Unsupervised deeply-forged face video detection method based on strong enhancement contrast learning

Abstract

The invention discloses an unsupervised deeply-forged face video detection method based on strong enhancement contrast learning, which comprises the following steps: firstly, preprocessing face video data, cutting out a full-size label-free image and a face region image, then preliminarily classifying the label-free data and endowing the label-free data with pseudo labels by using natural physiological characteristics in face features; a strong enhancement contrast learning strategy is adopted, a plurality of basic enhancement operations are randomly combined to generate a strong enhancement view with larger structure change and semantic difference, strong enhancement contrast learning of the face image is completed, and finally accurate detection of a counterfeit face video is realized through diclustering and authenticity discrimination. Compared with an existing unsupervised detection method, the method has the advantages that multi-level feature fusion and strong enhancement contrast learning are combined, deep counterfeit traces can be captured more effectively, higher robustness and generalization ability are shown in complex scenes such as noise and compression, and the method is high in robustness and generalization performance. And high-precision detection of the deeply-forged face video under an unsupervised condition is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of computer vision and deep learning, and specifically relates to an unsupervised deep fake face video detection method based on strongly enhanced contrast learning. Background Art

[0002] With the rapid development of mobile communications technology and the widespread adoption of smart devices, images and videos have gradually replaced text as the primary means for people to access information. In this era of digitalization, the rise of social media and various digital platforms has greatly facilitated the sharing and access of digital media content. Simultaneously, the rapid growth of the internet and improvements in computer performance have provided the foundation for the rapid development of deep learning and computer vision. Against this backdrop, deepfake technology, which combines computer vision and deep learning, has also rapidly gained popularity.

[0003] However, the misuse of deepfake technology has also raised numerous social and ethical issues, particularly regarding the spread of false information and the violation of personal privacy. The primary danger of deepfake technology is its facilitation of telecommunications fraud and its serious threat to public property safety. Deepfakes are also commonly used in pornographic photos and films, replacing the faces of protagonists with those of unrelated individuals, infringing on the privacy and reputation of the individuals involved.

[0004] Early deepfake detection methods focused on clues left by facial manipulation techniques and made great progress thanks to supervised learning. However, supervised learning requires a large number of labels, and labeling is time-consuming and tedious, which is often difficult to achieve in reality, making the training process complex and costly. Unsupervised deepfake detection does not require the use of labeled data, thereby reducing the time and cost of data preparation. Although existing unsupervised detection methods do not require massive amounts of labeled data, they often rely on weakly enhanced contrastive learning strategies, which have difficulty capturing deeper and more subtle forgery traces and perform poorly in noisy or compressed scenarios. Therefore, there is an urgent need for an unsupervised detection technology that can integrate richer forms of data augmentation and is noise-robust. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides an unsupervised deep fake face video detection method based on strong enhancement contrastive learning. On the basis of pseudo label generation, the "strong enhancement + distribution divergence minimization" strategy is introduced to explore more subtle fake features and improve detection robustness.

[0006] The technical solution adopted by the present invention is: an unsupervised deep fake face video detection method based on strong enhancement contrast learning, the specific steps are as follows:

[0007] S1. Input unlabeled deep fake face video and real face video data, perform preprocessing, and divide them into training set and test set;

[0008] A multi-task cascade convolutional neural network (MTCNN) model is used to detect faces in the input video, and the detected face area is cropped out to generate full-size unlabeled images and face area images. Finally, the face video is divided into training and test sets according to the actual ratio.

[0009] S2. Extract features from the full-size unlabeled images and face area image data in the training set obtained in step S1, fuse the extracted features, perform preliminary clustering, and generate pseudo labels by assigning pseudo labels;

[0010] First, the visual artifact feature VAF and local texture feature LTF are extracted simultaneously from the full-size unlabeled image.

[0011] Among them, VAF is based on high-frequency residual analysis and extraction. It captures visual artifacts introduced in the deep fake process by converting the image from the spatial domain to the frequency domain and analyzing abnormal patterns in the spectrum; LTF extracts the texture features of the image by combining local binary patterns and Gabor filter groups.

[0012] The extracted VAF and LTF features are then fused to construct a hybrid feature space. The fusion method is to first standardize the two types of features and then splice them in the feature dimension. In the hybrid feature space, the K-means++ algorithm is used to perform preliminary clustering on the samples, dividing them into two categories.

[0013] Finally, based on the distance between the sample and the cluster center, the samples are further divided into three categories: high-confidence samples, boundary samples, and outliers. Corresponding label assignment strategies are adopted for different types of samples to generate reliability pseudo labels, and the same pseudo labels are assigned to the face area images in the same frame as the full-size unlabeled image.

[0014] Among them, for high-confidence samples, the pseudo labels of the corresponding categories are directly assigned; for boundary samples, the relative distance ratio between them and the cluster centers is calculated, and the threshold is set to assign pseudo labels; for outliers, no pseudo labels are assigned for the time being.

[0015] S3. By randomly combining multiple weak and strong enhancement operations on the pseudo-labeled face region images in the training set, an enhanced view with structural deformation and semantic difference is constructed. Then, a contrastive learning network architecture including a double-layer projection head is adopted, and a distribution divergence minimization (DDM) strategy is introduced to build a confidence sample selection module. High-confidence samples are dynamically selected to participate in the training of the contrastive learning network, completing the strong enhancement contrastive learning of face images.

[0016] S4. Use the backbone network of the contrastive learning network trained in step S3 to extract the features of the data samples in the test set divided in step S1, and perform binary clustering based on the K-means++ algorithm. Finally, use the inter-frame correlation to distinguish the authenticity of the classification results, thereby achieving accurate detection of forged face videos.

[0017] Furthermore, the step S3 is specifically as follows:

[0018] First, weak enhancements T and T' are applied simultaneously to the pseudo-label face region image X in the training set to generate weakly enhanced views V and V'; and strong enhancement transformation S is applied to image X to generate a strongly enhanced view V". V and V' are converted using the key-value encoder f θ Get its embedded representation z = f θ (V) and z ′ =f θ (V ′ ). At the same time, V” is passed through the query encoder Get the corresponding embedding representation Finally, each batch will add z into the queue to form a group including z1, z2, ..., z k , a representation library for z.

[0019] Among them, weak enhancement operations include: random cropping, horizontal flip, color jitter and Gaussian blur; strong enhancement is generated by randomly combining multiple basic transformation operations, including: shear X(Y) (-0.3, 0.3), translation X(Y) (-0.3, 0.3), rotation (-30, 30), auto contrast (0 or 1), inversion (0 or 1), equalization (0 or 1), solarization (0, 256), color separation (4, 8), contrast (0.05, 0.95), brightness (0.05, 0.95), sharpness (0.05, 0.95).

[0020] Where X(Y) represents the vertical and horizontal directions, and the numerical value represents the operation amplitude.

[0021] Then construct a contrastive learning network architecture including a double-layer projection head. The first projection head g u By reducing the dimension to eliminate feature redundancy, the second projection head g f Perform comparative spatial mapping.

[0022] Among them, the double-layer projection head consists of two fully connected layers and a normalization layer connected in series. The first layer is a 2048-dimensional hidden layer, followed by a ReLU activation function, and the second layer is the output layer without an activation function. The former reduces the dimension of the output, and the latter maps it to a low-dimensional space that is friendly to contrastive learning.

[0023] The loss function used by the dual-layer projection head is The calculation expression is as follows:

[0024]

[0025] in, represents the mathematical expectation of the sample index i in a batch B; K represents the total number of key embeddings; q(z k ∣z i ′ ) represents the query embedding z i ′ Embed z with any key k The conditional distribution between q(z i ∣z i ′ ) represents the ideal target distribution; when j = i, p(z i ∣z i ′ ) represents the likelihood of the positive sample pair; when j = k and k ≠ i, it represents the relative negative likelihood between the query and the negative sample; τ represents the temperature parameter.

[0026] Then, the distribution divergence minimization DDM strategy is introduced to calculate the similarity of the weakly enhanced view V' and the strongly enhanced view V" in all samples in the representation library, respectively, to obtain the similarity vector. Then, the KL divergence between the two similarity vectors is calculated as a measure of the distribution difference, and the divergence is minimized as the loss function. The expression is as follows:

[0027]

[0028] The total loss function is defined as the weighted sum of the standard contrast loss and the distribution divergence minimization loss, and the expression is as follows:

[0029]

[0030] Among them, β∈[0,1] represents the weighting coefficient.

[0031] Finally, a confidence sample selection module is constructed based on the feature space distribution density. By calculating the cosine similarity between the sample and the cluster center, an ascending queue is constructed, and high-confidence samples are dynamically selected to participate in the training of the comparative learning network.

[0032] Furthermore, the step S4 is specifically as follows:

[0033] First, feature extraction is performed on the test set data samples. This involves using the backbone network trained in step S3 to extract discriminant features for the test samples. The facial region images from the test video are fed into the backbone network to obtain their representations in feature space. Based on the extracted feature representations, the K-means++ algorithm is then used for binary clustering, dividing the test samples into two categories.

[0034] Then, inter-frame correlation analysis is introduced by calculating the Spearman correlation coefficient ρ between consecutive frames.l , to distinguish the true from the false of the clustering results, that is, for the samples in each category, calculate the average value of the Spearman correlation coefficient between the features of adjacent frames The expression is as follows:

[0035]

[0036] Among them, ρ l represents the correlation between the lth frame and the l+1th frame; f l,n represents the nth feature of the lth frame; L represents the total number of frames in the video clip, and N represents the total number of features in each frame.

[0037] Ultimately, the category with a lower correlation coefficient is judged to be a real video, and the category with a higher correlation coefficient is judged to be a fake face video.

[0038] The beneficial effects of the present invention are as follows: the method of the present invention first pre-processes the facial video data, crops out full-size unlabeled images and facial area images, and then uses the natural physiological characteristics of facial features to preliminarily classify the unlabeled data and assign pseudo-labels, and adopts a strong enhancement contrast learning strategy to generate a strong enhancement view with greater structural changes and semantic differences by randomly combining multiple basic enhancement operations, thereby breaking through the existing contrast learning's dependence on weak transformations, mining potential and subtle forgery traces in images, and completing the strong enhancement contrast learning of facial images. Finally, accurate detection of forged facial videos is achieved through binary clustering and authenticity discrimination, that is, based on a well-trained feature extractor and the distribution law of the feature space, forged samples and real samples are clustered, and the clustering results are further authenticated by calculating the facial motion trajectory and micro-expression change pattern between adjacent frames. Compared with existing unsupervised detection methods, the method of the present invention can more effectively capture deep fake traces by combining multi-level feature fusion with strongly enhanced contrast learning, and exhibit stronger robustness and generalization ability in complex scenarios such as noise and compression, thereby achieving high-precision detection of deep fake face videos under unsupervised conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flowchart of an unsupervised deep fake face video detection method based on strongly enhanced contrast learning of the present invention.

[0040] Figure 2 This is a framework diagram of the pseudo-label generation stage in an embodiment of the present invention.

[0041] Figure 3 This is a framework diagram of the strong enhanced contrast learning stage in an embodiment of the present invention.

[0042] Figure 4 This is a framework diagram of the binary clustering and authenticity discrimination stage in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The method of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0044] like Figure 1 As shown in the figure, the flow chart of the unsupervised deep fake face video detection method based on strong enhanced contrast learning of the present invention is as follows:

[0045] S1. Input unlabeled deep fake face video and real face video data, perform preprocessing, and divide them into training set and test set;

[0046] A multi-task cascaded convolutional neural network (MTCNN) model is used to detect faces in the input video, and the detected face area (key frame) is cropped out to generate full-size unlabeled images and face area images. Finally, the face video is divided into training and test sets according to the actual ratio.

[0047] S2. Extract features from the full-size unlabeled images and face area image data in the training set obtained in step S1, fuse the extracted features, perform preliminary clustering, and generate pseudo labels by assigning pseudo labels;

[0048] like Figure 2 As shown in FIG, first, the visual artifact feature (VAF) and the local texture feature (LTF) are extracted simultaneously from the full-size unlabeled image.

[0049] VAF is based on high-frequency residual analysis and extraction. It converts images from the spatial domain to the frequency domain and analyzes abnormal patterns in the spectrum to capture visual artifacts introduced during deepfakes. LTF, on the other hand, extracts texture features of images by combining local binary patterns with Gabor filter banks. These features can reflect the consistency and subtle changes in facial skin texture.

[0050] The extracted VAF and LTF features are then fused to construct a hybrid feature space. The fusion method is to first standardize the two types of features and then splice them in the feature dimension. In the hybrid feature space, the K-means++ algorithm is used to perform preliminary clustering on the samples, dividing them into two categories.

[0051] Finally, based on the distance between the sample and the cluster center, the samples are further divided into three categories: high-confidence samples, boundary samples, and outliers. Corresponding label assignment strategies are adopted for different types of samples to generate reliability pseudo labels 0 and 1, and the same pseudo label is assigned to the face area image in the same frame as the full-size unlabeled image.

[0052] High-confidence samples (those close to their respective cluster centers) are directly assigned pseudo labels of the corresponding category. For boundary samples (those located at the boundary between two categories), the ratio of their relative distances to the cluster centers is calculated, and a threshold is set to assign pseudo labels. For outliers (those far from all cluster centers), no pseudo labels are assigned. This strategy ensures the reliability of pseudo labels.

[0053] S3. By randomly combining multiple weak and strong enhancement operations on the pseudo-labeled face region images in the training set, an enhanced view with structural deformation and semantic difference is constructed. Then, a contrastive learning network architecture including a double-layer projection head is adopted, and a distribution divergence minimization (DDM) strategy is introduced to build a confidence sample selection module. High-confidence samples are dynamically selected to participate in the training of the contrastive learning network, completing the strong enhancement contrastive learning of face images.

[0054] S4. Use the backbone network of the contrastive learning network trained in step S3 to extract the features of the data samples in the test set divided in step S1, and perform binary clustering based on the K-means++ algorithm. Finally, use the inter-frame correlation to distinguish the authenticity of the classification results, thereby achieving accurate detection of forged face videos.

[0055] like Figure 3 As shown, in this embodiment, the step S3 is specifically as follows:

[0056] First, weak enhancements T and T' are applied simultaneously to the pseudo-label face region image X in the training set to generate weakly enhanced views V and V'; and strong enhancement transformation S is applied to image X to generate a strongly enhanced view V". V and V' are converted using the key-value encoder f θ Get its embedded representation z = f θ (V) and z ′ =f θ (V ′ ). At the same time, V” is passed through the query encoder Get the corresponding embedding representation Finally, each batch will add z into the queue to form a group including z1, z2, ..., z k , a representation library of z. This strong enhancement strategy can produce views with greater structural deformation and semantic differences, which helps to expose deeper and more subtle forgery traces.

[0057] Among them, weak enhancement operations include: random cropping, horizontal flip, color jitter and Gaussian blur; strong enhancement is generated by randomly combining multiple basic transformation operations, including: shear X(Y) (-0.3, 0.3), translation X(Y) (-0.3, 0.3), rotation (-30, 30), auto contrast (0 or 1), inversion (0 or 1), equalization (0 or 1), solarization (0, 256), color separation (4, 8), contrast (0.05, 0.95), brightness (0.05, 0.95), sharpness (0.05, 0.95).

[0058] Where X(Y) represents the vertical and horizontal directions, and the numerical value represents the operation amplitude.

[0059] Then construct a contrastive learning network architecture including a double-layer projection head. The first projection head g u By reducing the dimension to eliminate feature redundancy, the second projection head g f Perform comparative spatial mapping.

[0060] Among them, the double-layer projection head consists of two fully connected layers and a normalization layer connected in series. The first layer is a 2048-dimensional hidden layer, followed by a ReLU activation function, and the second layer is the output layer without an activation function. The former reduces the dimension of the output, and the latter maps it to a low-dimensional space that is friendly to contrastive learning.

[0061] The loss function used by the dual-layer projection head is The calculation expression is as follows:

[0062]

[0063] in, represents the mathematical expectation of the sample index i in a batch B; K represents the total number of key embeddings; q(z k ∣z i ′ ) represents the query embedding z i ′ Embed z with any key k The conditional distribution between q(z i ∣z i ′ ) represents the ideal target distribution; when j = i, p(z i ∣z i ′ ) represents the likelihood of the positive sample pair; when j = k and k ≠ i, it represents the relative negative likelihood between the query and the negative sample; τ represents the temperature parameter.

[0064] Then, a distribution divergence minimization (DDM) strategy is introduced to align the similarity distributions of the weakly enhanced and strongly enhanced views in the representation library. The similarities of all samples of the weakly enhanced view V' and the strongly enhanced view V" in the representation library are calculated respectively to obtain similarity vectors. The KL divergence between the two similarity vectors is then calculated as a measure of the distribution difference, and the divergence is minimized as a loss function. This allows the strongly enhanced view to inherit the discriminative information contained in the weakly enhanced view while not being disturbed by excessive transformations, thus achieving discriminative feature alignment. The expression is as follows:

[0065]

[0066] The total loss function is defined as the weighted sum of the standard contrast loss and the distribution divergence minimization loss, and the expression is as follows:

[0067]

[0068] Among them, β∈[0,1] represents the weighting coefficient.

[0069] Finally, a confidence sample selection module is constructed based on the feature space distribution density. By calculating the cosine similarity between the sample and the cluster center, an ascending queue is constructed, and high-confidence samples are dynamically selected to participate in the training of the comparative learning network.

[0070] like Figure 4 As shown, in this embodiment, the step S4 is specifically as follows:

[0071] First, feature extraction is performed on the test set data samples. This involves using the backbone network trained in step S3 to extract discriminant features for the test samples. The facial region images from the test video are fed into the backbone network to obtain their representations in feature space. Based on the extracted feature representations, the K-means++ algorithm is then used for binary clustering, dividing the test samples into two categories.

[0072] Since fake videos often use a certain smoothing strategy or eliminate noise in the post-processing process during the generation process, which makes the features between consecutive frames appear too consistent, their inter-frame correlation is usually higher than that of real videos. Therefore, in order to determine which category corresponds to the real video and which category corresponds to the fake video, inter-frame correlation analysis is introduced. By calculating the Spearman correlation coefficient ρ between consecutive frames l , to distinguish the true from the false of the clustering results, that is, for the samples in each category, calculate the average value of the Spearman correlation coefficient between the features of adjacent frames The expression is as follows:

[0073]

[0074] Among them, ρ l represents the correlation between the lth frame and the l+1th frame; fl,n represents the nth feature of the lth frame; L represents the total number of frames in the video clip, and N represents the total number of features in each frame.

[0075] Ultimately, the category with a lower correlation coefficient is judged to be a real video, and the category with a higher correlation coefficient is judged to be a fake face video.

[0076] In experimental verification, the detection performance of the method of the present invention on multiple public datasets is better than that of existing unsupervised methods, and it shows strong robustness in complex scenarios.

[0077] In summary, the method of the present invention combines multi-level feature fusion with strong enhancement contrast learning, and uses strong enhancement contrast learning to extract the essential features of deep fake face videos, which can achieve high accuracy and generalization ability of deep fake detection and show strong robustness in complex scenarios. The method of the present invention adopts classic algorithms such as K-means++, cosine similarity and Spearman correlation coefficient, combined with advanced technologies such as contrast learning and distribution divergence minimization in deep learning. It not only has high theoretical value, but also has good practical application prospects. The strong enhancement strategy can mine deep fake patterns that are difficult to expose with conventional weak enhancement, and the DDM architecture effectively suppresses the negative effects of noise and excessive transformations, thereby improving the robustness and generalization of deep fake video face detection.

[0078] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. An unsupervised deep fake face video detection method based on strong enhanced contrastive learning, the specific steps are as follows: S1. Input unlabeled deep fake face video and real face video data, perform preprocessing, and divide them into training set and test set; Use the multi-task cascade convolutional neural network (MTCNN) model to detect faces in the input video, crop the detected face area, generate full-size unlabeled images and face area images, and finally divide the face video into training and test sets based on the actual ratio. S2. Extract features from the full-size unlabeled images and face area image data in the training set obtained in step S1, fuse the extracted features, perform preliminary clustering, and generate pseudo labels by assigning pseudo labels; First, the visual artifact feature VAF and local texture feature LTF are extracted simultaneously from the full-size unlabeled image; in, VAF is based on high-frequency residual analysis and extraction. It converts the image from the spatial domain to the frequency domain and analyzes abnormal patterns in the spectrum to capture visual artifacts introduced in the deep fake process. LTF extracts the texture features of the image by combining local binary patterns and Gabor filter banks. The extracted VAF and LTF features are then fused to construct a hybrid feature space. The fusion method is to first standardize the two types of features and then splice them in the feature dimension. In the hybrid feature space, the K-means++ algorithm is used to perform preliminary clustering on the samples and divide the samples into two categories. Finally, based on the distance between the sample and the cluster center, the samples are further divided into three categories: high-confidence samples, boundary samples, and outliers. Corresponding label assignment strategies are adopted for different types of samples to generate reliability pseudo labels. The same pseudo labels are assigned to the face area images in the same frame as the full-size unlabeled image. Among them, for high-confidence samples, the corresponding category pseudo labels are directly assigned; for boundary samples, the relative distance ratio between them and the cluster centers is calculated, and the threshold is set to assign pseudo labels; for outliers, no pseudo labels are assigned temporarily; S3. By randomly combining multiple weak and strong enhancement operations on the pseudo-labeled face region images in the training set, an enhanced view with structural deformation and semantic difference is constructed. Then, a contrastive learning network architecture including a double-layer projection head is adopted, and a distribution divergence minimization (DDM) strategy is introduced to build a confidence sample selection module. High-confidence samples are dynamically selected to participate in the training of the contrastive learning network, completing the strong enhancement contrastive learning of face images. S4. Use the backbone network of the contrastive learning network trained in step S3 to extract the features of the data samples in the test set divided in step S1, and perform binary clustering based on the K-means++ algorithm. Finally, use the inter-frame correlation to distinguish the authenticity of the classification results, thereby achieving accurate detection of forged face videos.

2. The unsupervised deep fake face video detection method based on strong enhanced contrast learning according to claim 1 is characterized in that The step S3 is specifically as follows: First, weak enhancements T and T' are applied simultaneously to the pseudo-label face region image X in the training set to generate weakly enhanced views V and V'; and strong enhancement transformation S is applied to image X to generate a strongly enhanced view V". V and V' are converted into strong enhanced views using the key-value encoder f θ Get its embedded representation z = f θ (V) and z ′ =f θ (V ′ ); at the same time, pass V” through the query encoder Get the corresponding embedding representation Finally, each batch will add z into the queue to form a group including z1, z2, ..., z k ,z’s representation library; Among them, weak enhancement operations include: random cropping, horizontal flip, color jitter and Gaussian blur; strong enhancement is generated by randomly combining multiple basic transformation operations, including: shear X (Y) (-0.3, 0.3), translation X (Y) (-0.3, 0.3), rotation (-30, 30), auto contrast (0 or 1), inversion (0 or 1), equalization (0 or 1), solarization (0, 256), toning (4, 8), contrast (0.05, 0.95), brightness (0.05, 0.95), sharpness (0.05, 0.95); Among them, X(Y) represents the vertical and horizontal directions, and the numerical value represents the operation amplitude; Then construct a contrastive learning network architecture including a double-layer projection head. The first projection head g u By reducing the dimension to eliminate feature redundancy, the second projection head g f Perform comparative spatial mapping; The dual-layer projection head consists of two series-connected fully connected layers and a normalization layer. The first layer is a 2048-dimensional hidden layer followed by a ReLU activation function, and the second layer is an output layer without an activation function. The former reduces the dimensionality of the output, while the latter maps it to a low-dimensional space that is friendly to contrastive learning. The loss function used by the dual-layer projection head is The calculation expression is as follows: in, represents the mathematical expectation of the sample index i in a batch B; K represents the total number of key embeddings; q(z k ∣z i ′ ) represents the query embedding z i ′ Embed z with any key k The conditional distribution between q(z i ∣z i ′ ) represents the ideal target distribution; when j = i, p(z i ∣z i ′ ) represents the likelihood of the positive sample pair; when j = k and k ≠ i, it represents the relative negative likelihood between the query and the negative sample; τ represents the temperature parameter; Then, the distribution divergence minimization DDM strategy is introduced to calculate the similarity of the weakly enhanced view V' and the strongly enhanced view V" in all samples in the representation library, respectively, to obtain the similarity vector. Then, the KL divergence between the two similarity vectors is calculated as a measure of the distribution difference, and the divergence is minimized as the loss function. The expression is as follows: The total loss function is defined as the weighted sum of the standard contrast loss and the distribution divergence minimization loss, and the expression is as follows: Among them, β∈[0,1] represents the weighting coefficient; Finally, a confidence sample selection module is constructed based on the feature space distribution density. By calculating the cosine similarity between the sample and the cluster center, an ascending queue is constructed, and high-confidence samples are dynamically selected to participate in the training of the comparative learning network.

3. The unsupervised deep fake face video detection method based on strong enhanced contrast learning according to claim 1 is characterized in that The step S4 is specifically as follows: First, feature extraction is performed on the test set data samples. That is, the backbone network trained in step S3 is used to extract the discriminant features of the test samples. The face area image in the test video is input into the backbone network to obtain its representation in the feature space. Then, based on the extracted feature representation, the K-means++ algorithm is used for binary clustering to divide the test samples into two categories. Then, inter-frame correlation analysis is introduced by calculating the Spearman correlation coefficient ρ between consecutive frames. l , to distinguish the true from the false of the clustering results, that is, for the samples in each category, calculate the average value of the Spearman correlation coefficient between the features of adjacent frames The expression is as follows: Among them, ρ l represents the correlation between the lth frame and the l+1th frame; f l,n represents the nth feature of the lth frame; L represents the total number of frames in the video clip, and N represents the total number of features in each frame; Ultimately, the category with a lower correlation coefficient is judged to be a real video, and the category with a higher correlation coefficient is judged to be a fake face video.

Images