Deepfake detection method and system, computer device and storage medium

Abstract

The application relates to a deepfake detection method, system, computer device and storage medium, the method comprising: acquiring a data set containing a deepfake image, the data set having multiple characteristic domains; performing a causal relationship analysis on the data set with respect to characteristics to obtain a characteristic causal diagram; constructing a deep neural network model for deepfake detection; inputting image data to be detected into the deep neural network model to perform deepfake detection, and outputting a deepfake detection result corresponding to the image data. The method is guided by a causal inference theory to analyze data bias, and an invariant risk minimization strategy is introduced as a loss constraint of the detection model, meanwhile, the detection task is disassembled into multiple subtasks, and specific information is retained and irrelevant bias is eliminated through a feature alignment and joint training strategy, efficient feature extraction and fusion are realized, and the detection accuracy of the deepfake detection model in a complex and changeable scene is improved.

Description

Technical Field

[0001] This invention is applicable to the field of artificial intelligence technology applications, and in particular relates to a deepfake detection method, system, computer equipment, and storage medium. Background Technology

[0002] Deepfake is a technique that combines and overlays images or videos onto source images or videos, splicing together a person's voice, facial expressions, and body movements to create fake content. While the widespread use of deepfake technology in recent years has brought diversity to media creation, its synthesized content is also easily abused in scenarios such as identity fraud, dissemination of misinformation, and digital harassment.

[0003] Most existing deepfake detection methods are based on deep neural networks, which are typically trained using the Empirical Risk Minimization (ERM) method. However, this method is prone to capturing irrelevant bias information in the data, resulting in poor generalization performance in real-world environments.

[0004] On the other hand, the datasets used by deep neural networks are usually constructed by simulating fake data from a large amount of real data. However, since the images or videos of people are affected by different ethnic deepfake techniques and other environmental factors, the data has obvious selective bias and confounding factors, which makes the existing detection methods less than ideal in practical applications. Summary of the Invention

[0005] This invention provides a deepfake detection method, system, computer equipment, and storage medium, aiming to solve the technical problems of data bias and poor detection accuracy in existing methods.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a deepfake detection method, comprising the following steps:

[0007] Obtain a dataset containing multiple deepfake images, the dataset having multiple feature domains;

[0008] A causal relationship analysis of the features is performed on the dataset to obtain a feature causal graph;

[0009] A deep neural network model for deepfake detection is constructed, comprising a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein:

[0010] The detection network is used to achieve deepfake detection;

[0011] The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains;

[0012] The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causality-invariant feature.

[0013] The image data to be detected is input into the deep neural network model for deepfake detection, and the deepfake detection result corresponding to the image data is output.

[0014] Furthermore, in the causal invariance learning module, the step of performing bias analysis based on the feature causal graph to identify causal invariant features includes the following sub-steps:

[0015] The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features.

[0016] Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

[0017] Furthermore, the causal invariance learning module is also used for:

[0018] The causal identification process of the feature causal graph is optimized by using a gradient weighting method.

[0019] Furthermore, the step of integrating and outputting all the aforementioned detection sub-results is specifically as follows:

[0020] Each detection sub-result is assigned a weight, and the different detection sub-results are then weighted and averaged according to the weights before being output; or,

[0021] The fusion layer is used to fuse all the detection sub-results and then output the data.

[0022] Furthermore, in the causal invariance learning module, when different subtasks perform feature alignment on the causal invariant features, the maximum mean difference alignment is used.

[0023] Furthermore, before the step of inputting the image data to be detected into the deep neural network model for deepfake detection and outputting the deepfake detection result corresponding to the image data, the method further includes:

[0024] The image data is preprocessed, and the preprocessing method includes at least one of data augmentation, labeling, and normalization.

[0025] Secondly, the present invention also provides a deepfake detection system, comprising:

[0026] A data acquisition unit is used to acquire a dataset containing multiple deepfake images, the dataset having multiple feature domains;

[0027] A causal construction unit is used to perform causal relationship analysis on the dataset regarding features to obtain a feature causal graph;

[0028] A model building unit is used to build a deep neural network model for deepfake detection. The deep neural network model includes a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein:

[0029] The detection network is used to achieve deepfake detection;

[0030] The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains;

[0031] The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causality-invariant feature.

[0032] The detection unit is used to input the image data to be detected into the deep neural network model for deepfake detection and output the deepfake detection result corresponding to the image data.

[0033] Furthermore, the causal invariance learning module is also used for:

[0034] The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features.

[0035] Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

[0036] Thirdly, the present invention also provides a computer device, comprising: a memory, a processor, and a deepfake detection program stored in the memory and executable on the processor, wherein the processor, when executing the deepfake detection program, implements the steps of the deepfake detection method as described in any of the above embodiments.

[0037] Fourthly, the present invention also provides a storage medium storing a deepfake detection program, wherein the deepfake detection program, when executed by a processor, implements the steps of the deepfake detection method as described in any of the above embodiments.

[0038] The beneficial effects achieved by this invention are that it proposes a deepfake detection method. This method uses causal inference theory to guide data bias analysis, introduces an invariant risk minimization strategy as a loss constraint for the detection model, and decomposes the detection task into multiple sub-tasks. It also retains specific information and eliminates irrelevant biases through feature alignment and joint training strategies, thereby achieving efficient feature extraction and fusion and improving the detection accuracy of the deepfake detection model in complex and variable scenarios. Attached Figure Description

[0039] Figure 1 This is a schematic flowchart of the deepfake detection method provided in the embodiments of the present invention;

[0040] Figure 2 This is a schematic diagram of a cause-effect graph provided in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the logical structure of the deep neural network model provided in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the structure of the deepfake detection system provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a deepfake detection method provided in an embodiment of the present invention. The deepfake detection method includes the following steps:

[0046] S101. Obtain a dataset containing multiple deepfake images, wherein the dataset has multiple feature domains.

[0047] A feature domain is a definition and description of a certain aspect of data. Each feature domain represents a specific attribute or feature type of the data. In practice, since it is almost impossible to collect large-scale unbiased data directly from the real world, existing deepfake image datasets are generated by using different forgery methods on real images. Therefore, the datasets often have multi-domain characteristics due to factors such as the race of people in the images, the forgery method, and the shooting equipment. That is, each domain has a specific feature distribution.

[0048] S102. Perform causal relationship analysis on the dataset regarding features to obtain a feature causal graph.

[0049] Causal relationship analysis is a method used to determine causal relationships between features, aiming to understand how one feature influences another, and the mechanism and extent of this influence. In deepfake images, certain features exhibit causal patterns different from those in normal images. For example, the lighting effects of objects in a fake image may not conform to natural causal relationships with the lighting characteristics of the surrounding environment. Causal analysis of these features can help provide a basis for subsequent detection of whether an image has been tampered with, and if so, where and how. A causal graph is a graphical tool used to visually represent causal relationships between variables. It presents the causal connections and influence paths between various factors in a graphical way. An example of a causal graph is shown below. Figure 2 As shown, Figure 2 This illustrates the causal relationships between features Z, S, X, and Y. Solid arrows pointing from Z to X and Y indicate that Z is the cause of X and Y; arrows pointing from S to X... S The solid arrow pointing to Y indicates that S is X. S And the reason for Y; and from X V Pointing to Y and from X S A dashed arrow pointing to Y may indicate a potential, yet-to-be-defined, or further-explored relationship, contrasting with a solid arrow indicating a clear causal relationship.

[0050] In summary, feature cause-effect graphs can intuitively represent the causal relationships between various features, which is beneficial for data analysis.

[0051] S103. Construct a deep neural network model for deepfake detection. The deep neural network model includes a detection network, a causal invariance learning module, and a multi-domain causality preservation module, specifically, as follows: Figure 3 As shown, Figure 3 This is a schematic diagram of the logical structure of the deep neural network model provided in an embodiment of the present invention, wherein:

[0052] The detection network is used to achieve deepfake detection;

[0053] The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method (IRM), so that the detection network focuses on the causal invariant features when performing detection in different feature domains;

[0054] The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causal invariant feature.

[0055] In this embodiment of the invention, the detection network serves as the backbone network for image detection and discrimination. It can be implemented based on existing model architectures, such as Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs). This embodiment of the invention adds a Causal Invariance Learning (CIL) module and a Multi-domain Causal Preservation (MCP) module to the backbone network for optimization.

[0056] Specifically, the step of performing bias analysis based on the feature causal invariance learning module to identify causal invariant features includes the following sub-steps:

[0057] The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features.

[0058] Based on the bias analysis of the causal and non-causal features of different multi-domain datasets, the causal invariant features that maintain the causal relationship in different feature domains are obtained.

[0059] Bias analysis is a data processing method that includes statistical analysis and analysis of variance. In this embodiment of the invention, the original dataset is first divided into multiple multi-domain datasets by identifying different feature domains present in the dataset. For example, deepfake images may include information such as color, texture, semantics, and geometry. Then, based on the feature causal graph, the features represented by each node in the graph and the causal relationship directions represented by the edges are analyzed to identify causal and non-causal features. Causal features are those that have a causal influence on other features, while non-causal features are those that do not have a causal influence on other features. Although non-causal features do not directly participate in causal relationships, they can be used as control variables or covariates to help eliminate interference from other factors, thereby more accurately inferring causal relationships.

[0060] Subsequently, bias analysis and comparison were performed on different multi-domain datasets to obtain causal invariant features. Causal invariant features refer to causal features that maintain a relatively stable causal relationship in multi-domain datasets with different feature domains.

[0061] It is understandable that the causal invariant features identified in different multi-domain datasets have a certain degree of universality and stability across different datasets. Therefore, detecting and identifying these causal invariant features can effectively improve the robustness of the detection model.

[0062] The causal invariance learning module is also used for:

[0063] The causal identification process of the feature causal graph is optimized by using a gradient weighting method.

[0064] Gradient weighting is a commonly used optimization technique that can enhance the influence of important features and suppress the interference of minor features by calculating gradients and assigning weights during the causal identification process of feature causal graphs in the detection model. This improves the accuracy and reliability of causal identification.

[0065] The multi-domain causality preservation module breaks down the detection task of the detection model into multiple sub-tasks, and while sharing global causal invariant features among the sub-tasks through feature alignment, it preserves local domain features. For example... Figure 3 As shown, each subtask executes the detection task by relying on the loss function constructed based on the invariant risk minimization method, and obtains the corresponding detection sub-result.

[0066] The specific steps for integrating and outputting all the aforementioned detection sub-results are as follows:

[0067] Each detection sub-result is assigned a weight, and the different detection sub-results are then weighted and averaged according to the weights before being output; or,

[0068] The fusion layer is used to fuse all the detection sub-results and then output the data.

[0069] By integrating all the detection sub-results, the overall deep neural network model achieves both high generalization and sensitivity to local details.

[0070] In the causal invariance learning module, when different subtasks perform feature alignment on the causal invariant features, the maximum mean difference alignment is used.

[0071] During implementation, the constructed deep neural network model still needs to undergo a training process based on the aforementioned loss function and multi-domain training set in order to obtain a model with accurate detection capabilities.

[0072] S104. Input the image data to be detected into the deep neural network model to perform deepfake detection, and output the deepfake detection result corresponding to the image data.

[0073] Before the step of inputting the image data to be detected into the deep neural network model for deepfake detection and outputting the deepfake detection result corresponding to the image data, the method further includes:

[0074] The image data is preprocessed, and the preprocessing method includes at least one of data augmentation, labeling, and normalization.

[0075] The beneficial effects achieved by this invention are that it proposes a deepfake detection method. This method uses causal inference theory to guide data bias analysis, introduces an invariant risk minimization strategy as a loss constraint for the detection model, and decomposes the detection task into multiple sub-tasks. It also retains specific information and eliminates irrelevant biases through feature alignment and joint training strategies, thereby achieving efficient feature extraction and fusion and improving the detection accuracy of the deepfake detection model in complex and variable scenarios.

[0076] This invention also provides a deepfake detection system 200, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the structure of the deepfake detection system provided in an embodiment of the present invention, which includes:

[0077] The data acquisition unit 201 is used to acquire a dataset containing multiple deepfake images, the dataset having multiple feature domains;

[0078] The causal construction unit 202 is used to perform causal relationship analysis on the dataset regarding features to obtain a feature causal graph;

[0079] Model building unit 203 is used to build a deep neural network model for deepfake detection. The deep neural network model includes a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein:

[0080] The detection network is used to achieve deepfake detection;

[0081] The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains;

[0082] The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causality-invariant feature.

[0083] The detection unit 204 is used to input the image data to be detected into the deep neural network model for deepfake detection and output the deepfake detection result corresponding to the image data.

[0084] Furthermore, the causal invariance learning module is also used for:

[0085] The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features.

[0086] Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

[0087] The deepfake detection system 200 can implement the steps in the deepfake detection method as described in the above embodiments and can achieve the same technical effect, as described in the above embodiments, and will not be repeated here.

[0088] This invention also provides a computer device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. The computer device 300 includes: a memory 302, a processor 301, and a deepfake detection program stored in the memory 302 and capable of running on the processor 301.

[0089] The processor 301 calls the deepfake detection program stored in the memory 302 and executes the steps in the deepfake detection method provided in this embodiment of the invention. Please refer to... Figure 1 The deepfake detection method specifically includes the following steps:

[0090] S101. Obtain a dataset containing multiple deepfake images, wherein the dataset has multiple feature domains.

[0091] S102. Perform causal relationship analysis on the dataset regarding features to obtain a feature causal graph.

[0092] S103. Construct a deep neural network model for deepfake detection. The deep neural network model includes a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein:

[0093] The detection network is used to achieve deepfake detection;

[0094] The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains;

[0095] The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causal invariant feature.

[0096] Specifically, the step of performing bias analysis based on the feature causal invariance learning module to identify causal invariant features includes the following sub-steps:

[0097] The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features.

[0098] Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

[0099] The causal invariance learning module is also used for:

[0100] The causal identification process of the feature causal graph is optimized by using a gradient weighting method.

[0101] The specific steps for integrating and outputting all the aforementioned detection sub-results are as follows:

[0102] Each detection sub-result is assigned a weight, and the different detection sub-results are then weighted and averaged according to the weights before being output; or,

[0103] The fusion layer is used to fuse all the detection sub-results and then output the data.

[0104] In the causal invariance learning module, when different subtasks perform feature alignment on the causal invariant features, the maximum mean difference alignment is used.

[0105] S104. Input the image data to be detected into the deep neural network model to perform deepfake detection, and output the deepfake detection result corresponding to the image data.

[0106] Before the step of inputting the image data to be detected into the deep neural network model for deepfake detection and outputting the deepfake detection result corresponding to the image data, the method further includes:

[0107] The image data is preprocessed, and the preprocessing method includes at least one of data augmentation, labeling, and normalization.

[0108] The computer device 300 provided in this embodiment of the invention can implement the steps in the deepfake detection method as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0109] This invention also provides a storage medium storing a deepfake detection program. When the deepfake detection program is executed by a processor, it implements the various processes and steps in the deepfake detection method provided in this invention and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0110] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by instructing related hardware (such as mobile phones, computers, servers, air conditioners, or network devices, etc.) through a deepfake detection program. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0111] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0112] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form without departing from the spirit and scope of the claims of the present invention, and all such changes are within the protection scope of the present invention.

Claims

1. A deepfake detection method, characterized in that, Includes the following steps: Obtain a dataset containing multiple deepfake images, the dataset having multiple feature domains; A causal relationship analysis of the features is performed on the dataset to obtain a feature causal graph; A deep neural network model for deepfake detection is constructed, comprising a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein: The detection network is used to achieve deepfake detection; The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains; The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causality-invariant feature. The image data to be detected is input into the deep neural network model for deepfake detection, and the deepfake detection result corresponding to the image data is output.

2. The deepfake detection method according to claim 1, characterized in that, The step of performing bias analysis based on the feature causal invariance learning module to identify causal invariance features includes the following sub-steps: The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features. Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

3. The deepfake detection method according to claim 2, characterized in that, The causal invariance learning module is also used for: The causal identification process of the feature causal graph is optimized by using a gradient weighting method.

4. The deepfake detection method according to claim 1, characterized in that, The specific steps for integrating and outputting all the aforementioned detection sub-results are as follows: Each detection sub-result is assigned a weight, and the different detection sub-results are then weighted and averaged according to the weights before being output; or, The fusion layer is used to fuse all the detection sub-results and then output the data.

5. The deepfake detection method according to claim 1, characterized in that, In the causal invariance learning module, when different subtasks perform feature alignment on the causal invariant features, the maximum mean difference alignment is used.

6. The deepfake detection method according to claim 1, characterized in that, Before the step of inputting the image data to be detected into the deep neural network model for deepfake detection and outputting the deepfake detection result corresponding to the image data, the method further includes: The image data is preprocessed, and the preprocessing method includes at least one of data augmentation, labeling, and normalization.

7. A deepfake detection system, characterized in that, include: A data acquisition unit is used to acquire a dataset containing multiple deepfake images, the dataset having multiple feature domains; A causal construction unit is used to perform causal relationship analysis on the dataset regarding features to obtain a feature causal graph; A model building unit is used to build a deep neural network model for deepfake detection. The deep neural network model includes a detection network, a causal invariance learning module, and a multi-domain causality preservation module, wherein: The detection network is used to achieve deepfake detection; The causal invariance learning module is used to perform bias analysis based on the feature causal graph, identify causal invariant features, and construct a loss function for the detection network to perform detection in different feature domains based on the invariant risk minimization method, so that the detection network focuses on the causal invariant features when performing detection in different feature domains; The multi-domain causality preservation module is used to divide the detection task of the detection network into multiple sub-tasks. Each sub-task corresponds to a feature for detection and outputs a corresponding detection sub-result. All the detection sub-results are integrated and output. Among them, different sub-tasks perform feature alignment on the causality-invariant feature. The detection unit is used to input the image data to be detected into the deep neural network model for deepfake detection and output the deepfake detection result corresponding to the image data.

8. The deepfake detection system according to claim 7, characterized in that, The causal invariance learning module is also used for: The dataset is divided according to the feature domain to obtain multiple multi-domain datasets. Causal identification is performed on the features in the multi-domain datasets according to the feature causal graph to obtain the causal features and non-causal features. Based on the causal and non-causal features of different multi-domain datasets, bias analysis is performed to obtain features in which the causal relationship remains unchanged in different feature domains, and these features are defined as the causal invariant features.

9. A computer device, characterized in that, include: The device includes a memory, a processor, and a deepfake detection program stored in the memory and executable on the processor, wherein the processor, when executing the deepfake detection program, implements the steps of the deepfake detection method as described in any one of claims 1-6.

10. A storage medium, characterized in that, The storage medium stores a deepfake detection program, which, when executed by a processor, implements the steps of the deepfake detection method as described in any one of claims 1-6.

Images