An unsupervised approach to face forgery assessment

Abstract

The invention discloses an unsupervised face forgery assessment method, comprising: taking the coarse-grained features of a face image as a weight parameter, and extracting facial action features, image color features and fine-grained features of different face images of a single person to construct a reconstructed face image of the person; on this basis, by combining the reconstruction loss of the reconstructed face image and the original face image, the perceptual loss based on the enhanced features of the reconstructed face image and the original face image, the DSVDD loss based on the fine-grained features of the original face image and the related face images, and the cosine similarity loss based on the facial action features and image color features of the original face image, a coarse-grained feature extraction module, a fine-grained feature generation module and a reconstruction model of a reconstruction module are jointly trained, so as to improve the robustness of the model, that is, to improve the accuracy of the face image reconstructed by the reconstruction model, and to realize face forgery assessment based on the difference between the accurate reconstructed face image and the original face image.

Description

Technical Field

[0001] The invention belongs to the field of face detection, and in particular relates to an unsupervised face forgery assessment method. Background Art

[0002] Face forgery detection is ultimately a classification problem, and its purpose is to distinguish real images from forged images, so currently supervised image classification models can be used for classification. Common image classification models include VGG, XceptionNet, EfficientNet, etc. These models are trained based on the data of the face forgery sample library, and the accuracy of the test in the library is considerable. These supervised models have the limitations of difficult to obtain training samples and the fixed way of generating training samples, which leads to inaccurate detection of samples generated by non-training samples by supervised models.

[0003] In view of the limitations of the above-mentioned supervised detection model, an unsupervised approach is used for model training. The biggest difference between unsupervised and supervised detection models is that they do not use labels to constrain the output results, but cluster the given data through the learning ability of the network, where the accuracy of clustering depends on the network's own feature extraction ability and the discrimination of the samples. Among them, the single-classification model is a more common anomaly detection method in unsupervised algorithms, that is, when only positive samples are input, they are clustered into one category. If there are samples that are not within the minimum boundary of the class, they are considered abnormal samples.

[0004] In the field of face generation detection, unsupervised models can be used as anomaly detection models. When detecting faces, the detector pays more attention to distinguishing true and false faces, and it is easier to obtain positive samples. We can choose to train only positive samples and perform clustering training on positive samples. In this way, when abnormal samples (false samples) are detected, their feature expressions are not within the expression range of normal samples, and the authenticity of the samples can be distinguished. This method is not targeted at a specific generation method. As long as the difference between the generation method and the real sample is extracted, the method can be detected. This solves the classification limitations of supervised models.

[0005] However, the existing unsupervised models cannot be directly applied to face generation detection and classification. This is because the current unsupervised detection models are generally applied to image classifications with obvious differences, such as handwriting recognition. The features of these classifications are relatively easy to extract, and their features are more obvious. This makes the latent vectors reconstructed by the reconstruction method inconsistent when using a single-classification network for classification, because the reconstruction method uses a method similar to PCA to perform nonlinear dimensionality reduction on the input data to obtain a representative vector. Due to the large differences in the content of images of different categories, the feature vectors after dimensionality reduction also have obvious differences.

[0006] For generated faces and natural faces, their main contents are similar, and their basic structures are both faces, facial features, and facial expressions. Even for forged faces, due to the development of neural networks, the current generation methods can often be indistinguishable from the real ones in the eyes of the human eye, which shows that in terms of basic features, the difference between forged faces and natural image faces is not much. In fact, the differences in generated faces are mostly concentrated in high-order features, namely texture inconsistency, facial blur, texture mismatch, etc., which do not affect the content of the face image itself. In addition, since the generated faces are mainly based on subtle differences in small areas and appear in uncertain positions, methods such as introducing attention mechanisms to enhance network learning are not applicable. If the reconstruction method is used for training, since the network is more sensitive to the content of the image, it will ignore subtle differences during dimensionality reduction, which makes the latent vector not contain real difference features, resulting in classification failure. The direct learning method is similar to it. In image classification, the face itself is classified into one category. In the absence of prior knowledge, it is difficult for the network to grasp subtle features for analysis. This limits the application of single classification methods in the field of face generation and detection. Summary of the invention

[0007] In view of the above, the purpose of the present invention is to provide an unsupervised face forgery assessment method, which reconstructs face images through a generative model trained in an unsupervised manner, performs forgery assessment based on the difference between the original face image and the reconstructed face image, and improves the accuracy of face assessment.

[0008] In order to achieve the above-mentioned invention object, the present invention provides the following technical solutions:

[0009] An unsupervised face forgery assessment method, comprising the following steps:

[0010] Obtain a data set of people to be protected, take a face image of each person as the original face image, other face images of the same person as related face images, and take image pairs of the original face image and the related face image as samples;

[0011] Constructing a reconstruction model including a coarse-grained feature extraction module, a fine-grained feature generation module and a reconstruction module, wherein the coarse-grained feature extraction module includes two branches, which are respectively used to extract facial action features and image color features of a face image; the fine-grained feature generation module is used to generate fine-grained features of a specified person; and the reconstruction module is used to generate a reconstructed face image by combining facial action features, image color features and fine-grained features;

[0012] Construct a loss function including generation loss and single classification loss, wherein the generation loss includes reconstruction loss of the reconstructed face image and the original face image, and perceptual loss based on enhanced features of the reconstructed face image and the original face image; the single classification loss includes DSVDD loss based on fine-grained features of the original face image and related face images, and cosine similarity loss based on facial action features and image color features of the original face image;

[0013] The reconstruction model is used to reconstruct the face image to be tested to obtain a reconstructed face image, and the difference between the reconstructed face image and the face image to be tested is compared to realize face forgery assessment.

[0014] In one embodiment, each branch of the coarse-grained feature extraction module uses MnasNet, the MnasNet of one branch extracts facial action features of a face image, and the MnasNet of another branch extracts image color features.

[0015] In one embodiment, the fine-grained feature generation module uses a multi-layer perceptron structure to generate fine-grained features from a given person identification code.

[0016] In one embodiment, the reconstruction module includes two feature untangling modules of two branches and a reconstruction decoder, wherein the two feature decoupling modules of the two branches perform feature untangling on the image color features and the facial action features respectively, and the feature reconstruction encoder uses the facial coarse-grained features after feature untangling to first fuse with the facial fine-grained features, and the obtained first fused feature features are then fused with the image color features processed by the feature untangling, and the obtained second fused features are mapped to generate a reconstructed face image.

[0017] In one embodiment, the reconstruction module and the fine-grained feature generation module are implemented using a transposed convolution upsampling layer, and each feature fusion uses an AdaIN structure to perform style affine transformation.

[0018] In one embodiment, the reconstruction loss adopts an MSE loss constructed based on the reconstructed face image and the original face image;

[0019] The perceptual loss includes the perceptual loss of the style features of the original face image and the style features of the reconstructed face image, and the perceptual loss of the content features of the original face image and the content features of the reconstructed face image;

[0020] The DSVDD loss for: The corresponding clustered DSVDD objective function is: Among them, (x i ) j and (x i ) kdenote the original face image and the related face image of the i-th person respectively. The subscripts j and k are used as the indexes of the face images to distinguish the original face image and the related face image. Φ F (·) represents the fine-grained features generated by the fine-grained feature generation module, represents the L2 distance, ||·|| 2 represents the Euclidean distance, n is the total number of face images of the i-th person, and Respectively represent (x i ) j and (x i ) k The corresponding fine-grained features, λ represents the adjustment parameter, W l represents the matrix formed by the weight parameter w of the lth layer in the fine-grained feature generation module, L represents the total number of layers, Represents the L2 norm of the network weight, and minimizes this term in the formula, that is, finds the minimum clustering boundary that satisfies the sample;

[0021] The cosine similarity loss for: The corresponding objective function is: Among them, Φ m ((x i ) j ) and Φ S ((x i ) j ) respectively represent (x i ) j facial action features and image color features.

[0022] In one embodiment, reconstructing the face image to be tested by using the reconstruction model to obtain the reconstructed face image includes:

[0023] The coarse-grained feature extraction module is used to extract the facial action features and image color features of the face image to be tested, the fine-grained feature generation module is used to generate the idealized fine-grained features of the person to be tested, and the reconstruction module is used to reconstruct the face image based on the facial action features, the image color features and the idealized fine-grained features to obtain a reconstructed face image of the face image to be tested;

[0024] The idealized fine-grained features are the average values ​​of the fine-grained features generated by the fine-grained feature generation module for all the facial images of the person in the person data set.

[0025] In one embodiment, the difference between the reconstructed face image and the face image to be tested is used to implement face forgery assessment, including:

[0026] When the difference between the reconstructed face image and the face image to be tested is greater than a preset threshold, the face image to be tested is considered to be a forged face image.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The facial action features, image color features and fine-grained features of different facial images of a single person are extracted to construct a reconstructed facial image of the person. On this basis, the reconstruction loss of the reconstructed facial image and the original facial image, the perceptual loss based on the reconstructed facial image and the original facial image, the DSVDD loss based on the fine-grained features of the original facial image and the related facial images, and the cosine similarity loss based on the facial action features and image color features of the original facial image are combined to jointly train the coarse-grained feature extraction module, the fine-grained feature generation module and the reconstruction module, covering all elements of the facial image and improving the accuracy of the facial image reconstructed by the reconstruction model. Based on the difference between the accurate reconstructed facial image and the original facial image, the face forgery assessment is realized. The process adopts an unsupervised learning method and combines the single classification method of clustering, so that the constructed reconstruction model is suitable for the assessment of forged faces generated by any generation method, with a wide range of applications and universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 creative work.

[0030] Figure 1 is a flow chart of an unsupervised face forgery assessment method provided by an embodiment;

[0031] Figure 2 is a schematic diagram of the structure of a reconstruction system provided in an embodiment;

[0032] Figure 3 is a schematic diagram of the structure of a reconstruction decoder provided in an embodiment;

[0033] Figure 4 The embodiment provides Figure 3 Schematic diagram of the structure of the upsampling layer of the reconstruction decoder; DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0035] In order to detect forged face images generated in various ways, an embodiment provides an unsupervised face forgery assessment method. Figure 1 : is a flowchart of the unsupervised face forgery assessment method provided by the embodiment. Figure 1 As shown, the unsupervised face forgery assessment method provided by the embodiment includes the following steps:

[0036] Step 1: Obtain a character dataset and construct training samples.

[0037] The character data set used includes facial images of each character at multiple times and in multiple states. In the embodiment, to construct a reconstruction model for face reconstruction, any facial image of the character is extracted as an original facial image, other facial images of the character are extracted as related facial images, and an image pair of the original facial image and the related facial image is used as a training sample.

[0038] Step 2: construct a reconstruction model including a coarse-grained feature extraction module, a fine-grained feature generation module and a reconstruction module.

[0039] Figure 2 Schematic diagram of the structure of the reconstruction model provided in the embodiment. Figure 2 As shown in the figure, the coarse-grained feature extraction module includes two branches, which are used to extract facial action features (including parameters such as character expression and head orientation) and image color features (including parameters such as brightness and hue) of face images; the fine-grained feature generation module is used to generate fine-grained features of a specified character; the reconstruction module is used to generate a reconstructed face image by combining facial action features, image color features, and fine-grained features;

[0040] like Figure 2 As shown, each branch network of the coarse-grained feature extraction module adopts MnasNet, which is represented by Φ m and Φ S The fine-grained feature generation module converts the identity code of a given person into the fine-grained features of the person, represented as Φ F Since the fine-grained features of face images have consistent characteristics in different face images of people, in order to ensure that the extracted features meet the above characteristics, the training network Φ F The input is compared within the sample of the same person, requiring the branch network Φ m and Φ S The input is the original face image and the input is the same, and the output is the original face image x i Facial action features and image color features The image color features Including image brightness, hue, etc. At the same time, the branch network Φ FThe output is the related face image x of the same training sample i Fine-grained features

[0041] like Figure 2 As shown in Figure 2, the reconstruction module includes two feature disentanglement modules of two branches and a reconstruction decoder. The two feature coupling modules of the two branches both use fully connected layers (FCL). FCL is similar to the mapping network in Style-GAN. and Perform feature unwrapping.

[0042] The input to the reconstruction decoder consists of and like Figure 3 As shown, the reconstruction decoder uses multiple upsampling layers, where the first upsampling layer converts facial action features into The first image feature is obtained by dimension change, and the first image feature is input to the second upsampling layer. The second upsampling layer disentangles the image feature of the first layer from the fine-grained feature processed by the FCL feature. The corresponding dimension features are fused to obtain the second fused features. The second fused features can be obtained by fusion of features of multiple second upsampling layers. The processing objects of each feature fusion are the output of the previous upsampling layer and the fine-grained feature fusion of each FCL feature disentanglement processing. Figure 2 The second fusion feature is processed by four second upsampling layers. The second fusion feature is input to the third upsampling layer, and the third upsampling layer uses the coarse-grained feature as a weight parameter and the image color feature processed by feature unwrapping. The third fused feature is obtained by fusion, and the third fused feature is mapped through a convolutional layer to generate a reconstructed face image.

[0043] Figure 4 The embodiment provides Figure 3 Schematic diagram of the structure of the upsampling layer of the reconstruction decoder. Figure 4 As shown in the figure, the first, second and third upsampling layers all adopt a combination of deconvolution layer (Conv Transpose), convolution layer (Conv) and AdaIN, where the deconvolution layer is used to extract image super-resolution, and the fine-grained features are regarded as image style features. The affine transformation of AdaIN is used to enable the fine-grained features to be combined with the incoming features.

[0044] Different from the description of the effect of network layer size on face generation content in Style-GAN, fine-grained features Not only does it exist in the fine-grained features it defines, such as the size of the eyes and their position in the face, which are also unique to a specific person, so the AdaIN structure is used in all upsampling layers, such as Figure 4 As shown in the modeling It should be consistent and a constant. Due to differences in lighting of faces in the character dataset, these are also part of the style but should not be counted. According to the theory of Style-GAN, subtle features such as color belong to fine-grained features, so the input of the last layer of the AdaIN structure is designed to be the image style feature. The design of the reconstruction module not only corresponds to the function of face modeling, but also takes into account the necessary processes in the reconstruction of the image in practice. The reconstruction training task includes two tasks: one is to reconstruct the original face image, and the other is to ensure that the FCL In order to converge to a constant vector as much as possible, both the single-class reconstruction model and the one-class DSVDD model are used. This design separates the coarse-grained features from the fine-grained features and uses the DSVDD model only for the fine-grained features.

[0045] Step 3: Construct the loss function of the reconstruction model and train the reconstruction model.

[0046] Combined with the reconstruction model constructed in step 2, the constructed loss function includes two parts, namely, generation loss and single classification loss. Among them, the generation loss includes the reconstruction loss of the reconstructed face image and the original face image, the perceptual loss based on the enhanced features of the reconstructed face image and the original face image, and the single classification loss includes the DSVDD loss based on the fine-grained features of the original face image and the related face images, and the cosine similarity loss based on the facial action features of the original face image and the image color features.

[0047] Among them, the reconstruction loss The MSE (mean-square error) loss constructed based on the reconstructed face image and the original face image is adopted.

[0048] Perceived loss It includes the perceptual loss of the style features of the original face image and the style features of the reconstructed face image, and the perceptual loss of the content features of the original face image and the content features of the reconstructed face image. Reconstruction loss and perceived loss It ensures that the difference between the reconstructed face image and the original face image can be minimized.

[0049] For single-classification loss, the single-classification objective function of DSVDD is prepared as the loss function. However, in general single-classification, the center of the hypersphere is generally a constant. The method for determining this constant is generally to pass all samples through the network that has only been initialized, and calculate the mean obtained through the network. This will cause a problem. The description of the ideal vector represented by this center value is unknown. It is likely that the common features or meaningless features of normal samples and abnormal samples are taken, which will cause classification failure in the subsequent abnormality detection. Some fine-grained normal samples require a supervised model to converge. Therefore, the loss function that converges to the constant c is improved in DSVDD. The embodiment draws on the idea of ​​clustering, but unlike other clustering methods, all training samples should be classified into the same category. It is hoped that this ideal feature can be a semantic feature, and the feature vector obtained by each training sample through the network has the smallest intra-class spacing. In the process of training clustering DSVDD, since the reconstruction model needs to use this latent vector as the input of the style feature, it also has the semantic features of the style in the clustering process. This ensures that the features can be gathered in a hypersphere and that the vectors in the sphere have a certain meaning. Therefore, the designed clustering form of DSVDD objective function is as follows:

[0050]

[0051] in, represents the L2 distance, ||·|| 2 represents the Euclidean distance, n is the total number of face images of the i-th person, and Respectively represent (x i ) j and (x i ) k The corresponding fine-grained features, the second term of the sum is the regularization term of the weight matrix W of the network layer l∈L, λ represents the regularization attenuation term parameter, W l represents the matrix formed by the weight parameter w of the lth layer in the fine-grained feature generation module, L represents the total number of layers, It means that the weight is the smallest under the L2 norm, that is, finding the minimum cluster boundary;.

[0052] DSVDD loss for:

[0053]

[0054] Among them, (x i ) j and (x i ) kdenote the original face image and the related face image of the i-th person respectively. The subscripts j and k are used as the indexes of the face images to distinguish the original face image and the related face image. Φ F (·) represents the fine-grained features generated by the fine-grained feature generation module.

[0055] In the embodiment, facial action features and image color features are required to be independent of each other. However, the input of the two branch networks is the same image; and although the two feature extraction networks are independent of each other, their network structures are consistent. In order to avoid confusing common features in the two latent vectors, the cosine similarity loss is introduced, which can be used to calculate the similarity between facial action features and image color features. Therefore, the objective function and the corresponding cosine similarity loss function are for:

[0056]

[0057]

[0058] Among them, Φ m ((x i ) j ) and Φ S ((x i ) j ) respectively represent (x i ) j facial action features and image color features.

[0059] Another advantage of the cosine similarity loss function is that it establishes a relationship between facial movements and image colors. This means that if the image has color mismatches due to artifacts, etc., the loss function can affect the relationship between facial movement features and image colors, resulting in greater facial reconstruction disturbances, resulting in obvious facial disharmony in the reconstructed face image.

[0060] In summary, the loss function of the reconstruction model is expressed as:

[0061]

[0062] Among them, α, β, γ, ρ are hyperparameters in the loss function.

[0063] When training the reconstruction model, the training samples constructed in step 1 are used, and the constructed loss function is adopted to optimize the reconstruction model parameters.

[0064] Step 4: Reconstruct the face image to be tested using the reconstruction model to obtain a reconstructed face image, and compare the difference between the reconstructed face image and the face image to be tested to implement face forgery assessment.

[0065] In the embodiment, all training samples are transmitted through the network to obtain a fine-grained feature set, and the average value of the fine-grained feature set is calculated as an idealized fine-grained feature. When applied, for each face image to be tested, the facial action features of the face image to be tested are generated using the fine-grained feature generation module, and the reconstruction module is used to reconstruct based on the facial action features and the idealized fine-grained features to obtain a reconstructed face image of the face image to be tested.

[0066] Then, the difference between the reconstructed face image and the face image to be tested is calculated. The embodiment uses the structural similarity (SSIM) value as the difference between the reconstructed face image and the face image to be tested, and judges whether it is a forged face image based on the difference. Specifically, when the difference is greater than a preset threshold, the face image to be tested is considered to be a forged face image.

[0067] The unsupervised face forgery assessment method (OC-pDet for short) provided in the embodiment is compared with the current newer single classification methods (One-class DSVDD, Soft-boundary DSVDD, OC-FakeDect-2) on two public datasets, DeepFakeDetection and Celeb-DF. The comparison results are shown in Table 1. It can be seen from Table 1 that the detection effect is greatly improved compared with the existing methods.

[0068] Table 1

[0069]

[0070] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An unsupervised face forgery assessment method, characterized in that: The following steps are involved: Obtain a data set of people to be protected, take a face image of each person as the original face image, other face images of the same person as related face images, and take image pairs of the original face image and the related face image as samples; Constructing a reconstruction model including a coarse-grained feature extraction module, a fine-grained feature generation module and a reconstruction module, wherein the coarse-grained feature extraction module includes two branches, which are respectively used to extract facial action features and image color features of a face image, the fine-grained feature generation module is used to generate fine-grained features of a specified person, and the reconstruction module is used to combine facial action features, image color features and fine-grained features to generate a reconstructed face image; Construct a loss function including generation loss and single classification loss, wherein the generation loss includes reconstruction loss of the reconstructed face image and the original face image, and perceptual loss based on enhanced features of the reconstructed face image and the original face image; the single classification loss includes DSVDD loss based on fine-grained features of the original face image and related face images, and cosine similarity loss based on facial action features and image color features of the original face image; The reconstruction model is used to reconstruct the face image to be tested to obtain a reconstructed face image, and the difference between the reconstructed face image and the face image to be tested is compared to realize face forgery assessment.

2. The unsupervised face forgery assessment method according to claim 1, characterized in that: Each branch of the coarse-grained feature extraction module adopts MnasNet, the MnasNet of one branch extracts facial action features of face images, and the MnasNet of another branch extracts image color features.

3. The unsupervised face forgery assessment method according to claim 1, characterized in that: The fine-grained feature generation module adopts a multi-layer perceptron structure to generate fine-grained features from a given person identification code.

4. The unsupervised face forgery assessment method according to claim 1, characterized in that: The reconstruction module includes two feature untangling modules of two branches and a reconstruction decoder, wherein the two feature decoupling modules of the two branches respectively untangle the image color features and the facial action features, and the feature reconstruction encoder uses the facial coarse-grained features after feature untangle to first fuse with the facial fine-grained features, and then fuses the obtained first fused feature features with the image color features processed by feature untangle, and the obtained second fused features are mapped to generate a reconstructed face image.

5. The unsupervised face forgery assessment method according to claim 3, characterized in that: The reconstruction module and the fine-grained feature generation module are implemented using a transposed convolution upsampling layer, and each feature fusion uses an AdaIN structure to perform style affine transformation.

6. The unsupervised face forgery assessment method according to claim 3, characterized in that: The feature disentanglement module adopts a fully connected layer.

7. The unsupervised face forgery assessment method according to claim 1, characterized in that: The reconstruction loss adopts the MSE loss constructed based on the reconstructed face image and the original face image; The perceptual loss includes the perceptual loss of the style features of the original face image and the style features of the reconstructed face image, and the perceptual loss of the content features of the original face image and the content features of the reconstructed face image; The DSVDD loss for: The corresponding clustered DSVDD objective function is: Among them, (x i ) j and (x i ) k denote the original face image and the related face image of the i-th person respectively. The subscripts j and k are used as the indexes of the face images to distinguish the original face image and the related face image. Φ F (·) represents the fine-grained features generated by the fine-grained feature generation module, represents the L2 distance, ||·|| 2 represents the Euclidean distance, n is the total number of face images of the i-th person, and Respectively represent (x i ) j and (x i ) k The corresponding fine-grained features, λ represents the adjustment parameter, W l represents the matrix formed by the weight parameter w of the lth layer in the fine-grained feature generation module, L represents the total number of layers, Represents the L2 norm of the weight parameter; The cosine similarity loss for: The corresponding objective function is: Among them, Φ m ((x i ) j ) and Φ S ((x i ) j ) respectively represent (x i ) j facial action features and image color features.

8. The unsupervised face forgery assessment method according to claim 1, characterized in that: The method of reconstructing the face image to be tested by using the reconstruction model to obtain the reconstructed face image comprises: The coarse-grained feature extraction module is used to extract the facial action features and image color features of the face image to be tested, the fine-grained feature generation module is used to generate the idealized fine-grained features of the person to be tested, and the reconstruction module is used to reconstruct the face image based on the facial action features, the image color features and the idealized fine-grained features to obtain a reconstructed face image of the face image to be tested; The idealized fine-grained features are the average values ​​of the fine-grained features generated by the fine-grained feature generation module for all the facial images of the person in the person data set.

9. The unsupervised face forgery assessment method according to claim 1, characterized in that: The difference between the reconstructed face image and the face image to be tested realizes face forgery assessment, including: When the difference between the reconstructed face image and the face image to be tested is greater than a preset threshold, the face image to be tested is considered to be a forged face image.

Images