Deep forgery detection method and system for removing identity information

Abstract

The invention relates to a deep forgery detection method and system for removing identity information, and the method comprises the following steps: S10, positioning a face region from a to-be-detected image, and obtaining a face region image; s20, extracting human face shallow layer features from the human face region image to obtain a shallow layer feature map; s30, extracting and enhancing texture features from the shallow feature map to obtain a texture enhanced feature map; s40, extracting semantic features from the shallow feature map to obtain a deep semantic feature map; s50, generating a plurality of attention maps and semantic feature maps based on the deep semantic feature map; s60, respectively carrying out artifact detection on the texture enhancement feature map and the deep semantic feature map through multi-scale anchor points; s70, fusing the attention map, the texture enhancement feature map, the semantic feature map and the artifact detection result to obtain local features and global features; and S80, performing judgment based on the global features and the local features. The detection precision can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a deep forgery detection method and system that removes identity information. Background Art

[0002] Facial recognition is a commonly used technology for identity verification, such as face unlocking and face payment. However, the widespread use of facial recognition technology has also led to the emergence of forged faces. Therefore, the ability to detect forged faces is crucial for the further development of facial recognition technology.

[0003] With the development of deep learning technology, deep learning-based face forgery detection has emerged. For example, the Chinese invention patent CN 116311430 A discloses a "Deep Forgery Detection Method and Device Based on Diversified Image Features." This method extracts texture, semantic, and temporal features through channel attention, spatial attention, and temporal attention mechanisms, respectively, and employs a bidirectional GRU to capture frame-level differences. Finally, classification is performed by channel-wise fusion of features. While this approach can improve feature diversity, it may miss subtle forgery traces in complex backgrounds or low-resolution images, resulting in inaccurate detection results. Furthermore, its feature fusion method, which relies on simple channel-wise splicing, makes it difficult to dynamically weight key areas, potentially causing the model to misuse identity-related features, affecting cross-dataset generalization performance. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep fake detection method and system that removes identity information to improve local detection accuracy and identity information protection capabilities.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for detecting deep fakes by removing identity information, comprising the following steps:

[0007] S10, locating a face region from the image to be detected and obtaining a face region image;

[0008] S20, extracting shallow facial features from the face region image to obtain a shallow feature map;

[0009] S30, extracting and enhancing texture features from the shallow feature map to obtain a texture enhanced feature map;

[0010] S40, extracting semantic features from the shallow feature map to obtain a deep semantic feature map;

[0011] S50, generating a plurality of attention maps based on the deep semantic feature map conversion, and extracting semantic features from the deep semantic feature map again to obtain a semantic feature map;

[0012] S60, performing artifact detection on the texture enhancement feature map and the deep semantic feature map respectively through multi-scale anchor points, and obtaining artifact detection results of the texture enhancement feature map and artifact detection results of the deep semantic feature map respectively;

[0013] S70, fusing the attention map, the texture enhancement feature map, and the artifact detection results of the texture enhancement feature map to obtain local features; fusing the artifact detection results of the attention map, the semantic feature map, and the deep semantic feature map to obtain global features;

[0014] S80: The discriminator makes a judgment based on the global features and the local features to obtain a probability that the face region image is a forged image.

[0015] In a second aspect, the present invention provides a deep fake detection system that removes identity information, comprising:

[0016] The face region positioning module is used to locate the face region from the image to be detected and obtain the face region image;

[0017] A shallow feature extraction module is used to extract shallow facial features from the face area image to obtain a shallow feature map;

[0018] A texture feature extraction module is used to extract and enhance texture features from the shallow feature map to obtain a texture enhanced feature map;

[0019] A semantic feature extraction module is used to extract semantic features from the shallow feature map to obtain a deep semantic feature map;

[0020] A semantic feature conversion module is used to generate multiple attention maps based on the deep semantic feature map conversion, and to extract semantic features from the deep semantic feature map again to obtain a semantic feature map;

[0021] A multi-scale artifact detection module is used to perform artifact detection on the texture enhancement feature map and the deep semantic feature map through multi-scale anchor points, and obtain artifact detection results of the texture enhancement feature map and the deep semantic feature map respectively;

[0022] The feature fusion module is used to fuse the artifact detection results of the attention map, texture enhancement feature map, and texture enhancement feature map to obtain local features; and fuse the artifact detection results of the attention map, semantic feature map, and deep semantic feature map to obtain global features;

[0023] The discriminator is used to make a judgment based on global features and local features to obtain the probability that the face area image is a forged image.

[0024] In a third aspect, the present invention provides a computer program product comprising computer-readable instructions, wherein the computer-readable instructions, when executed by a processor, implement the steps of the deep fake detection method for removing identity information of the present invention.

[0025] In a fourth aspect, the present invention provides a computer-readable storage medium comprising computer-readable instructions, characterized in that when the computer-readable instructions are executed by a processor, the steps of the deep counterfeit detection method for removing identity information of the present invention are implemented.

[0026] In a fifth aspect, the present invention provides an electronic device comprising: a memory storing program instructions; a processor connected to the memory, executing the program instructions in the memory, and implementing the steps of the deep fake detection method for removing identity information of the present invention.

[0027] Compared with the prior art, the present invention has the following technical advantages:

[0028] (1) By generating multiple attention maps to focus on different areas of the face image and fusing global features and local features for discrimination, the dependence on identity information is avoided, thereby reducing the misuse of identity information and paying attention to subtle texture features, thereby improving the generalization ability and detection performance of the model.

[0029] (2) The location of the forged area is detected by multi-scale anchor points, and the texture features are efficiently enhanced, so that the model can better focus on local forged features, improving the accuracy and robustness of detection.

[0030] For other advantages of the present invention, please refer to the relevant description in the embodiment section. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 This is a flowchart of a deep fake detection method that removes identity information provided in an embodiment.

[0033] Figure 2 This is a logical architecture diagram of a deep fake detection model that removes identity information provided in an embodiment.

[0034] Figure 3 This is a block diagram of a deep fake detection system that removes identity information provided in an embodiment.

[0035] Figure 4 This is a block diagram of the composition of an electronic device. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0037] See also Figure 1 , a deep fake detection method for removing identity information provided in this embodiment includes the following steps:

[0038] S10, locating the face region from the image to be detected, and obtaining a face region image.

[0039] See Figure 2 , the detection method of this embodiment can detect both single images and videos. If the object to be detected is a video, it is necessary to first extract frames from the video to obtain multiple images, and then filter out images containing faces from the multiple images. The present invention mainly detects deep fakes of facial images, and some of the image frames that make up the video may not contain the face area or do not contain the complete face area. Therefore, after extracting the video frames, it is necessary to filter and remove image frames that do not contain complete faces and blurred faces (unrecognizable). The filtered images containing faces are the images to be detected ( Figure 2 If the object to be detected is directly a single image, then this image is also the image to be detected.

[0040] During this step, RetinaFace can be used to perform face detection, locate the face region, and obtain a face region image. The face region image can also be cropped, for example, to a size of 380×380, to facilitate subsequent processing by unifying the size.

[0041] S20, extracting shallow facial features from the face region image to obtain a shallow feature map.

[0042] In the specific implementation, the face area image is input into the backbone network 1 ( Figure 2Backbone Layers 1 in the

[15] (using the Efficient-B4 architecture) undergoes processing through multiple convolutional layers and activation functions to obtain shallow facial features of the facial region image. These features primarily include facial features distribution and color features. Facial features distribution includes facial contour and key point spacing (interpupillary distance), while color features include skin color and hair color. The extracted shallow facial features are then fed into multiple convolutional and pooling layers to produce a shallow feature map Sf (Shallow Feature) with a certain degree of abstraction.

[0043] S30, extracting and enhancing texture features from the shallow feature map to obtain a texture enhanced feature map.

[0044] The processed shallow feature map Sf will be input into the texture feature enhancement module ( Figure 2 The TextureHighlight block in the image is used to extract and enhance texture features, and output a texture enhanced feature map THf (Texture Enhanced Feature Map).

[0045] The texture feature enhancement module includes a texture feature extraction module and a CondenseNet module. The texture feature extraction module first downsamples the shallow feature map Sf, reducing its resolution to better focus on local features. It then performs average pooling on the downsampled shallow feature map to obtain a new feature map D. This process helps extract the average information from the shallow feature map of the face. Finally, residual processing is performed on the shallow feature map Sf and the feature map D. This process involves subtracting Sf from the feature map D to obtain the texture feature map Tf (Texture Feature Map), which primarily contains the image's texture details.

[0046] The texture feature map Tf will be input into the CondenseNet module for texture feature enhancement. The enhanced texture feature can be called texture enhanced feature.

[0047] The CondeseNet module consists of an input layer, a BN-ReLu layer, a 1x1 convolution layer, a BN-ReLu layer, a 3x3 convolution layer, and an output layer. First, the 1x1 convolution layer (convolution layer one) of the CondeseNet module performs the first feature extraction and compression on Tf to reduce redundant features; then the feature map after convolution is normalized to make the distribution of the texture feature map more stable. Then, the texture feature map is fused and enhanced through the Relu activation function and the 3x3 convolution layer (convolution layer two) to improve the richness of texture features. Finally, the enhanced texture feature map is pooled, and during the training process, redundant connections are eliminated by pruning connections to improve the efficiency of feature extraction. After processing by the CondeseNet module, the texture enhanced feature map THf (Texture Enhanced Feature Map) is obtained. This feature map retains and enhances the complete subtle texture features, providing richer feature information for subsequent artifact detection.

[0048] S40, extracting semantic features from the shallow feature map to obtain a deep semantic feature map.

[0049] See Figure 2 , the shallow feature map Sf passes through the backbone network 2( Figure 2 Backbone Layers2 in

[15] uses the Efficient-B4 structure) and the attention layer to extract deep semantic features from the shallow feature map Sf and generate a basic deep semantic feature map.

[0050] S50, generating multiple attention maps based on the deep semantic feature map conversion, and extracting semantic features from the deep semantic feature map again to obtain a semantic feature map.

[0051] See Figure 2 The deep semantic feature map is transformed into multiple attention maps through the attention block, which consists of a 1x1 convolutional layer, a batch normalization layer, and a Relu layer. Each attention map can highlight a key area in the image. Generating multiple attention maps allows us to focus on different important areas, allowing us to focus on key local features of the face.

[0052] In deepfake detection, traces of forgery may appear in different areas of the image. By generating multiple attention maps, we can focus on multiple different local areas at the same time, thereby capturing forgery characteristics more comprehensively. Each attention map can focus on different parts of the image, such as key areas such as the eyes, nose, and mouth. This fine-grained attention method helps detect more subtle traces of forgery. By using multiple attention maps, we can focus on local features rather than global identity information, which helps reduce "implicit identity leakage" and improve the generalization ability of the model.

[0053] The deep semantic feature map passes through the backbone network 3( Figure 2 Backbone Layers3 in the

[15] adopts the Efficient-B4 structure to extract semantic features again and obtain a semantic feature map (Semantic Feature Map). This semantic feature map has richer semantic features than the basic deep semantic feature map.

[0054] S60, performing artifact detection on the texture enhancement feature map and the deep semantic feature map respectively through multi-scale anchor points, and obtaining artifact detection results of the texture enhancement feature map and artifact detection results of the deep semantic feature map respectively.

[0055] The texture enhancement feature map THf and the deep semantic feature map will be input into a multi-scale artifact detection module ( Figure 2 The APM module in the

[15] (APM stands for "Artifact Perception Module"). The multi-scale artifact detection module consists of a backbone network, multi-scale extra layers, and fully connected layers. By adding four extra layers of different scales to the end of the backbone network (testing has shown that setting four extra layers of different scales ensures that the model can focus well on both the overall picture and details, with sizes of 7×7, 5×5, 3×3, and 1×1, respectively), the size of the input image (texture-enhanced feature map or deep semantic feature map) decreases with the addition of these extra layers. These feature maps of different scales help detect artifact regions of different sizes. During training, multi-scale default anchors are used to detect artifact regions in fake images. Each feature map grid is associated with multiple default anchors of different scales on the input image. The output is the position offset (m×4) and category confidence (m×2, i.e., fake or real) of each default anchor on the image. If the intersection over union (IoU) between the anchor box and the ground truth of the artifact region is greater than a threshold, the default anchor box is annotated as a fake anchor. Therefore, after the texture enhancement feature map is processed by additional layers of different scales, it is possible to identify whether there is an artifact area or locate the position of the artifact area.

[0056] When this step is executed, the location of the artifact area in the texture enhancement feature map THf and the deep semantic feature map, as well as the presence of the artifact area, are detected based on the multi-scale anchor points. A short connection is created between the final 1×1 feature map and the end of the backbone network, and the output is sent to a fully connected layer to generate the final prediction. The output here refers to the position offset information of each default anchor box and the category confidence of each default anchor box. For each anchor box, the multi-scale artifact detection module outputs an m×4 quantity and an m×2 quantity, where m is the number of anchor boxes, 4 represents the coordinate offset of the anchor box (such as the center coordinate, width and height offset), and 2 represents the confidence of the two categories (artifact or non-artifact). The final prediction refers to the fully connected layer outputting a probability value, indicating whether there is an artifact area in the texture enhancement feature map. This result is a scalar value indicating the probability that the image belongs to the artifact category (forged image) or the non-artifact category (real image).

[0057] S70, fusing the attention map, the texture enhancement feature map, and the artifact detection results of the texture enhancement feature map to obtain local features; fusing the artifact detection results of the attention map, the semantic feature map, and the deep semantic feature map to obtain global features.

[0058] See Figure 2 , through a bilinear attention pooling layer ( Figure 2 In the BAP module (BAP stands for "Bilinear Attention Pooling"), the texture-enhanced feature map and the semantic feature map are used as input during the fusion process. The attention map guides the focus on different areas of the image. The weighted pooling layer performs a pooling operation, which weights the different parts of each feature map to generate a comprehensive feature representation. The comprehensive feature representation of the texture-enhanced feature map is fused with the artifact detection results of the texture-enhanced feature map to obtain local features. The comprehensive feature representation of the semantic feature map is fused with the artifact detection results of the deep semantic feature map to obtain global features.

[0059] S80: The discriminator makes a judgment based on the global features and the local features to obtain a probability that the face region image is a forged image.

[0060] The fused global and local features are fed into the discriminator, which processes these features through a series of fully connected layers and activation functions. The discriminator uses the nonlinear relationships between features learned during training to distinguish between real and fake images. The final layer of the discriminator is a fully connected layer whose output dimension matches the number of categories. The output is normalized using a softmax function to obtain the predicted probability for each category (real or fake).

[0061] Global features are primarily extracted through deep feature maps in the backbone network, reflecting the overall image information and semantics. Local features are primarily generated through texture enhancement by processing shallow feature maps in the texture feature enhancement module, as well as the position and confidence information obtained from artifact regions detected by the multi-scale artifact detection module. Furthermore, attention maps are generated to focus on key areas in the image. Both global and local features are input into the discriminator, allowing the model to combine texture and semantic information to improve its ability to detect artifact regions. This multi-layered feature extraction and fusion strategy helps improve the detection accuracy and generalization capabilities of the overall model.

[0062] During the training phase, the cross entropy loss of the classification can be calculated based on the comparison between the predicted probability and the true label to measure the accuracy of the model classification. Then, the gradient descent is performed through the composite loss function to update the model parameters. The composite loss function L is composed of the detection loss L test , regional independence loss L multi and cross entropy loss L cross The composite loss function is:

[0063] L=L test +λ1·L multi +λ2·L cross (1)

[0064]

[0065] Among them, λ1 and λ2 are the weights of region independence loss and cross entropy loss, respectively, and the detection loss consists of confidence loss and location loss:

[0066]

[0067] The symbols involved in the above formula are shown in the following table:

[0068]

[0069] In the above methods, multi-scale anchor points are used to locate forged areas, accurately capturing artifact features at different scales and significantly improving detection comprehensiveness and accuracy, especially in complex scenarios. CondenseNet pruning enhances texture features, strengthening texture extraction while lightweighting the model, reducing computational complexity and sensitively capturing texture changes. Bilinear attention pooling guides feature fusion, combining the advantages of global and local features, highlighting key areas and reducing reliance on identity information, improving recognition capabilities and information protection. The composite loss function integrates detection, regional independence, and cross-entropy loss to optimize classification and positioning, enhancing adaptability and stability. The innovative integration of these technical points makes the model more robust in local detection accuracy, identity information protection, and complex cross-dataset scenarios, with significant advantages.

[0070] See also Figure 3 Based on the same inventive concept, an embodiment of the present invention further provides a deep fake detection system that removes identity information, including:

[0071] The face region positioning module is used to locate the face region from the image to be detected and obtain the face region image;

[0072] A shallow feature extraction module is used to extract shallow facial features from the face area image to obtain a shallow feature map;

[0073] A texture feature extraction module is used to extract and enhance texture features from the shallow feature map to obtain a texture enhanced feature map;

[0074] A semantic feature extraction module is used to extract semantic features from the shallow feature map to obtain a deep semantic feature map;

[0075] A semantic feature conversion module is used to generate multiple attention maps based on the deep semantic feature map conversion, and to extract semantic features from the deep semantic feature map again to obtain a semantic feature map;

[0076] A multi-scale artifact detection module is used to perform artifact detection on the texture enhancement feature map and the deep semantic feature map through multi-scale anchor points, and obtain artifact detection results of the texture enhancement feature map and the deep semantic feature map respectively;

[0077] The feature fusion module is used to fuse the artifact detection results of the attention map, texture enhancement feature map, and texture enhancement feature map to obtain local features; and fuse the artifact detection results of the attention map, semantic feature map, and deep semantic feature map to obtain global features;

[0078] The discriminator is used to make a judgment based on global features and local features to obtain the probability that the face area image is a forged image.

[0079] For the specific execution operations of each module, please refer to the relevant steps of the aforementioned method embodiment, which will not be repeated here.

[0080] like Figure 4 As shown, this embodiment also provides an electronic device, which may include a processor 41 and a memory 42, wherein the memory 42 is coupled to the processor 41. It is worth noting that this figure is exemplary, and other types of structures may be used to supplement or replace this structure to implement data extraction, report generation, communication or other functions.

[0081] like Figure 4 As shown, the electronic device may further include: an input unit 43, a display unit 44 and a power supply 45. It is worth noting that the electronic device does not necessarily have to include Figure 4 In addition, electronic devices may also include Figure 4 For components not shown, reference may be made to the prior art.

[0082] The processor 41 is sometimes also called a controller or an operation control, and may include a microprocessor or other processor devices and / or logic devices. The processor 41 receives inputs and controls the operations of various components of the electronic device.

[0083] The memory 42 may be, for example, one or more of a cache, flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable devices, and may store information such as configuration information of the processor 41 and instructions executed by the processor 41. The processor 41 may execute programs stored in the memory 42 to implement information storage or processing. In one embodiment, the memory 42 also includes a buffer memory to store intermediate information.

[0084] An embodiment of the present invention further provides a computer program product, comprising computer-readable instructions. When the computer-readable instructions are executed in an electronic device, the program product enables the electronic device to perform the operating steps included in the method of the present invention.

[0085] An embodiment of the present invention further provides a storage medium storing computer-readable instructions, wherein the computer-readable instructions enable an electronic device to execute the operation steps included in the method of the present invention.

[0086] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0087] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0088] The above-described embodiments are merely specific implementations of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications, substitutions, and improvements within the technical scope disclosed in the present invention, and such modifications, substitutions, and improvements are intended to be encompassed within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A deep fake detection method that removes identity information, characterized in that: The following steps are involved: S10, locating a face region from the image to be detected and obtaining a face region image; S20, extracting shallow facial features from the face region image to obtain a shallow feature map; S30, extracting and enhancing texture features from the shallow feature map to obtain a texture enhanced feature map; S40, extracting semantic features from the shallow feature map to obtain a deep semantic feature map; S50, generating a plurality of attention maps based on the deep semantic feature map conversion, and extracting semantic features from the deep semantic feature map again to obtain a semantic feature map; S60, performing artifact detection on the texture enhancement feature map and the deep semantic feature map respectively through multi-scale anchor points, and obtaining artifact detection results of the texture enhancement feature map and artifact detection results of the deep semantic feature map respectively; S70, fusing the attention map, the texture enhancement feature map, and the artifact detection results of the texture enhancement feature map to obtain a local feature; The artifact detection results of the attention map, semantic feature map and deep semantic feature map are integrated to obtain the global feature; S80: The discriminator makes a judgment based on the global features and the local features to obtain a probability that the face region image is a forged image.

2. A deep fake detection method for removing identity information according to claim 1, characterized in that: The processing of S30 includes: First, the shallow feature map is downsampled, and then the downsampled shallow feature map is average pooled to obtain a new feature map. Finally, the shallow feature map and the new feature map are subjected to residual processing to obtain a texture feature map. The first convolution layer extracts and compresses the texture feature map and performs normalization processing; then the normalized feature map is subjected to feature fusion and enhancement through the Relu activation function and the second convolution layer; finally, the enhanced feature map is pooled to obtain a texture enhanced feature map.

3. A deep fake detection method for removing identity information according to claim 1, characterized in that: The processing of S50 includes: the deep semantic feature map is converted into multiple attention maps through the attention block, and each attention map focuses on a local area of the face; the deep semantic feature map is passed through the backbone network of the Efficient-B4 structure to extract semantic features again to obtain a semantic feature map.

4. A deep fake detection method for removing identity information according to claim 1, characterized in that: The processing of S60 includes: detecting the positions of artifact areas on the texture enhancement feature map and the deep semantic feature map based on multi-scale anchor points, and outputting the position offset and category confidence of each default anchor point on the texture enhancement feature map and the deep semantic feature map.

5. The method for detecting deep fakes by removing identity information according to claim 1, characterized in that: In S70, using the attention map as a guide, a bilinear attention pooling layer is used to perform weighted summation on different parts of the input image to generate a comprehensive feature representation, wherein the input image is a texture enhancement feature map and a semantic feature map; the comprehensive feature representation of the texture enhancement feature map is fused with its artifact detection result to obtain a local feature; The comprehensive feature representation of the semantic feature map is fused with the artifact detection results of the deep semantic feature map to obtain the global feature.

6. A deep fake detection method for removing identity information according to claim 1, characterized in that: In the process of training the discriminator, gradient descent is performed through the composite loss function to update the model parameters of the discriminator. The composite loss function is: L = L test +λ1·L multi +λ2·L cross , λ1 and λ2 are the weights of regional independence loss and cross entropy loss respectively, y i is the true label, is the predicted probability, V ij is the facial feature vector, c j is the feature center, m in is the boundary within the class, m out is the boundary between classes, N is the number of positive samples, B is the batch size, M is the number of attention heads, and L conf is the confidence loss, L loc is the position loss, α is the weight, x ij is the coordinate of the real frame, are the predicted coordinates.

7. A deep fake detection system that removes identity information, characterized in that: include: The face region positioning module is used to locate the face region from the image to be detected and obtain the face region image; A shallow feature extraction module is used to extract shallow facial features from the face area image to obtain a shallow feature map; A texture feature extraction module is used to extract and enhance texture features from the shallow feature map to obtain a texture enhanced feature map; A semantic feature extraction module is used to extract semantic features from the shallow feature map to obtain a deep semantic feature map; A semantic feature conversion module is used to generate multiple attention maps based on the deep semantic feature map conversion, and to extract semantic features from the deep semantic feature map again to obtain a semantic feature map; A multi-scale artifact detection module is used to perform artifact detection on the texture enhancement feature map and the deep semantic feature map through multi-scale anchor points, and obtain artifact detection results of the texture enhancement feature map and the deep semantic feature map respectively; Feature fusion module, used to fuse the attention map, texture enhancement feature map and artifact detection results of the texture enhancement feature map to obtain local features; The artifact detection results of the attention map, semantic feature map and deep semantic feature map are integrated to obtain the global feature; The discriminator is used to make a judgment based on global features and local features to obtain the probability that the face area image is a forged image.

8. A computer program product comprising computer-readable instructions, characterized in that: When executed by a processor, the computer-readable instructions implement the steps of the deep fake detection method for removing identity information as described in any one of claims 1-6.

9. A computer-readable storage medium comprising computer-readable instructions, characterized in that: When executed by a processor, the computer-readable instructions implement the steps of the deep fake detection method for removing identity information as described in any one of claims 1-6.

10. An electronic device, characterized in that: include: Memory, which stores program instructions; A processor is connected to the memory and executes program instructions in the memory to implement the steps of the deep counterfeit detection method for removing identity information as described in any one of claims 1 to 6.

Images