 style=";text-align:left;direction:ltr">F′<h2 style=";text-align:left;direction:ltr"> (i,j) <h2 style=";text-align:left;direction:ltr"> =Mat(Sotfmax(Atta<h2 style=";text-align:left;direction:ltr"> (i,j) <h2 style=";text-align:left;direction:ltr"> V<h2 style=";text-align:left;direction:ltr"> L(i,j) <h2 style=";text-align:left;direction:ltr"> )) Among them, Atta (i,j) is the local aggregate attention map, V L(i,j) represents all values ​​in the local window centered at (i, j), Mat(·) represents matrix multiplication and weighted summation; the obtained attention features F′ are output from each layer (i,j) , after a 3×3 Conv convolution layer and a ReLU activation function, more abstract multi-level difference feature information H is extracted l , where l∈[1,L] represents a hierarchical network branch and L represents the number of hierarchical network layers.

5. The frequency-guided hierarchical difference-aware interactive face forgery detection algorithm according to claim 1, characterized in that: The frequency-guided hierarchical interaction module in step 3 includes the following steps: Step 3.1: Connect all levels of difference features H l , and use the fast Fourier transform FFT to transform the features of the image domain into the frequency domain to obtain all clustered frequency domain features G; G=FFT(Concat(H1,H2,...H L )) Among them, FFT(·) represents the fast Fourier transform, which can quickly transform features into the frequency domain, but requires less computing time. For a given input feature H l ∈R C×H×W , C, H, and W represent the dimension, width, and height of the feature map, respectively. The input features through FFT(·) will be projected into frequency features, where each frequency component contains the relationship in a specific frequency of the input component; Step 3.2: After transformation, use the learnable global filter K∈R K×L×C To exchange all frequency information, and further use the inverse FFT, that is Convert the features back to the original features X; in, represents element-by-element multiplication, G represents all clustering features, X∈R K×L×C , L is the number of layers in the hierarchical network, and C is the number of channels; at the same time, we further set up jump connections between input and output to speed up the convergence process of the network; Step 3.3: Aggregate all features together through multi-head clustering projection and reduce them to a single feature vector to generate the prediction result Y for face forgery detection. Y=σ(w·ψ(X)+b) Where ψ(·) represents a multi-head clustering projection operation, and the number of clusters is set to 1, which can focus on reducing redundant features to 1 centralized cluster, achieving the effect of hierarchical aggregation of features from all layers. σ(·) represents an activation function, where w and b represent the weight matrix and bias term of the classification activation function, and Y is the prediction result of 0 for a real face or 1 for a fake face.

Citation Information

Patent Citations

  • Face forgery detection method based on image block disruption

    CN113269167A

  • Deep fake face detection method based on frequency learning

    CN114898437A

  • Deep forgery detection method and corresponding device

    CN115311525A

  • Face forgery detection method based on feature decoupling

    CN116416686A

  • Counterfeit face detection method and device

    CN117894053A

Cited By

  • Crack semantic segmentation model training method, segmentation method and related device

    CN119478421A

  • A method for training a crack semantic segmentation model, a segmentation method, and related devices

    CN119478421B