Image depth forgery tracing method based on GANs model

Abstract

The invention discloses an image depth forgery tracing method based on a GANs model, and belongs to the technical field of artificial intelligence. Comprising the following steps of data set preparation, feature engineering, generator and discriminator construction, generator and discriminator training, tracking and traceability analysis and continuous learning and iteration. By training the antagonism model, more complex image structures and deep features can be understood, so that the method is more accurate in distinguishing highly real forged images. Through continuous learning of the antagonism generation process, the method can better adapt to a novel counterfeit technology theoretically, and keeps the effectiveness of tracking and tracing. Specific statistical laws or generation modes of images generated by GANs can be identified through special training, so that the tracking precision is improved. An automatic learning framework is utilized, once model training is completed, a large number of images can be rapidly and automatically processed, and the tracking and analysis efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for tracking and tracing deep forgeries of images based on a GANs model. Background Art

[0002] Image forgery tracking and tracing technology aims to track and confirm the source, production process, and possible tampering history of the image. This technology is implemented through a variety of methods, including but not limited to the use of information traceability identification (such as QR codes, barcodes, RFID) to improve credibility and enhance product reputation.

[0003] Most conventional methods currently rely on lower-level image features such as pixel-level comparison and compression trace analysis. These methods are prone to failure when faced with high-precision deep fake images, because images generated by GANs may be very similar to real images at the pixel level. Existing technologies are often designed for known fake technologies, and have poor adaptability when faced with new or evolving fake technologies. They are unable to capture the unique generation traces of GANs-generated images, such as pattern repetition, texture anomalies and other deep fake features. They require more manual intervention and feature engineering, and the analysis process is relatively cumbersome and time-consuming. The speed of technical updates of conventional methods may not keep up with the rapid development of deep fake technology.

[0004] This shows that the existing technology still has problems such as low counterfeit identification accuracy, poor adaptability, difficulty in capturing deep fake features, and the need for more manual intervention and feature engineering. Summary of the invention

[0005] The purpose of the present invention is to provide an image deep fake tracking and tracing method based on the GANs model, which solves the problems of low forgery recognition accuracy, poor adaptability, difficulty in capturing deep fake features, and the need for more manual intervention and feature engineering.

[0006] To achieve the above object, the present invention provides a method for tracking and tracing deep forgeries of images based on a GANs model, comprising the following steps:

[0007] S1. Dataset preparation: Use crawler technology to automatically crawl a large number of real images and known forged images on the Internet as training sets, clean and annotate the data sets, and distinguish real samples from forged samples;

[0008] S2, feature engineering, using normalization or scaling to preprocess the image to make it suitable for the model input specifications, using data enhancement technology to increase data diversity and reduce overfitting;

[0009] S3, construct a generator and a discriminator, define the output dimension according to the resolution and number of color channels of the image, use random noise as input to build a CNN model to generate a forged image; define the output dimension according to the data type of the generator output, use the image output by the generator as input to build a CNN model to determine whether the input image is forged;

[0010] S4, training the generator and the discriminator, inputting the real image and the forged image into the discriminator at the same time, and after the discriminator is trained to a certain degree, starting to jointly train the generator and the discriminator;

[0011] S5. Tracking and tracing analysis: Tracking and tracing images through metadata analysis, network analysis and pattern recognition. Based on the spread trend, hot spots and time patterns of forged images, the ARIMA model is used to predict possible forgery activities in the future.

[0012] S6. Continuous learning and iteration. The system needs to continuously absorb new data to update the feature library and model to adapt to the evolution of deep fake technology.

[0013] Preferably, in step S3, the construction steps of the generator are:

[0014] The input layer accepts 256x256 RGB images, and its output dimension is 256x256x3;

[0015] The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1.

[0016] (to ensure that the output size is the same as the input size);

[0017] Define the number of input channels I, thereby determining the number of output channels O, and the number of parameters of each network layer is I*O*K, where K is the number of parameters of the convolution kernel;

[0018] Use boundary expansion method to keep the output size the same as the input size;

[0019] After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows:

[0020]

[0021] There are 4 fully connected layers in the network, the last of which is the output layer, which uses the Softmax activation function for multi-category classification as follows:

[0022] Calculate the exponential function: For each element in the input vector, calculate its exponential value

[0023] y=e z

[0024] Each element of the input vector is z, e z is an exponential function;

[0025] Softmax activation function:

[0026]

[0027] Among them, Z i is the i-th element in the input vector, K is the total number of categories, and e is the base of the natural logarithm.

[0028] Preferably, in step S3, the discriminator is constructed by:

[0029] According to the data type of the generator output, the input dimension of the discriminator is set to 256x256x3;

[0030] The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1;

[0031] Multiple convolutional layers are used to extract features, where the calculation formula for the convolution output result is as follows:

[0032] N=(W-F+2P) / S+1

[0033] Among them, N is the output size, W is the input size, F is the convolution kernel size, P is the padding value size, and S is the stride size;

[0034] After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows:

[0035]

[0036] There are 4 fully connected layers in the network, and the output layer uses the Sigmoid activation function for recognition. The expression of the Sigmoid activation function is as follows:

[0037]

[0038] Here, X is the input of the function and σ(X) is the output of the function.

[0039] Preferably, in step S4, at the initial stage of training, the discriminator is mainly trained, and the real image and the forged image are simultaneously input into the discriminator, and the parameters of the discriminator are adjusted through the back propagation algorithm so that it can correctly distinguish the real image from the forged image.

[0040] Preferably, in step S5, in the tracking and tracing analysis, the specific steps of metadata analysis are as follows:

[0041] Image metadata extraction: using a metadata extraction tool or ExifTool to extract metadata of the image in step S4, the metadata including file name, file size, resolution, shooting time, shooting location and GPS positioning information;

[0042] Based on the extracted timestamps and GPS coordinates, a time-location correlation map is constructed to analyze the shooting location and time sequence of the image, and to track the source and propagation path of the image;

[0043] Use image recognition technology to identify specific objects, signs or scenes in images and look for patterns or anomalies, identifying signs of forgery or tampering.

[0044] Preferably, in step S5, in the tracking and tracing analysis, the specific steps of network analysis are as follows:

[0045] Network graph construction: each user, web page, or account that publishes or forwards the image is considered a node. If there is a clear forwarding or citation relationship, an edge is established between the corresponding nodes. If there is no direct relationship but the time and content are similar, it is considered a weak connection and an edge is constructed.

[0046] Use network centrality indicators to identify key nodes, discover different communication communities through the Louvain algorithm, and understand the dissemination patterns and acceptance of images in different groups;

[0047] Use time series analysis to track changes in the speed and scope of image dissemination over time, and identify peak and decline periods of dissemination; combine user IP addresses or self-declared location information to analyze the geographic dissemination path of the image and identify the regional characteristics of the dissemination.

[0048] Preferably, in step S5, in the tracking and tracing analysis, the specific steps of pattern recognition are as follows:

[0049] Use feature matching algorithms to compare the similarity between the target image and the images in the database to find the same or similar image sources;

[0050] Based on the matching results, trace back to the source or key dissemination node where the image first appeared, and analyze the initial upload time, location, and uploader information of the image;

[0051] Analyze the image content, identify the objects, scenes, text, etc., and understand the communication context and intention of the image based on the context information;

[0052] Machine learning models are applied to identify whether image dissemination follows a specific propagation path, whether it is associated with a specific event or a popular trend over time. If the above characteristics exist, a feedback report is generated to display the path, change points and key nodes of image dissemination.

[0053] Preferably, in step S5, the specific steps of using the ARIMA model to predict possible counterfeiting activities in the future are:

[0054] Count the propagation trends, hot spots, and time patterns of forged images, draw time series graphs and autocorrelation graphs, and observe the stability of time series data;

[0055] Check whether the data is white noise, determine the model order based on the autocorrelation function ACF and partial autocorrelation function PACF graphs, and determine the optimal ARIMA model parameters p, b, q;

[0056] According to the order determination results, the ARIMA model is established by combining the autoregressive model AR, the moving average model MA and the difference I, and the prediction is carried out according to the selected model.

[0057] Preferably, the mathematical formula of the autocorrelation coefficient is as follows:

[0058]

[0059] Among them, ρ(t,s) is the autocorrelation coefficient, which represents the correlation between time point t and time point s;

[0060] γ(t,s) is the time series {X t}, which is used to measure the covariance between time point t and time point s;

[0061] DX t and DX s They represent the variance at time point t and time point s respectively.

[0062] Therefore, the present invention adopts a method for tracking and tracing deep forgeries of images based on a GANs model using the above structure, which has the following beneficial effects:

[0063] By training the adversarial model, the present invention can understand more complex image structures and deep features, and is therefore more accurate in distinguishing highly realistic forged images. By continuously learning the adversarial generation process, it can theoretically better adapt to new forgery techniques and maintain the effectiveness of tracking and tracing.

[0064] The present invention can be specially trained to identify the statistical regularities or generation patterns that are unique to these images generated by GANs, thereby improving tracking accuracy. Using an automated learning framework, once the model training is completed, a large number of images can be quickly and automatically processed, improving the efficiency of tracking and analysis.

[0065] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1This is a flow chart of a method for tracking and tracing deep fake images based on a GANs model in the present invention. DETAILED DESCRIPTION

[0067] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0068] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0069] Example

[0070] As shown in the figure, the present invention provides a method for tracking and tracing deep forgery of images based on a GANs model, comprising the following steps:

[0071] S1. Dataset preparation: Use crawler technology to automatically crawl a large number of real images and known forged images on the Internet as training sets, clean and annotate the data sets, and distinguish real samples from forged samples;

[0072] S2, feature engineering, using normalization or scaling to preprocess the image to make it suitable for the model input specifications, using data enhancement technology to increase data diversity and reduce overfitting;

[0073] S3, construct a generator and a discriminator, define the output dimension according to the resolution and number of color channels of the image, use random noise as input to build a CNN model to generate a forged image; define the output dimension according to the data type of the generator output, use the image output by the generator as input to build a CNN model to determine whether the input image is forged;

[0074] S4, training the generator and the discriminator, inputting the real image and the forged image into the discriminator at the same time, and after the discriminator is trained to a certain degree, starting to jointly train the generator and the discriminator;

[0075] S5. Tracking and tracing analysis: Tracking and tracing images through metadata analysis, network analysis and pattern recognition. Based on the spread trend, hot spots and time patterns of forged images, the ARIMA model is used to predict possible forgery activities in the future.

[0076] S6. Continuous learning and iteration. The system needs to continuously absorb new data to update the feature library and model to adapt to the evolution of deep fake technology.

[0077] In step S3, the generator construction steps:

[0078] The input layer accepts 256x256 RGB images, and its output dimension is 256x256x3;

[0079] The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1.

[0080] (to ensure that the output size is the same as the input size);

[0081] Define the number of input channels I, thereby determining the number of output channels O, and the number of parameters of each network layer is I*O*K, where K is the number of parameters of the convolution kernel;

[0082] Use boundary expansion method to keep the output size the same as the input size;

[0083] After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows:

[0084]

[0085] There are 4 fully connected layers in the network, the last of which is the output layer, which uses the Softmax activation function for multi-category classification as follows:

[0086] Calculate the exponential function: For each element in the input vector, calculate its exponential value

[0087] y=e z

[0088] Each element of the input vector is z, e z is an exponential function;

[0089] Softmax activation function:

[0090]

[0091] Among them, Z i is the i-th element in the input vector, K is the total number of categories, and e is the base of the natural logarithm.

[0092] In step S3, the discriminator is constructed as follows:

[0093] According to the data type of the generator output, the input dimension of the discriminator is set to 256x256x3;

[0094] The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1;

[0095] Multiple convolutional layers are used to extract features, where the calculation formula for the convolution output result is as follows:

[0096] N=(W-F+2P) / S+1

[0097] Among them, N is the output size, W is the input size, F is the convolution kernel size, P is the padding value size, and S is the stride size;

[0098] After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows:

[0099]

[0100] There are 4 fully connected layers in the network, and the output layer uses the Sigmoid activation function for recognition. The expression of the Sigmoid activation function is as follows:

[0101]

[0102] Here, X is the input of the function and σ(X) is the output of the function.

[0103] In step S4, at the initial stage of training, the discriminator is mainly trained, and the real image and the forged image are simultaneously input into the discriminator, and the parameters of the discriminator are adjusted through the back propagation algorithm so that it can correctly distinguish the real image from the forged image.

[0104] In the tracking and tracing analysis, the specific steps of metadata analysis are as follows:

[0105] Image metadata extraction: Use a metadata extraction tool or ExifTool to extract the metadata of the image in step S4.

[0106] Metadata includes file name, file size, resolution, shooting time, shooting location and GPS location information, which can be used to perform preliminary screening of images and quickly exclude images that do not meet the requirements.

[0107] Extract features from the filtered images, including color features, texture features, shape features, etc., and then use these features to compare and match images. According to the extracted timestamps and GPS coordinates, build a time and location association map, analyze the shooting location and time sequence of the image, and track the source and propagation path of the image.

[0108] The search results are verified through the information provided by metadata, such as shooting location, shooting equipment, etc., to ensure the accuracy and reliability of the search results.

[0109] Use image recognition technology to identify specific objects, signs or scenes in images and look for patterns or anomalies, identifying signs of forgery or tampering.

[0110] In step S5, in the tracking and tracing analysis, the specific steps of network analysis are as follows:

[0111] Network graph construction: each user, web page, or account that publishes or forwards the image is considered a node. If there is a clear forwarding or citation relationship, an edge is established between the corresponding nodes. If there is no direct relationship but the time and content are similar, it is considered a weak connection and an edge is constructed.

[0112] Use network centrality indicators to identify key nodes, discover different communication communities through the Louvain algorithm, and understand the dissemination patterns and acceptance of images in different groups;

[0113] Use time series analysis to track changes in the speed and scope of image dissemination over time, and identify peak and decline periods of dissemination; combine user IP addresses or self-declared location information to analyze the geographic dissemination path of the image and identify the regional characteristics of the dissemination.

[0114] In step S5, in the tracking and tracing analysis, the specific steps of pattern recognition are as follows:

[0115] Use feature matching algorithms to compare the similarity between the target image and the images in the database to find the same or similar image sources;

[0116] Based on the matching results, trace back to the source or key dissemination node where the image first appeared, and analyze the initial upload time, location, and uploader information of the image;

[0117] Analyze the image content, identify the objects, scenes, text, etc., and understand the communication context and intention of the image based on the context information;

[0118] Machine learning models are applied to identify whether image dissemination follows a specific propagation path, whether it is associated with a specific event or a popular trend over time. If the above characteristics exist, a feedback report is generated to display the path, change points and key nodes of image dissemination.

[0119] In step S5, the specific steps of using the ARIMA model to predict possible counterfeiting activities in the future are as follows:

[0120] Statistical analysis of the spread trend, hot spots, and time patterns of forged images, and observe the stability of time series data by drawing time series graphs, autocorrelation graphs, etc. If the data is not stable, differential processing is required to make it stable and observe the stability of time series data;

[0121] The model order is determined based on the autocorrelation function (ACF) and partial autocorrelation function (PACF) graphs. This step is to determine the parameters p (number of autoregressive terms), d (number of difference orders) and q (number of moving average terms) in the ARIMA model. In addition, the grid search method can also be used for order determination. ;

[0122] The mathematical formula for the autocorrelation coefficient is as follows:

[0123]

[0124] Among them, ρ(t,s) is the autocorrelation coefficient, which represents the correlation between time point t and time point s;

[0125] γ(t,s) is the time series {X t}, which is used to measure the covariance between time point t and time point s;

[0126] DX t and DX s They represent the variance at time point t and time point s respectively.

[0127] Based on the determined parameters p, d, and q, an ARIMA model is established. This process involves analyzing the data and fitting the model. The established model is tested to ensure that the model's prediction effect meets expectations. This may include testing the model's residuals to see if it is white noise, and the model's prediction accuracy.

[0128] According to the order determination results, the ARIMA model is established by combining the autoregressive model AR, the moving average model MA and the difference I, and the prediction is carried out according to the selected model.

[0129] Use the established ARIMA model to make forecasts. This can be a forecast for the future or a fitted forecast of historical data. Depending on the specific needs, an appropriate forecast period can be selected. Evaluate the forecast results to see how accurate and reliable the forecast is. This may involve comparing the difference between the forecast and the actual value, and evaluating the performance indicators of the model.

[0130] Therefore, the present invention adopts the above-mentioned image deep forgery tracking and tracing method based on the GANs model. By training the adversarial model, it can understand more complex image structures and deep features, and is therefore more accurate in distinguishing highly realistic forged images. By continuously learning the adversarial generation process, it can theoretically better adapt to new forgery technologies and maintain the effectiveness of tracking and tracing. It can be specially trained to identify the statistical laws or generation patterns unique to these images generated by GANs, thereby improving tracking accuracy. Using an automated learning framework, once the model training is completed, a large number of images can be quickly and automatically processed to improve the efficiency of tracking and analysis.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for tracing and identifying deep fake images based on a GANs model, characterized in that: The following steps are involved: S1. Dataset preparation: Use crawler technology to automatically crawl a large number of real images and known forged images on the Internet as training sets, clean and annotate the data sets, and distinguish real samples from forged samples; S2, feature engineering, using normalization or scaling to preprocess the image to make it suitable for the model input specifications, using data enhancement technology to increase data diversity and reduce overfitting; S3, construct a generator and a discriminator, define the output dimension according to the resolution and number of color channels of the image, use random noise as input to build a CNN model to generate a forged image; define the output dimension according to the data type of the generator output, use the image output by the generator as input to build a CNN model to determine whether the input image is forged; S4, training the generator and the discriminator, inputting the real image and the forged image into the discriminator at the same time, and after the discriminator is trained to a certain degree, starting to jointly train the generator and the discriminator; S5. Tracking and tracing analysis: Tracking and tracing images through metadata analysis, network analysis and pattern recognition. Based on the spread trend, hot spots and time patterns of forged images, the ARIMA model is used to predict possible forgery activities in the future. S6. Continuous learning and iteration. The system needs to continuously absorb new data to update the feature library and model to adapt to the evolution of deep fake technology.

2. According to claim 1, a method for tracking and tracing deep forgery of images based on a GANs model is characterized by: In step S3, the generator construction step: The input layer accepts 256x256 RGB images, and its output dimension is 256x256x3; The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1. (to ensure that the output size is the same as the input size); Define the number of input channels I, thereby determining the number of output channels O, and the number of parameters of each network layer is I*O*K, where K is the number of parameters of the convolution kernel; Use boundary expansion method to keep the output size the same as the input size; After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows: There are 4 fully connected layers in the network, the last of which is the output layer, which uses the Softmax activation function for multi-category classification as follows: Calculate the exponential function: For each element in the input vector, calculate its exponential value and=e z Each element of the input vector is z, e z is an exponential function; Softmax activation function: Among them, Z i is the i-th element in the input vector, K is the total number of categories, and e is the base of the natural logarithm.

3. According to claim 1, a method for tracing and identifying deep forgeries of images based on a GANs model is characterized by: In step S3, the discriminator is constructed as follows: According to the data type of the generator output, the input dimension of the discriminator is set to 256x256x3; The convolution layer uses a small 3x3 convolution kernel and the stride is set to 1; Multiple convolutional layers are used to extract features, where the calculation formula for the convolution output result is as follows: N=(W-F+2P) / S+1 Among them, N is the output size, W is the input size, F is the convolution kernel size, P is the padding value size, and S is the stride size; After each convolutional layer, a 2x2 max pooling layer is used for spatial downsampling, which is expressed as follows: There are 4 fully connected layers in the network, and the output layer uses the Sigmoid activation function for recognition. The expression of the Sigmoid activation function is as follows: Here, X is the input of the function and σ(X) is the output of the function.

4. According to claim 1, a method for tracking and tracing deep forgery of images based on a GANs model is characterized by: In step S4, at the initial stage of training, the discriminator is mainly trained, and the real image and the forged image are simultaneously input into the discriminator, and the parameters of the discriminator are adjusted through the back propagation algorithm so that it can correctly distinguish the real image from the forged image.

5. According to claim 1, a method for tracing and identifying deep forgeries of images based on a GANs model is characterized by: In step S5, in the tracking and tracing analysis, the specific steps of metadata analysis are as follows: Image metadata extraction: using a metadata extraction tool or ExifTool to extract metadata of the image in step S4, the metadata including file name, file size, resolution, shooting time, shooting location and GPS positioning information; Based on the extracted timestamps and GPS coordinates, a time-location correlation map is constructed to analyze the shooting location and time sequence of the image, and to track the source and propagation path of the image; Use image recognition technology to identify specific objects, signs or scenes in images and look for patterns or anomalies, identifying signs of forgery or tampering.

6. According to claim 1, a method for tracking and tracing deep forgery of images based on a GANs model is characterized by: In step S5, in the tracking and tracing analysis, the specific steps of network analysis are as follows: Network graph construction: each user, web page, or account that publishes or forwards the image is considered a node. If there is a clear forwarding or citation relationship, an edge is established between the corresponding nodes. If there is no direct relationship but the time and content are similar, it is considered a weak connection and an edge is constructed. Use network centrality indicators to identify key nodes, discover different communication communities through the Louvain algorithm, and understand the dissemination patterns and acceptance of images in different groups; Use time series analysis to track changes in the speed and scope of image dissemination over time, and identify peak and decline periods of dissemination; combine user IP addresses or self-declared location information to analyze the geographic dissemination path of the image and identify the regional characteristics of the dissemination.

7. According to claim 1, a method for tracking and tracing deep forgery of images based on a GANs model is characterized by: In step S5, in the tracking and tracing analysis, the specific steps of pattern recognition are as follows: Use feature matching algorithms to compare the similarity between the target image and the images in the database to find the same or similar image sources; Based on the matching results, trace back to the source or key dissemination node where the image first appeared, and analyze the initial upload time, location, and uploader information of the image; Analyze the image content, identify the objects, scenes, text, etc., and understand the communication context and intention of the image based on the context information; Machine learning models are applied to identify whether image dissemination follows a specific propagation path, whether it is associated with a specific event or a popular trend over time. If the above characteristics exist, a feedback report is generated to display the path, change points and key nodes of image dissemination.

8. According to claim 1, a method for tracking and tracing deep forgery of images based on a GANs model is characterized by: In step S5, the specific steps of using the ARIMA model to predict possible counterfeiting activities in the future are as follows: Count the propagation trends, hot spots, and time patterns of forged images, draw time series graphs and autocorrelation graphs, and observe the stability of time series data; Check whether the data is white noise, determine the model order based on the autocorrelation function ACF and partial autocorrelation function PACF graphs, and determine the optimal ARIMA model parameters p, b, q; According to the order determination results, the ARIMA model is established by combining the autoregressive model AR, the moving average model MA and the difference I, and the prediction is carried out according to the selected model.

9. The method for tracing and identifying deep forgeries of images based on a GANs model according to claim 8, characterized in that: The mathematical formula for the autocorrelation coefficient is as follows: Among them, ρ(t,s) is the autocorrelation coefficient, which represents the correlation between time point t and time point s; γ(t,s) is the time series {X t }, which is used to measure the covariance between time point t and time point s; DX t and DX s Represent the variance at time point t and time point s respectively.

Images