A false face detection model compression method based on joint distillation
By constructing and training a student neural network model based on a joint distillation method, the problem of deploying fake face detection models on resource-constrained devices is solved, achieving efficient fake face detection while maintaining good detection performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 10TH RES INST OF CETC
- Filing Date
- 2023-03-01
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for detecting fake face videos involve deep neural network models with large parameter sizes and high computational costs, making them difficult to deploy on devices with limited computing resources. Furthermore, existing model compression algorithms struggle to effectively extract forensic knowledge learned by intermediate network layers, resulting in performance loss.
A joint distillation-based approach is adopted to construct a student neural network model by reducing the number of network layers and parameters in the teacher's neural network. The student neural network model is then trained using a joint distillation loss function to ensure dimensional alignment of the intermediate layer feature maps. Cross-entropy and gradient information are combined to assist the feature distillation loss function in training.
It enables the deployment of fake face detection models on devices with limited computing resources, significantly improving computational efficiency while slightly reducing or maintaining detection performance, and is applicable to neural network detection algorithms with different structures.
Smart Images
Figure CN116341649B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of compression technology, and in particular to a compression method for a fake face detection model based on combined distillation. Background Technology
[0002] In recent years, with the rapid development of computer vision and artificial intelligence technologies, the technology for creating fake face videos has achieved new breakthroughs, with Deepfake technology being the most representative. Deepfake technology can replace the face in a source video with the face of a target person, while maintaining a relatively consistent expression. As the technology for creating fake face videos continues to be optimized and updated, these videos are now difficult for the human eye to distinguish. On the one hand, this technology can provide strong technical support for industries such as film and digital entertainment. For example, various face video editing software programs launched in recent years, including FaceSwap and ZAO, have been released. On the other hand, the open-source nature of this technology has lowered the barrier to entry, allowing ordinary users to easily create highly realistic fake face videos. These fake face videos have been used by criminals to commit fraud or spread rumors, causing significant harm to personal finances, online public opinion stability, and national security.
[0003] To address the potential threat of fake face videos, research institutions and internet companies worldwide have proposed various detection methods. Most of these methods employ deep neural network models to achieve good detection performance, combining this with various tampering traces in the fake face videos. While deep neural network models possess data-driven feature learning capabilities, and the aforementioned methods have achieved good detection performance on standard fake face video datasets, the deep neural network models used in existing detection methods are becoming increasingly complex, with large model parameter scales and high computational costs, making them difficult to deploy on devices with limited computing resources and unable to handle the massive video detection scenarios on the internet. Therefore, model compression for fake face video detection methods based on neural network models has significant research and application value, and can effectively promote the practical application of fake face video detection. However, there is a significant lack of model compression algorithms for fake face video detection. Only the paper "Towards Generalizable DEEPFAKEFaceForgery Detection with Semi-Supervised Learning and Knowledge Distillation" by Lin et al. proposes a basic method for model compression in fake face video detection. This method uses standard knowledge distillation on the numerical vectors (logits) output by the model's output layer to achieve model compression, preserving some of the original model's detection capabilities while compressing model parameters. However, for fake face video detection tasks, the above method struggles to effectively mine the forensic knowledge learned by the intermediate layers of the neural network model and cannot adaptively perform knowledge distillation based on the importance of different feature channels, resulting in a significant performance loss even after compression. Summary of the Invention
[0004] In view of this, the present invention provides a method for compressing a fake face detection model based on combined distillation to solve the above-mentioned technical problems.
[0005] This invention discloses a method for compressing a fake face detection model based on combined distillation, which includes the following steps:
[0006] Step 1: Train the known fake face detection neural network model using the training set to obtain the trained fake face detection neural network model;
[0007] Step 2: Let the trained fake face detection neural network model be the teacher neural network model. By reducing the number of network layers and the number of parameters in the network layers of the teacher neural network model, construct the student neural network model.
[0008] Step 3: Train the student neural network model using the joint distillation loss function to obtain a compressed fake face detection model.
[0009] Furthermore, in step 1:
[0010] Fake face dataset As the training set, the cross-entropy loss function is used, and the stochastic gradient descent algorithm is employed to train a known fake face detection neural network model. The training continues until the loss function value is less than a preset threshold T. r Or the number of training iterations exceeds a preset threshold N r Training ends at time; where x i Let y represent the i-th sample in the training set. i Let y represent the label of the i-th sample. i ∈{0,1}, where 0 and 1 represent that the sample belongs to a real face sample and a fake face sample, respectively, and N t This represents the total number of samples.
[0011] Furthermore, the cross-entropy loss function is:
[0012]
[0013] in, Let θ be the cross-entropy loss function, T be a known fake face detection neural network model, x be the input sample, y be the label corresponding to the input sample, and y ∈ {0,1}; t This represents the parameter set of a known neural network model for detecting fake faces.
[0014] After training T is completed, the neural network model is obtained. in, The optimal parameter set obtained through training;
[0015] For an input sample x, input it into T and perform a feedforward operation to obtain the probability that x belongs to a fake face. When s>T f If the input sample is true, it is considered a fake face video; otherwise, it is considered a real face video. Where T... f The probability threshold for determining authenticity.
[0016] Furthermore, in step 2:
[0017] The reduction of the number of network layers in the teacher's neural network model and the reduction of the number of parameters in the network layers should satisfy the following:
[0018] The number of network parameters in the student's neural network model is less than the number of network parameters in the teacher's neural network model, that is: Furthermore, both the student's neural network model and the teacher's neural network model contain intermediate layers with the same feature map dimension as their outputs, i.e., the feature map dimension alignment condition is: there exists 1 <m t <K t And 1 <m s <K s , Satisfy dim(f) t ) = dim(f s ); where dim(·) represents the function to calculate the dimension of a multidimensional vector, f t =T sub (x), f s =S sub (x);
[0019] The trained fake face detection neural network model is as follows: The network layers of the teacher's neural network model; the student's neural network model is... θ s This represents the parameter set of the student neural network model. For the network layers of the student neural network model.
[0020] Furthermore, in step 3:
[0021] Fake face dataset As the training set, the training method is as follows: using a joint distillation loss function, the student neural network model is trained using the stochastic gradient descent algorithm. When the loss function value is less than a preset threshold T... d Or the number of training iterations exceeds N d Training ends at that time.
[0022] Furthermore, the combined distillation loss function is:
[0023]
[0024] Among them, L JD For the combined distillation loss function, Let L be the cross-entropy loss function. GFD The feature distillation loss function is used to assist gradient information;
[0025]
[0026]
[0027]
[0028] in, and The d-th term of the student neural network models,1 and d p,1 There are network layers, and d s,1 <d p,1 ; and The d-th term of the teacher's neural network model s,2 and d p,2 There are network layers, and d s,2 <d p,2 ; This indicates the computation of network layer l in the student neural network model. s The output feature map and the network layer l in the teacher's neural network model t Feature distillation loss function of the output feature map, l s and l t The output feature maps all have the same dimensions, H×W×C; C represents the number of channels in the feature map, and H and W are the height and width of the feature map, respectively; for the input sample, This represents the network layer l in the teacher's neural network model. t Output the data of the k-th channel of the feature map and This represents the network layer l in the student neural network model. s Output the data of the k-th channel of the feature map. L MSE It is the mean square error function; w l,k It is based on the teacher's neural network model during backpropagation. The weights are calculated from the gradient information.
[0029] Furthermore, the weights w of network layer l k for:
[0030]
[0031]
[0032] Among them, y c This represents the predicted score of the c-th category output by the teacher's neural network model for the input sample. M k (i,j) is The element indexed by (i,j) in the middle. It represents the average intensity of the gradient information corresponding to the k-th feature map; lg(·) represents the logarithmic function with base 10; This indicates that the gradient information is obtained using the backpropagation algorithm.
[0033] Furthermore, the known neural network model for detecting fake faces is either XceptionNet or ResNet.
[0034] Furthermore, reducing the number of parameters in the network layer means either reducing the number of convolution kernel parameters in the convolutional layer or reducing the number of neurons in the fully connected layer.
[0035] Because of the adoption of the above technical solution, the present invention has the following advantages:
[0036] 1. This invention enables the training of student neural network models with lightweight parameters from teacher neural network models that have a large number of parameters. The student neural network model exhibits significantly improved computational efficiency compared to the teacher neural network model, while detection performance shows only a slight decrease or remains unchanged. This invention can be better deployed in application scenarios with limited computing resources.
[0037] 2. Feature distillation can be performed between teacher and student neural network models with different structures, as long as the intermediate network layers of the teacher and student neural network models satisfy the feature map dimension alignment condition. This invention can be flexibly applied to different neural network-based fake face detection algorithms. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0039] Figure 1 This is a schematic diagram of the process for obtaining the combined distillation function according to an embodiment of the present invention;
[0040] Figure 2 This is a flowchart illustrating a method for compressing a fake face detection model based on combined distillation, according to an embodiment of the present invention. Detailed Implementation
[0041] The present invention will be further described in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.
[0042] See Figure 2 This invention provides an embodiment of a method for compressing a fake face detection model based on combined distillation, which includes the following steps:
[0043] S1: Train the known fake face detection neural network model using the training set to obtain the trained fake face detection neural network model;
[0044] S2: Let the trained fake face detection neural network model be the teacher neural network model. By reducing the number of network layers and the number of parameters in the network layers of the teacher neural network model, a student neural network model is constructed.
[0045] S3: The student neural network model is trained using a joint distillation loss function to obtain a compressed fake face detection model.
[0046] In this embodiment, in S1:
[0047] Fake face dataset As the training set, the cross-entropy loss function is used, and the stochastic gradient descent algorithm is employed to train a known fake face detection neural network model. The training continues until the loss function value is less than a preset threshold T. r Or the number of training iterations exceeds a preset threshold N r Training ends at time; where x i Let y represent the i-th sample in the training set. i Let y represent the label of the i-th sample. i ∈{0,1}, where 0 and 1 represent that the sample belongs to a real face sample and a fake face sample, respectively, and N t This represents the total number of samples.
[0048] In this embodiment, the cross-entropy loss function is:
[0049]
[0050] in, Let θ be the cross-entropy loss function, Y be a known fake face detection neural network model, x be the input sample, y be the label corresponding to the input sample, and y ∈ {0,1}; t This represents the parameter set of a known neural network model for detecting fake faces.
[0051] After training T is completed, the neural network model is obtained. in, The optimal parameter set obtained through training;
[0052] For an input sample x, input it into T and perform a feedforward operation to obtain the probability that x belongs to a fake face. When s>Y f If Y is true, the input sample is a fake face video; otherwise, it is a real face video. f The probability threshold for determining authenticity.
[0053] In this embodiment, in S2:
[0054] Reducing the number of network layers and the number of parameters in the network layers of the teacher's neural network model should satisfy the following:
[0055] The number of network parameters in the student's neural network model is less than the number of network parameters in the teacher's neural network model, that is: Furthermore, both the student's neural network model and the teacher's neural network model contain intermediate layers with the same feature map dimension as their outputs, i.e., the feature map dimension alignment condition is: there exists 1 <m t <K t And 1 <m s <K s , Satisfy dim(f) t ) = dim(f s ); where dim(·) represents the function to calculate the dimension of a multidimensional vector, f t =T sub (x), f s =S sub (x);
[0056] Among them, the trained neural network model for detecting fake faces is The network layers of the teacher's neural network model; the student's neural network model is... θ s This represents the parameter set of the student neural network model. For the network layers of the student neural network model.
[0057] In this embodiment, in S3:
[0058] Fake face dataset As the training set, the training method is as follows: using a joint distillation loss function, the student neural network model is trained using the stochastic gradient descent algorithm. When the loss function value is less than a preset threshold T... d Or the number of training iterations exceeds N d Training ends at that time.
[0059] In this embodiment, the combined distillation loss function is:
[0060]
[0061] Among them, L JD For the combined distillation loss function, Let L be the cross-entropy loss function. GFD The feature distillation loss function is used to assist gradient information;
[0062]
[0063]
[0064]
[0065] in, and The d-th term of the student neural network model s,1 and d p,1 There are network layers, and d s,1 <d p,1 ; and The d-th term of the teacher's neural network model s,2 and d p,2 There are network layers, and d s,2 <d p,2 ; This indicates the computation of network layer l in the student neural network model. s The output feature map and the network layer l in the teacher's neural network model t Feature distillation loss function of the output feature map, l s and l t The output feature maps all have the same dimensions, H×W×C; C represents the number of channels in the feature map, and H and W are the height and width of the feature map, respectively; for the input sample, This represents the network layer l in the teacher's neural network model. t Output the data of the k-th channel of the feature map and This represents the network layer l in the student neural network model. s Output the data of the k-th channel of the feature map. L MSE It is the mean square error function; w l,k It is based on the teacher's neural network model during backpropagation. The weights are calculated from the gradient information.
[0066] In this embodiment, the weights w of network layer l k for:
[0067]
[0068]
[0069] Among them, y c This represents the predicted score of the c-th category output by the teacher's neural network model for the input sample. M k (i,j) is The element indexed by (i,j) in the middle. It represents the average intensity of the gradient information corresponding to the k-th feature map; lg(·) represents the logarithmic function with base 10; This indicates that the gradient information is obtained using the backpropagation algorithm.
[0070] In this embodiment, the known neural network models for detecting fake faces are XceptionNet or ResNet.
[0071] In this embodiment, reducing the number of parameters in the network layer means either reducing the number of convolution kernel parameters in the convolutional layer or reducing the number of neurons in the fully connected layer.
[0072] For ease of understanding, the present invention provides a more specific embodiment:
[0073] Step 1, given a known fake face detection neural network model XceptionNetT(·θ) t The model structure is shown in Table 1, where θ t This represents the parameter set of the neural network model T.
[0074] Table 1. Structure of the Teacher's Neural Network Model
[0075]
[0076]
[0077]
[0078] Using fake face datasets Where x i Let y represent the i-th sample in the dataset. i Indicates the corresponding label, y i ∈{0,1}, where “0” and “1” represent that the sample belongs to a real face sample and a fake face sample, respectively. Specifically, x is an image containing a face region cropped from the decompressed frame of the input video, with a size of 256×256. A cross-entropy loss function is applied to T. Training is performed using stochastic gradient descent, and the loss function value is less than a preset threshold T. r Or the number of training iterations exceeds N r End training at the designated time. T is recommended. r =10 -3 and N r =10 5 The cross-entropy loss function is calculated as follows:
[0079]
[0080] After the fake face detection neural network model T is trained, the neural network model is obtained. in The optimal parameter set is obtained through training. For an input sample x, performing a feedforward operation on its input T yields the probability that x belongs to a fake face. When s>T fWhen the input sample is displayed correctly, it is considered a fake face video. Conversely, it is considered a real face video. It is recommended that T... f =0.5.
[0081] Step 2, the fake face detection neural network model T consists of multiple network layers, namely K t This represents the total number of network layers in the neural network model. A student neural network model is constructed by reducing the number of network layers in the teacher's neural network model and reducing the number of parameters within each layer. K s The total number of network layers in the student neural network model. Ways to reduce the number of parameters in a network layer include: reducing the number of convolutional kernel parameters in convolutional layers, reducing the number of neurons in fully connected layers, etc. The following two conditions must be met: (1) For the number of parameters in a neural network model, the number of network parameters in S must be less than the number of network parameters in T, i.e.: (2) For the feature map dimension of the intermediate layer output of the neural network model, it is required that there exists a feature map dimension of a certain layer in S that is the same as the feature map dimension of a certain layer in T, i.e., the feature map dimension alignment condition. Specifically, it can be expressed as: there exists 1 <m t <K t And 1 <m s <K s , Satisfy dim(f) t ) = dim(f s ), where dim(·) represents the function to calculate the dimension of the vector, f t =T sub (x), f s =S sub (x). The student neural network model S can be constructed as shown in Table 2, and is called XceptionNet_S.
[0082] Table 2 Structure of the student neural network model
[0083]
[0084]
[0085] Step 3, design the combined distillation loss function L JD The student neural network model S is trained. The loss function L is used. JD It can be represented as:
[0086]
[0087] in, L represents the cross-entropy loss function. GFD This represents the feature distillation loss function aided by gradient information.
[0088]
[0089]
[0090]
[0091] in, and The d-th term of the student neural network model S s,1 and d p,1 There are network layers, and d s,1 <d p,1 . and The d-th term of the teacher's neural network model T s,2 and d p,2 There are network layers, and d s,2 <d p,2 . and These are the shallow network layers of the student and teacher neural network models, respectively. and These are the deep network layers of the student and teacher neural network models, respectively. This indicates the computation of network layer l in S. s The output feature map and network layer l in T t Feature distillation loss function of the output feature map, l s and l t The output feature maps have the same dimensions, H×W×C; C represents the number of channels in the feature map, and H and W are the height and width of the feature map, respectively. For the teacher's neural network model XceptionNet and the student's neural network model XceptionNet_S, it is recommended... and Set them to the values in Table 2 respectively. and suggestion and Set them to the values in Table 1 respectively. and
[0092] For the input sample, This represents the network layer l in the teacher's neural network model. t Output the data of the k-th channel of the feature map and This represents the network layer l in the student neural network model. s Output the data of the k-th channel of the feature map. L MSE It is the mean square error function. l,k It is based on the teacher's neural network model during backpropagation. The weights are calculated using gradient information. l,k The calculation method is as follows, with the layer number l omitted for clarity:
[0093]
[0094]
[0095] Among them, y c This represents the predicted score of the c-th class output by T for the input sample. M k (i,j) is The element indexed by (i,j) in the middle. It represents the average intensity of the gradient information corresponding to the k-th feature map. lg(·) denotes a logarithmic function with base 10. This indicates that the gradient information is obtained using the backpropagation algorithm.
[0096] The joint distillation loss function L is applied to the neural network model S. JD The algorithm is used for training, and the loss function value is less than a preset threshold T. d Or the number of training iterations exceeds N d Training ends at time T, yielding the trained student neural network model. Recommendation T d =10 -4 and N d =10 5 After the fake face detection neural network model S is trained, the student neural network model is obtained. in The optimal parameter set is obtained through training. For an input sample x, inputting it into S and performing a feedforward operation yields the probability that x belongs to a fake face. When s>T f When the input sample is displayed correctly, it is considered a fake face video. Otherwise, it is considered a real face video. It is recommended to use T... f =0.5.
[0097] This embodiment uses the FaceForensics++ (FF++) public dataset for training and testing. FF++ consists of three parts: a training set, a development set, and a validation set. For the videos in the dataset, the videos are first decompressed into video frames. 10% of the video frames are randomly selected from each video. For each video frame, the face region is extracted using the open-source face extraction tool Dlib and scaled to a resolution of 256×256. Detection accuracy (%) is used as the performance evaluation metric, while floating point operations (FLOPs) and the number of model parameters (#Param) are used as the evaluation metrics for the computational efficiency of the neural network model. The experimental results are shown in Table 3.
[0098] Table 3 Comparison of neural network performance before and after model compression.
[0099]
[0100] As shown in Table 3, compared to the teacher's neural network model XceptionNet, the student neural network model XceptionNet_S trained using joint distillation exhibits comparable detection performance (the difference is less than 1%), but with a significant improvement in computational efficiency. Specifically, the number of floating-point operations is reduced to approximately 20% of the original, while the number of model parameters is only about 13% of the original. Furthermore, simply reducing the size of the neural network would have a significant negative impact on detection performance. Therefore, the joint distillation-based fake face detection model compression algorithm proposed in this invention can effectively improve the efficiency of the detection algorithm while maintaining good detection performance.
[0101] Furthermore, this embodiment uses the classic neural network model ResNet-50 as the teacher neural network model and ResNet-18 as the student neural network model for performance analysis. When calculating the feature distillation loss function, and They were set as the 10th and 40th convolutional layers of ResNet-50, respectively. and These were set as the 5th and 13th convolutional layers of ResNet-18, respectively. The remaining experimental steps were the same as the previous experiment, and the results are shown in Table 4.
[0102] Table 4. Performance comparison of neural networks before and after model compression.
[0103]
[0104] As shown in Table 4, this invention still demonstrates good model compression capabilities for teacher and student neural networks with different network structures. While significantly reducing the number of model parameters and computational complexity, it maintains comparable detection performance between the student neural network model and the original teacher neural network model. The experimental results indicate that the method of this invention has a certain degree of versatility.
[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for compressing a fake face detection model based on combined distillation, characterized in that, Includes the following steps: Step 1: Train the known fake face detection neural network model using the training set to obtain the trained fake face detection neural network model; Step 2: Let the trained fake face detection neural network model be the teacher neural network model. By reducing the number of network layers and the number of parameters in the network layers of the teacher neural network model, construct the student neural network model. Step 3: Train the student neural network model using the joint distillation loss function to obtain a compressed fake face detection model; In step 3: Data set of fake faces As a training set, in which Represents the first in the training set One sample, Indicates the first The label of each sample, 0 and 1 represent that the sample belongs to a real face sample and a fake face sample, respectively. This represents the total number of samples. The training method is as follows: a joint distillation loss function is used, and the student neural network model is trained using the stochastic gradient descent algorithm. When the loss function value is less than a preset threshold... Or the number of training iterations exceeds a preset threshold Training ends at this time; The combined distillation loss function is: in, For the combined distillation loss function, Let cross-entropy be the loss function. The feature distillation loss function is used to assist gradient information; in, and The first character of the student neural network model and Each network layer, and ; and The first term of the teacher's neural network model and Each network layer, and ; This indicates the computation of network layers in the student neural network model. The output feature map and the network layers in the teacher's neural network model Feature distillation loss function of the output feature map and The output feature maps have the same dimension, which is 1. ; This represents the number of channels in the feature map. and These represent the height and width of the feature map, respectively; for the input sample, This represents the network layers in the teacher's neural network model. The first output feature map Data from each channel, Represents the network layers in the student neural network model The first output feature map Data from each channel, ; It is the mean square error function; It is based on the teacher's neural network model during backpropagation. The weights are calculated using gradient information. This represents the parameter set of the student neural network model. For the input sample, For the labels corresponding to the input samples, For student neural network models.
2. The method according to claim 1, characterized in that, In step 1: Fake face dataset As the training set, the cross-entropy loss function is used, and the stochastic gradient descent algorithm is employed to train a known fake face detection neural network model. The training continues until the loss function value is less than a preset threshold. Or the number of training iterations exceeds a preset threshold Training ends at that time.
3. The method according to claim 2, characterized in that, The cross-entropy loss function is: in, Let cross-entropy be the loss function. For known neural network models for detecting fake faces, For the input sample, For the labels corresponding to the input samples, ; This represents the parameter set of a known neural network model for detecting fake faces. When on After training, a neural network model is obtained. ,in, The optimal parameter set obtained through training; For input samples Enter it Performing feedforward operation yields Probability of a fake face ;when If the input sample is false, it is considered a fake face video; otherwise, it is considered a real face video. The probability threshold for determining authenticity.
4. The method according to claim 1, characterized in that, In step 2: The reduction of the number of network layers in the teacher's neural network model and the reduction of the number of parameters in the network layers satisfy the following: The number of network parameters in the student's neural network model is less than the number of network parameters in the teacher's neural network model, that is: Furthermore, both the student's neural network model and the teacher's neural network model contain intermediate layers with the same feature map dimension as their outputs, i.e., the feature map dimension alignment condition exists. and , , ,satisfy ;in, This represents a function that calculates the dimensions of a multidimensional vector. , ; The trained fake face detection neural network model is as follows: , The network layers of the teacher's neural network model; the student's neural network model is... , For the network layers of the student neural network model.
5. The method according to claim 1, characterized in that, Network layer weight for: in, This indicates that for an input sample, the output of the teacher's neural network model is the first... Predicted scores for each category; , yes China and Israel The element at index , It is the first The average intensity of gradient information corresponding to each feature map; Represents the logarithmic function with base 10; This indicates that the gradient information is obtained using the backpropagation algorithm.
6. The method according to claim 1, characterized in that, The known neural network models for detecting fake faces are XceptionNet or ResNet.
7. The method according to claim 1, characterized in that, The reduction of the number of parameters in the network layer is achieved by either reducing the number of convolution kernel parameters in the convolutional layer or reducing the number of neurons in the fully connected layer.