A low-quality deepfake detection method based on high-frequency enhancement
By constructing local and global high-frequency information enhancement branches, combined with channel attention mechanism and multi-level wavelet decomposition, the problem of low accuracy in low-quality Deepfake detection is solved, and efficient detection on highly compressed data is achieved.
Patent Information
- Application Number
- CN202310011974.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-05
- Publication Date
- 2026-01-06
- Estimated Expiration
- 2043-01-05
AI Technical Summary
Existing deepfake detection methods perform poorly in detecting low-quality deepfakes, especially under high compression conditions, making it difficult to effectively detect forged face images.
By constructing local high-frequency information enhancement branches and global high-frequency information enhancement branches, and utilizing channel attention mechanism and multi-level wavelet decomposition, local and global high-frequency information is extracted and fused, and detection is performed in conjunction with backbone network features.
It improves the accuracy of low-quality Deepfake detection, especially performing well on highly compressed data, and does not rely on additional information, with high computational efficiency.
Smart Images
Figure CN116863513B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of Deepfake defense, and particularly relates to a low-quality Deepfake detection method based on high-frequency enhancement. BACKGROUND
[0002] With the continuous development of technology and the rapid improvement of computing power, artificial intelligence theory and technology are constantly maturing. At present, it has been applied in many fields such as city service, intelligent transportation, industrial manufacturing and military security, and has promoted the human world to enter the intelligent era. As one of the core fields of artificial intelligence, deep learning (DL) performs excellently in many tasks, such as target detection in the field of computer vision (CV) and speech recognition in the field of natural language processing. However, while deep learning technology leads a new round of artificial intelligence wave, it also faces serious potential threats.
[0003] Existing Deepfake defense technology mainly focuses on high-quality Deepfake detection, that is, false content detection of uncompressed or lightly compressed false content. Specifically, given a clear human face image, the Deepfake detector is committed to extracting various features and making true or false classification according to these features. Although these methods have a particularly high detection accuracy, once the data is compressed, the detection method is limited, because most Deepfake detection methods rely on the manipulation details brought by the forgery process, and compression will bring artifacts, making the forged content blurred, resulting in a sharp decline in the performance of the existing detection method. Therefore, the false face detection for low-quality Deepfake becomes a problem to be solved.
[0004] Although there are some works on low-quality Deepfake detection, such as document 1 "J. Hu, X. Liao, W. Wang, and Z. Qin, "Detecting Compressed Deepfake Videos in Social Networks Using Frame-Temporality Two-Stream Convolutional Network," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 3, pp. 1089-1102, March 2022, doi: 10.1109 / TCSVT.2021.3074259." proposes a detection method for Deepfake compressed videos, by analyzing the intra-frame features and inter-frame features of compressed Deepfake videos, a two-stream method is proposed, which uses time-level stream to extract time correlation features, and uses frame-level stream to gradually prune the network to prevent the model from fitting compression noise, finally, the prediction results of the two streams are fused. Document 2 "Le, Binh M., and Simon S. Woo. "ADD: Frequency Attention and Multi-View based Knowledge Distillation to Detect Low-Quality Compressed Deepfake Images." arXiv preprint arXiv:2112.03553 (2021)." applies frequency domain learning and optimal transport theory to knowledge distillation (KD), explores the transfer learning ability in KD to enable the student network to effectively learn discriminative features from low-quality images, thereby improving the detection of low-quality compressed Deepfake images. However, the above methods all need additional information, such as time information and the guidance of high-quality images. In addition, document 3 "Qian Y, Yin G, Sheng L, et al. Thinking in frequency: Face forgery detection by mining frequency-aware clues [C] / / European conference on computer vision. Springer, Cham, 2020: 86-103." uses frequency domain information to improve the detection performance of low-quality fake content. SUMMARY
[0005] In order to solve the problems of the prior art, the application provides a low-quality Deepfake detection method based on high-frequency enhancement, comprising the following steps:
[0006] Obtaining a face image set;
[0007] Extracting feature information, local high-frequency information and global high-frequency information of face images in the face image set;
[0008] Fusing the feature information and the local high-frequency information and the global high-frequency information respectively to obtain first fusion information and second fusion information;
[0009] Fusing the first fusion information and the second fusion information to obtain final fusion feature information;
[0010] Inputting the final fusion feature information into a classifier for classification prediction to realize detection of false faces.
[0011] Further, the extraction of the feature information comprises:
[0012] Extracting feature information of an RGB face image by using a backbone network Xception;
[0013] Preferably, the extraction of the feature information of the RGB face image by using the Xception as the backbone network comprises:
[0014] Dividing the backbone network Xception into modules, setting a feature extraction module and a feature fusion module;
[0015] Inputting the face image into the feature extraction module to extract feature information of the face image;
[0016] Setting an insertion position of features to be fused in the feature fusion module, inserting the local high-frequency information or the global high-frequency information at the insertion position, fusing the feature information and the local high-frequency information and the global high-frequency information respectively to obtain first fusion information and second fusion information.
[0017] Further, the extraction of the local high-frequency information comprises:
[0018] Segmenting the face image into a plurality of sub-block images, performing discrete cosine transform on each sub-block image to obtain corresponding DCT coefficients;
[0019] Performing cascade processing on the DCT coefficients of each sub-block image to obtain multi-dimensional DCT coefficients of the sub-block image; arranging all the sub-block images according to the original relative positions to obtain a DCT coefficient tensor;
[0020] performing a channel attention mechanism on the DCT coefficient tensor, adding a DCT coefficient attention mask in the DCT coefficient tensor to obtain a first enhanced DCT coefficient tensor;
[0021] performing a channel bottleneck module denoising operation on the first enhanced DCT coefficient tensor to obtain a second enhanced DCT coefficient tensor, and performing an inverse discrete cosine transform operation on the second enhanced DCT coefficient tensor to obtain the local high-frequency information.
[0022] Further, performing a channel attention mechanism on the DCT coefficient tensor, adding a DCT coefficient attention mask in the DCT coefficient tensor, comprises:
[0023] First, a max-pooling layer and an average-pooling layer are constructed to process the obtained DCT coefficient tensor to obtain the maximum and average values in the amplitudes corresponding to the DCT basis functions, i.e., to obtain the high-frequency amplitudes and average amplitudes;
[0024] Then, a shared multi-layer perception and a Sigmoid activation layer are constructed, the high-frequency amplitudes and average amplitudes are sent into the shared multi-layer perception, and then are sent into the Sigmoid activation layer after superposition, so as to generate the DCT coefficient attention mask, so that it pays more attention to the high-frequency coefficient channel.
[0025] Further, the extraction of the local high-frequency information further comprises: after obtaining the first enhanced DCT coefficient tensor, performing denoising processing on the first enhanced DCT coefficient tensor to obtain a new DCT coefficient attention mask, multiplying the new DCT coefficient attention mask with the DCT coefficient tensor to obtain a second enhanced DCT coefficient tensor;
[0026] Preferably, the denoising processing is performed by a bottleneck module neural network, specifically comprising: 4 consecutive blocks, each block being composed of a convolutional layer, a normalization layer and an activation function layer, each convolutional layer being provided with a 3*3 kernel, and the stride and padding of each convolutional layer being 1; the obtained new DCT coefficient tensor is cascaded with the channel bottleneck module to smooth the DCT coefficients and filter out the DCT coefficient noise;
[0027] Finally, the newly obtained DCT coefficient attention mask is multiplied with the obtained DCT coefficient tensor to obtain a new DCT coefficient tensor, so as to achieve the purpose of high-frequency enhancement.
[0028] Further, the acquisition of the global high-frequency information comprises:
[0029] performing wavelet decomposition on the face image to obtain a first-level high-frequency component and a first-level low-frequency component, wherein the first-level high-frequency component comprises high-frequency components in different directions;
[0030] The first high-frequency component is subjected to convolution module learning to obtain a first high-frequency component;
[0031] The first low-frequency component is subjected to wavelet decomposition to obtain a second high-frequency component and a second low-frequency component, the second high-frequency component is subjected to convolution module learning, and the learning result is fused with the first high-frequency component to obtain a first global high-frequency information, and the operation is repeatedly performed for multiple times to finally obtain the global high-frequency information.
[0032] Further, the global high-frequency information acquisition further comprises:
[0033] The first low-frequency component is subjected to wavelet decomposition, the high-frequency component obtained through the decomposition is subjected to convolution module learning, and the learning result is fused with the high-frequency component convolution module learning result of the previous level to obtain the global high-frequency information of the level;
[0034] The low-frequency component obtained through the decomposition is subjected to wavelet decomposition;
[0035] The above operation is repeatedly performed for multiple times to obtain the multi-level global high-frequency information.
[0036] Further, the components in different directions include horizontal components, vertical components and diagonal components.
[0037] Further, the components in different directions are concatenated before the convolution module learning.
[0038] Further, the acquisition of the face image set to be detected comprises the following steps: according to an existing Deepfake face forgery public video data set, face image cropping and extraction are performed by using a face detector MTCNN, and manual screening and removal of false detection data are performed to perform data cleaning.
[0039] The present application has the following advantages:
[0040] The present application has the following advantages:
[0041] In addition, the detected deepfake accuracy is obviously better than other methods, especially on highly compressed data. In order to verify the effectiveness of the method, different deepfake detection methods are used to perform experiments on the FF++ dataset, which includes three versions, raw, c23 and c40. The method achieves an accuracy of 99.36% on the raw version, 92.83% on the c23 version, and 71.84% on the c40 version. Compared with the benchmark model Xception, it is increased by 0.48%, 2.37% and 1.87% respectively. In addition, our method does not require additional time information and has high computational efficiency. BRIEF DESCRIPTION OF DRAWINGS
[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0043] Figure 1 Flowchart of the present application;
[0044] Figure 2 Local high-frequency information extraction process schematic diagram of the present application;
[0045] Figure 3 Global high-frequency information extraction process schematic diagram of the present application;
[0046] Figure 4 Information fusion process schematic diagram of the present application. DETAILED DESCRIPTION
[0047] In order to make the purpose, technical scheme and advantages of the present application more clear, the embodiments of the present application will be further described in detail below.
[0048] Reference Figures 1 to 4 A low-quality deepfake detection method based on high-frequency enhancement, comprising:
[0049] Prepare face dataset:
[0050] According to the existing Deepfake face forgery public video dataset, the face detector MTCNN is used to crop and extract the face image, and manual screening is performed to remove false detection data for data cleaning.
[0051] Specifically:
[0052] (a) For a given video dataset, first extract the consecutive N video frames of each of the K videos using the video reading algorithm, then extract the face image in each video frame using the face detection algorithm, and perform normalization and center cropping to obtain the RGB face image I(x,y), x=1,2,…,w,y=1,2,…,h with width and height w and h;
[0053] Where the video is extracted using the VideoReader class in python, and the face image is extracted using the face detector MTCNN, and the obtained face image is placed in the folder.
[0054] The face image obtained in step (a) is adjusted to an image with a size of 256*256*3, and center cropping is performed, and the face image is normalized using the mean [0.485, 0.456, 0.406] and variance [0.229, 0.224, 0.225], and finally an input image with a size of 224*224*3 is obtained;
[0055] The normalized image is packaged as a tensor x with [b,c,h,w] i ∈R b×c×h×w , R is a vector space, c is the number of channels of the input image, h and w are the height and width of the input image; b is the number of face images in each batch, and the label is y∈{0,1}, where 0 represents a real video and 1 represents a false video.
[0056] Data preprocessing:
[0057] (b) Preprocess the face image, filter out false detection face images, and perform data cleaning, and finally perform data segmentation.
[0058] Specifically:
[0059] (b-1) Manually remove non-face images detected by the face detector to obtain a relatively pure data set;
[0060] (b-2) Divide the face data set obtained in step (b-1) into a training set and a test set, wherein the training set accounts for 75% and the test set accounts for 25%; wherein the training set is used for subsequent model training, and the test set is used to judge the performance of the model.
[0061] The extraction of feature information includes:
[0062] (c) Extract features from the RGB face image using the backbone network Xception;
[0063] Specifically:
[0064] (c-1) the main network Xception is divided into a feature extraction module and a feature fusion module;
[0065] The main network Xception is divided into a feature extraction module and a feature fusion module. Figure 4
[0066] (c-2) the feature extraction module is used to extract the feature information of the face image.
[0067] The feature fusion module is used to set the insertion position of the to-be-fused feature, insert the local high-frequency information or the global high-frequency information into the insertion position, and fuse the feature information, the local high-frequency information and the global high-frequency information respectively to obtain first fusion information and second fusion information.
[0068] The extraction of the local high-frequency information includes:
[0069] (d) the local high-frequency information of the face image is extracted by using discrete cosine transform and inverse transform, and is fused with the feature extracted by the main network Xception in step (c);
[0070] Specifically:
[0071] (d-1) first, the face image with a size of 224*224*3 is converted into a YCbCr image; then the YCbCr image is divided into 28*28 sub-block images according to the size of 8*8, and each sub-block image is a 3-channel image;
[0072] (d-2) discrete cosine transform is performed on each sub-block image obtained in step (d-1) to obtain corresponding 64*3 DCT coefficients;
[0073] (d-3) the 64*3 DCT coefficients of each sub-block obtained in step (d-2) are concatenated to obtain corresponding 192-dimensional DCT coefficients; all sub-blocks are arranged according to the original relative position to finally obtain a DCT coefficient tensor with a size of 28*28*192;
[0074] (d-4) performing a channel attention mechanism on the DCT coefficient tensor obtained in step (d-3), the specific steps being: first, constructing a max-pooling layer and an average-pooling layer, processing the DCT coefficient tensor obtained in step (d-3) to obtain the maximum and average of the amplitudes corresponding to 192 DCT basis functions, that is, the high-frequency amplitudes and average amplitudes; then constructing a shared multi-layer perception and a Sigmoid activation layer, inputting the high-frequency amplitudes and average amplitudes into the shared multi-layer perception, and then inputting them into the Sigmoid activation layer after superposition, thereby generating a DCT coefficient attention mask, so that it pays more attention to the high-frequency coefficient channel;
[0075] (d-5) constructing a learnable channel bottleneck module, specifically including 4 consecutive blocks, each block being composed of a convolutional layer, a normalization layer and an activation function layer, each convolutional layer being provided with a 3*3 kernel, and the stride and padding of each convolutional layer being 1, the new DCT coefficient tensor obtained in step (d-4) being cascaded with the channel bottleneck module to smooth the DCT coefficients and filter out DCT coefficient noise; finally, multiplying the newly obtained DCT coefficient attention mask with the DCT coefficient tensor obtained in step (d-3) to obtain a new DCT coefficient tensor, so as to achieve the purpose of high-frequency enhancement;
[0076] (d-6) performing a discrete cosine inverse transform operation on the DCT coefficient tensor obtained in step (d-5) to restore it to the RGB spatial domain, thereby indirectly realizing a frequency-aware local spatial attention mechanism.
[0077] (d-7) extracting features from the RGB spatial domain image obtained in step (d-6);
[0078] (d-8) performing dual-domain fusion on the features of the RGB spatial domain image obtained in step (d-7) and the feature information of the face image obtained in step (c-2), as shown in FIG. 8. Figure 4
[0079] The acquisition of the global high-frequency information includes:
[0080] (e) extracting global high-frequency information from the face image by using a multi-level wavelet decomposition and fusion strategy, and fusing the extracted global high-frequency information with the features extracted by the backbone network in step (c);
[0081] Specifically:
[0082] (e-1) first performing wavelet decomposition on the face image with a size of 224*224*3 to obtain a first-level high-frequency component and a first-level low-frequency component, wherein the high-frequency component is divided into three directions, namely: a horizontal component, a vertical component, and a diagonal component;
[0083] (e-2) Cascade the three high frequency components of different directions obtained in step (e-1), construct a grouped convolution module to learn, obtain new horizontal component, vertical component and diagonal component;
[0084] (e-3) Continue to perform wavelet decomposition on the three high frequency components of different directions obtained in step (e-2), cascade the horizontal component, vertical component and diagonal component obtained after decomposition to construct corresponding convolution blocks to learn, to constitute new horizontal component, vertical component and diagonal component;
[0085] (e-4) Continue to perform wavelet decomposition on the first-level low frequency component obtained in step (e-1), obtain second-level high frequency component and second-level low frequency component, and the high frequency component is divided into three directions, i.e., horizontal component, vertical component and diagonal component;
[0086] (e-5) Cascade the three high frequency components of different directions obtained in step (e-4), construct a grouped convolution module to learn, obtain new horizontal component, vertical component and diagonal component;
[0087] (e-6) Correspondingly cascade and learn the three high frequency components of different directions obtained in step (e-3) and the three high frequency components of different directions obtained in step (e-5), obtain new horizontal component, vertical component and diagonal component, so as to realize first-level global high frequency information fusion;
[0088] (e-7) Continue to perform wavelet decomposition on the horizontal component, vertical component and diagonal component obtained in step (e-6), cascade the horizontal component, vertical component and diagonal component obtained after decomposition to construct corresponding convolution blocks to learn, to constitute new horizontal component, vertical component and diagonal component;
[0089] (e-8) Continue to perform wavelet decomposition on the second-level low frequency component obtained in step (e-4), obtain third-level high frequency component and third-level low frequency component, and the high frequency component is divided into three directions, i.e., horizontal component, vertical component and diagonal component;
[0090] (e-9) Cascade the three high frequency components of different directions obtained in step (e-8), construct a grouped convolution module to learn, obtain new horizontal component, vertical component and diagonal component;
[0091] (e-10) Correspondingly cascade and learn the horizontal component, vertical component and diagonal component obtained in step (e-7) and the three high frequency components of different directions obtained in step (e-9), obtain new horizontal component, vertical component and diagonal component, so as to realize second-level global high frequency information fusion;
[0092] (e-11) continue wavelet decomposition on the horizontal component, the vertical component and the diagonal component obtained in step (e-10), and cascade the horizontal component, the vertical component and the diagonal component obtained after the decomposition to construct a corresponding convolution block for learning, to form new horizontal component, vertical component and diagonal component;
[0093] (e-12) continue wavelet decomposition on the three-level low-frequency components obtained in step (e-8), to obtain four-level high-frequency components and four-level low-frequency components, and among them, the high-frequency components are divided into three directions, i.e., horizontal component, vertical component and diagonal component;
[0094] (e-13) cascade the high-frequency components in the three different directions obtained in step (e-12) to construct a group convolution module for learning, to obtain new horizontal component, vertical component and diagonal component;
[0095] (e-14) correspondingly cascade and learn the horizontal component, the vertical component and the diagonal component obtained in step (e-11) and the high-frequency components in the three different directions obtained in step (e-13), to obtain new horizontal component, vertical component and diagonal component, so as to realize three-level global high-frequency information fusion;
[0096] (e-15) cascade the horizontal component, the vertical component and the diagonal component obtained in step (e-14), and extract global high-frequency features through five convolution blocks;
[0097] (e-16) perform dual-domain fusion on the global high-frequency features obtained in step (e-15) and the feature information of the face image obtained in step (c-2), as shown in FIG. 8. Figure 4
[0098] (f) fuse the features obtained in step (e-16) and step (d-8), and input the fused features into a classifier for prediction, as shown in FIG. 9. Figure 4
[0099] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A low-quality Deepfake detection method based on high-frequency enhancement, characterized in that, The method comprises the following steps: obtaining a face image set; extracting feature information, local high-frequency information and global high-frequency information of face images in the face image set; fusing the feature information and the local high-frequency information and the global high-frequency information respectively to obtain first fusion information and second fusion information; fusing the first fusion information and the second fusion information to obtain final fusion feature information; inputting the final fusion feature information into a classifier for classification prediction to realize detection of a fake face; the extraction of the local high-frequency information comprises: segmenting the face image into a plurality of sub-block images, performing discrete cosine transform on each of the sub-block images to obtain corresponding DCT coefficients; performing cascade processing on the DCT coefficients of each of the sub-block images to obtain multi-dimensional DCT coefficients of the sub-block images; and arranging all the sub-block images according to the original relative positions to obtain a DCT coefficient tensor; performing a channel attention mechanism on the DCT coefficient tensor to add a DCT coefficient attention mask to the DCT coefficient tensor to obtain a first enhanced DCT coefficient tensor; performing a channel bottleneck module operation on the first enhanced DCT coefficient tensor to obtain a second enhanced DCT coefficient tensor; and performing an inverse discrete cosine transform operation on the second enhanced DCT coefficient tensor to obtain the local high-frequency information.
2. The low-quality Deepfake detection method based on high-frequency enhancement according to claim 1, wherein, the extraction of the feature information comprises: extracting feature information of an RGB face image by using a backbone network Xception; the extraction of the feature information of the RGB face image by using the Xception as the backbone network comprises: dividing the backbone network Xception into a feature extraction module and a feature fusion module; inputting the face image into the feature extraction module to extract feature information of the face image; setting an insertion position of a feature to be fused in the feature fusion module, inserting the local high-frequency information or the global high-frequency information at the insertion position, and fusing the feature information and the local high-frequency information and the global high-frequency information respectively to obtain first fusion information and second fusion information.
3. The low-quality Deepfake detection method based on high-frequency enhancement according to claim 1, wherein performing a channel attention mechanism on the DCT coefficient tensor to add a DCT coefficient attention mask to the DCT coefficient tensor comprises: firstly, constructing a max-pooling layer and an average-pooling layer to process the obtained DCT coefficient tensor to obtain a maximum value and an average value in an amplitude corresponding to a DCT basis function, i.e., a high-frequency amplitude and an average amplitude; then, constructing a shared multi-layer perceptron and a Sigmoid activation layer, inputting the high-frequency amplitude and the average amplitude into the shared multi-layer perceptron, and then inputting the superimposed high-frequency amplitude and average amplitude into the Sigmoid activation layer to generate a DCT coefficient attention mask, so that the DCT coefficient attention mask pays more attention to a high-frequency coefficient channel.
4. The low-quality Deepfake detection method based on high-frequency enhancement of claim 1, wherein, The extraction of the local high-frequency information further comprises: after obtaining the first enhanced DCT coefficient tensor, performing denoising processing on the first enhanced DCT coefficient tensor to obtain a new DCT coefficient attention mask, multiplying the new DCT coefficient attention mask with the DCT coefficient tensor to obtain a second enhanced DCT coefficient tensor; The denoising processing is performed by using a bottleneck module neural network, and specifically comprises: four continuous blocks, each block comprising a convolution layer, a normalization layer and an activation function layer, each convolution layer being provided with a 3*3 kernel, and the stride and padding of each convolution layer being 1; the obtained new DCT coefficient tensor is cascaded with a channel bottleneck module to smooth the DCT coefficients and filter the DCT coefficient noise. Finally, the newly obtained DCT coefficient attention mask is multiplied with the obtained DCT coefficient tensor to obtain a new DCT coefficient tensor, so as to achieve the purpose of high-frequency enhancement.
5. The low-quality Deepfake detection method based on high-frequency enhancement of claim 1, wherein, The acquisition of the global high-frequency information comprises: wavelet decomposition is performed on the face image to obtain a first high-frequency component and a first low-frequency component, wherein the first high-frequency component comprises high-frequency components in different directions; convolution module learning is performed on the first high-frequency component to obtain a first high-frequency component; wavelet decomposition is performed on the first low-frequency component to obtain a second high-frequency component and a second low-frequency component, and convolution module learning is performed on the second high-frequency component, and the learning result is fused with the first high-frequency component to obtain a first global high-frequency information, and the operation is repeatedly performed for multiple times to finally obtain the global high-frequency information.
6. The low-quality Deepfake detection method based on high-frequency enhancement according to claim 5, wherein, The acquisition of the global high-frequency information further comprises: wavelet decomposition is performed on the first low-frequency component, convolution module learning is performed on the obtained high-frequency component, the learning result is fused with the high-frequency component learning result of the previous stage to obtain the global high-frequency information of the stage; the obtained low-frequency component is further subjected to wavelet decomposition; the above operation is repeated for multiple times to obtain multiple levels of global high-frequency information.
7. The low-quality Deepfake detection method based on high-frequency enhancement of claim 5, wherein, The components in different directions comprise horizontal components, vertical components and diagonal components.
8. The low-quality Deepfake detection method based on high-frequency enhancement of claim 5, wherein, Before the convolution module learning, the components in different directions are cascaded.
9. The low-quality Deepfake detection method based on high-frequency enhancement of claim 1, wherein, The acquisition of the face image set comprises: according to an existing Deepfake face forgery public video dataset, face image cropping and extraction are performed by using a face detector MTCNN, and manual screening and removal of false detection data are performed for data cleaning.