A deep fake image detection method and system integrating noise perception

Abstract

The present disclosure belongs to the field of machine vision technology, and specifically relates to a deep fake image detection method and system integrating noise perception, comprising: obtaining an image to be detected; extracting features of an RGB stream according to the acquired image to be detected and dual attention calculation to obtain multi-scale enhanced features of the image; extracting noise domain features of the acquired image to be detected, and obtaining image noise stream features according to the obtained multi-scale enhanced features of the image and the extracted noise domain features; fusing the obtained image noise stream features with the extracted RGB stream features through dual-stream features to obtain image dual-stream fusion enhanced features; completing image forgery detection according to the obtained image dual-stream fusion enhanced features and local consistency calculation.

Description

Technical Field

[0001] The present disclosure belongs to the field of machine vision technology, and in particular relates to a deep fake image detection method and system integrating noise perception. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Deepfake refers to the use of deep learning methods such as generative adversarial networks to replace the face of a target person in the source video to create false audio and video information. Since this type of technology can generate high-quality fake videos that are difficult to distinguish with the naked eye, it can be easily abused by criminals and bring serious social impact. Therefore, the research on deepfake information detection is an important direction in the field of vision.

[0004] At present, various deep fake detection methods have emerged one after another, which can detect fake faces through manual features, such as inconsistent head posture, heartbeat rhythm or eye blinking, etc. With the widespread application of deep learning methods, researchers have begun to explore the use of deep neural networks to capture high-level semantic features of the spatial domain of fake images to improve detection effects.

[0005] According to the inventors, some deep fake algorithms use image compression post-processing algorithms to destroy artifacts, making them undetectable in color space, but still leaving traces of tampering in high-frequency information. In order to improve the robustness of the detection model, noise information is introduced into forged image detection as a high-frequency information. Real faces are coherent in all blocks, while forged faces are fused together from different faces, which produces inconsistent information in certain locations. Therefore, the concept of consistency learning, which is used in the field of general image detection to detect image copy, move, or copy and paste operations, is introduced into the field of deep fake detection. Consistency learning usually uses the method of calculating local similarity to measure the similarity between each block of the image to detect the consistency of the image spatial domain.

[0006] Although the above methods can effectively solve the problem of deep fake detection, the following problems still limit their performance improvement:

[0007] (1) Detection methods based on convolutional neural networks tend to focus on preferred salient areas, making it difficult to capture other artifacts that are not obvious but distinguishable, and are prone to overfitting.

[0008] (2) The high-frequency noise features currently introduced in deep fake detection are mostly extracted manually, which lacks certain flexibility and has poor adaptability when dealing with some post-processed fake videos.

[0009] (3) Deep fake detection methods based on local similarity features ignore global features that also contain discriminative information, thus limiting the ability to extract discriminative features. Summary of the invention

[0010] In order to solve the above problems, the present disclosure proposes a deep fake image detection method and system integrating noise perception. In view of the low accuracy and poor generalization caused by the inability of existing deep fake detection algorithms to fully extract discriminative features, a progressive multi-scale deep fake detection model integrating noise perception is constructed. Image detection is realized based on the constructed detection model, thereby improving detection accuracy and generalization ability.

[0011] According to some embodiments, a first solution of the present disclosure provides a deep fake image detection method integrating noise perception, which adopts the following technical solutions:

[0012] A deep fake image detection method integrating noise perception, comprising:

[0013] Acquire the image to be detected;

[0014] According to the acquired image to be detected and the dual attention calculation, the features of the RGB stream are extracted to obtain the multi-scale enhanced features of the image;

[0015] Extracting noise domain features from the acquired image to be detected, and obtaining image noise flow features based on the obtained image multi-scale enhancement features and the extracted noise domain features;

[0016] The noise stream features of the obtained image and the features of the extracted RGB stream are fused into two stream features to obtain the image two stream fusion enhanced features;

[0017] Image forgery detection is completed based on the obtained image dual-stream fusion enhanced features and local consistency calculation.

[0018] As a further technical limitation, in the process of extracting features of the RGB stream, a progressive spatial attention module is used to adaptively learn the enhancement weights and suppression weights of the image spatial domain features through dual attention calculations, and the original feature map is weightedly fused with the enhancement weights and suppression weights respectively to obtain an enhanced feature map and a suppressed feature map; the enhanced feature map is used for dual-stream feature fusion, and the suppressed feature map is re-input into the dual attention calculation of the next stage after convolution, so as to mine potential features of image areas other than attention.

[0019] As a further technical limitation, high-frequency noise information is obtained through an adaptive convolution kernel, and the obtained high-frequency noise information and phase spectrum are used together as the input of the noise stream to realize the extraction of noise domain features, and are complementary fused with the multi-scale enhanced features in the spatial domain obtained from the RGB stream to obtain the image dual-stream fusion enhancement features.

[0020] As a further technical limitation, the image to be detected is preprocessed to obtain a grayscale image, and the obtained grayscale image is converted into a frequency domain grayscale image based on discrete Fourier transform. The phase of the obtained frequency domain grayscale image is used as the phase spectrum information of the input image, and the phase spectrum reconstruction image is obtained in combination with discrete Fourier inverse transform.

[0021] As a further technical limitation, the acquired image to be detected is preprocessed to obtain a grayscale image, an adaptive SRM filter is used, a constrained convolution layer is introduced, and the obtained grayscale image is used as the input of the constrained convolution layer to obtain the SRM noise information map of the image.

[0022] As a further technical limitation, the dual-stream fusion enhancement features are complementarily fused to obtain multi-scale enhancement features, and the multi-scale enhancement features are input into the convolution block to obtain global features containing high-level semantic information. At the same time, the multi-scale enhancement features are divided into multiple local blocks, and local consistency information is obtained by calculating the similarity between each local block. The obtained global features and local consistency information are combined to complete forgery detection.

[0023] Furthermore, a bilinear interpolation algorithm is used to transform multi-scale enhanced feature maps into the same scale and splice them into feature maps along the channel dimension to obtain local consistency information.

[0024] According to some embodiments, a second solution of the present disclosure provides a deep fake image detection system integrating noise perception, which adopts the following technical solutions:

[0025] A deep fake image detection system integrating noise perception, comprising:

[0026] An acquisition module, configured to acquire an image to be detected;

[0027] A progressive spatial attention module is configured to extract features of the RGB stream according to the acquired image to be detected and the dual attention calculation to obtain multi-scale enhanced features of the image;

[0028] The noise perception module is configured to extract noise domain features from the acquired image to be detected, obtain image noise flow features according to the obtained image multi-scale enhancement features and the extracted noise domain features; perform dual-stream feature fusion on the obtained image noise flow features and the extracted RGB flow features to obtain image dual-stream fusion enhancement features;

[0029] The detection module is configured to complete the image forgery detection based on the obtained image dual-stream fusion enhancement features and local consistency calculation.

[0030] According to some embodiments, a third solution of the present disclosure provides a computer-readable storage medium, which adopts the following technical solution:

[0031] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the deep fake image detection method integrating noise perception as described in the first aspect of the present disclosure.

[0032] According to some embodiments, a fourth solution of the present disclosure provides an electronic device, which adopts the following technical solution:

[0033] An electronic device comprises a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for detecting deep fake images integrating noise perception as described in the first aspect of the present disclosure are implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention uses a progressive spatial attention module to successively focus on different areas of the image, and enhances the prominent areas at each stage to obtain multi-scale discriminative features, thereby improving the model's ability to capture artifacts; a two-stream detection model is constructed by combining the noise information and RGB information of the image to cope with the possible impact of post-processing algorithms in different scenarios; multi-scale local similarity features and global features obtained by multi-scale fusion are used to jointly perform forgery detection to improve the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0037] Figure 1 is a flow chart of a deep fake image detection method integrating noise perception in the first embodiment of the present disclosure;

[0038] Figure 2 is an architecture diagram of the image detection model in the first embodiment of the present disclosure;

[0039] Figure 3 is a dual attention calculation structure diagram in the first embodiment of the present disclosure;

[0040] Figure 4 It is a structural block diagram of the deep fake image detection system integrating noise perception in the second embodiment of the present disclosure. DETAILED DESCRIPTION

[0041] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0044] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0045] Embodiment 1

[0046] Embodiment 1 of the present disclosure introduces a deep fake image detection method integrating noise perception.

[0047] like Figure 1 A deep fake image detection method integrating noise perception is shown, comprising:

[0048] Acquire the image to be detected;

[0049] According to the acquired image to be detected and the dual attention calculation, the features of the RGB stream are extracted to obtain the multi-scale enhanced features of the image;

[0050] Extracting noise domain features from the acquired image to be detected, and obtaining image noise flow features based on the obtained image multi-scale enhancement features and the extracted noise domain features;

[0051] The noise stream features of the obtained image and the features of the extracted RGB stream are fused into two stream features to obtain the image two stream fusion enhanced features;

[0052] Image forgery detection is completed based on the obtained image dual-stream fusion enhanced features and local consistency calculation.

[0053] The image detection model in this embodiment adopts a progressive network learning model that integrates noise perception, and its architecture is shown in the following figure: Figure 2 In this embodiment, the image detection model is a two-stream network consisting of an RGB stream and an adaptive noise perception stream, specifically including:

[0054] Progressive spatial attention submodule: This module completes the feature extraction of the RGB stream, and its core is the dual attention calculation (Dual Feature Selective Module, DFSM for short). DFSM can adaptively learn two weights of image spatial domain features, namely enhancement and suppression weights, and perform weighted fusion with the original feature map to obtain enhanced feature maps and suppressed feature maps respectively; the enhanced feature map prepares for subsequent fusion, and the suppressed feature map is input again to the next stage of DFSM after convolution, forcing the model to mine other potential features of the image area outside of attention. After three DFSM learning cycles, three enhanced features of different scales are gradually obtained, providing the prerequisite for the subsequent dual-stream feature fusion;

[0055] (2) Noise-aware Module (NAM): This module uses the designed adaptive convolution kernel to obtain high-frequency noise information, and uses it together with the phase spectrum as the input of the noise stream to extract noise domain features. It then complementally fuses the spatial domain multi-scale features obtained from the RGB stream to obtain dual-stream fusion enhanced features, providing more comprehensive features for forgery detection.

[0056] Based on the progressive spatial attention submodule and the noise perception submodule, the dual-stream fusion enhanced features obtained by the noise perception submodule are first fused with each other. On the one hand, they are input into the convolution block to obtain global features covering high-level semantic information. On the other hand, they are divided into multiple local blocks, and the local consistency information is obtained by calculating the similarity between each block. The global features and local consistency information are combined to complete forgery detection.

[0057] In this embodiment, the progressive spatial attention submodule is as follows: Figure 2 As shown in the RGB stream branch in the figure, it realizes progressive multi-scale enhanced feature learning by inserting the dual attention calculation module between the convolution blocks of the network. The dual attention calculation structure is shown in Figure 3 As shown, assuming there is a face image I, the feature obtained by convolution is represented as X∈R H×W×C , the calculation process is as follows:

[0058] First, the feature X is divided into k spatial parts X along the width dimension. i ∈R H×(W / k)×C ,i∈[1,k]. Then it is input into the 1×1 convolution block for dimensionality reduction, and the initial weight representation S of each spatial part is obtained after the normalization layer and the ReLU function activation output i ∈R H×(Wk)×1 Inspired by the Convolutional Block Attention Module (CBAM), in S iThe average pooling and maximum pooling operations are used respectively and fused to obtain the standardized weight representation S of each part. softmax ∈R 1×1×W / k .

[0059] In order to better highlight the salient regions and guide the network to focus on other potential regional features, this embodiment designs a weight enhancement suppression mechanism to further enhance the normalized weight S softmax Learn to enhance the weight W boost and the suppression weight W suppress as follows:

[0060] Assume the input normalized weight S softmax =(s1,...,s L ), L = W / k, the learned enhancement weights and suppression weights are The enhancement weight is calculated as:

[0061]

[0062] Where i∈[1,L], T b is the enhancement threshold, taken from the normalized weight S softmax The nth largest weight in the s is set to L / 4 by default. i If it is greater than or equal to the threshold, it can be regarded as the enhanced spatial part, and its weight is set to the hyperparameter α; otherwise, it is set to 0, that is, the part is not enhanced.

[0063] The suppression weight is calculated as:

[0064]

[0065] Among them, T s is the suppression threshold, and the setting method is the same as the enhancement threshold. β is a hyperparameter that represents the weight of the suppression space part.

[0066] The learned enhancement weight W boost and the suppression weight W suppress The enhanced feature map X can be obtained by weighted fusion with the feature map X B and suppression feature map X S as follows:

[0067]

[0068]

[0069] Among them, σ controls the degree of enhancement, Represents element-wise dot product.

[0070] Enhanced feature map X BIt will be directly used as the output of DFSM to prepare for subsequent feature fusion; suppress feature map X S It is used as input to the next stage of the network, so that the network can learn the potential enhanced features of other regions. By inserting the DFSM module into different stages of the network, enhanced feature maps of different scales can be obtained.

[0071] In order to enhance the model's ability to discover artifact information and cope with the effects of large loss of artifact details caused by image post-processing, this embodiment introduces a noise stream branch for auxiliary decision-making, the core of which is the noise perception submodule; the input of this module is the noise information map and phase spectrum reconstruction map of the face image, which are successively fused and convolved with RGB enhanced features of different scales to obtain dual-stream complementary features of different scales, in preparation for comprehensive prediction.

[0072] Assume that there is a face image I, and convert it into a grayscale image I through an image processing algorithm gray ∈R H×W×1 Afterwards, the phase spectrum reconstruction diagram and noise information diagram are obtained as follows:

[0073] First, the grayscale image I gray Apply discrete Fourier transform to convert it to the frequency domain, take its phase to obtain the phase spectrum information of the input image, and then apply inverse discrete Fourier transform to obtain the phase spectrum reconstruction image I phase ∈R H×W×1 .

[0074] Noise map based on adaptive SRM filtering

[0075] As an effective noise extraction method, the manually designed SRM filter is often subject to fixed kernel parameters and is insensitive to post-processed image artifacts. Therefore, this embodiment introduces a constrained convolution layer to adaptively update the kernel parameters of the SRM through a network learning method to improve the robustness of the model.

[0076] The constrained convolution kernel rule is:

[0077]

[0078] Among them, C k represents the convolution kernel of the kth channel, C k (0,0) is the coordinate of its center, C k Update the grayscale image I gray As the input constrained convolution layer, the SRM noise information of the image can be obtained. srm ∈R H×W×3 .

[0079] like Figure 2 As shown in the noise flow branch in the figure, the phase spectrum is reconstructed into Figure I phaseand Noise Information Graph I srm Spliced ​​along the channel dimension into a 4-channel feature map I noise As the input of the noise stream, the convolution blocks of different layers obtain noise feature maps of different scales. In order to enhance the model's ability to detect artifacts, the output of each convolution block adds a fusion of noise features and corresponding scale RGB enhancement features, and strengthens the feature expression of each specific part by mining complementary information, and then inputs the next convolution block for higher-level feature learning.

[0080] Assume that the feature obtained by three convolutions of the noise flow is The three enhanced features obtained by the progressive attention module are The two features are straightened into two-dimensional vectors along the spatial dimension to obtain and Inspired by self-attention, the above two-stream complementary fusion method is:

[0081]

[0082]

[0083] Among them, ρ is a hyperparameter that controls the degree of dual-stream fusion, att BN represent for The complementary weights, att NB represent for The complementary weight of is calculated as:

[0084]

[0085]

[0086] Formula (7) gives Will be input to the next convolutional block to learn higher-level noise features Formula (6) gives It is retained as the input feature of the comprehensive prediction module, which integrates the complementary features of the two streams, enhances the expressiveness of the model features, and provides more detailed information for subsequent local similarity calculation and acquisition of global artifact information.

[0087] In view of the excellent detection effect of consistency learning in deep fake detection, this embodiment introduces it into the final prediction module. However, considering that this method is limited to local difference information, it is easy to ignore some global information, such as the color of artifacts in different facial regions and the contextual connection between each artifact, etc., it is combined with the global information to make a decision.

[0088] In order to further enhance the feature expression capability of the input comprehensive prediction module, this embodiment uses the different scale enhanced features obtained by formula (6) By performing pairwise fusion, the enhanced features of each scale aggregate the discriminative information from other scales, and fully combine the contextual information of the image. Here, the fusion method introduced in this embodiment is used to obtain a new multi-scale complementary enhanced feature as follows:

[0089]

[0090] Among them, μ is used to control the degree of fusion.

[0091] att i,j represent for The complementary weights are calculated as follows:

[0092]

[0093] in, and Yes and The two-dimensional vectors are pulled along the spatial dimension. In this fusion strategy, each scale of the enhanced feature map will be fused with other scales of feature maps to achieve comprehensive complementary enhancement, reduce semantic information loss, and provide more discriminative high-level features for classifier decision-making.

[0094] Here, local similarity calculation is introduced to obtain local consistency information. The feature uses a more representative multi-scale enhanced feature The bilinear interpolation algorithm is used to make the feature maps of different scales consistent in size, and then they are spliced ​​into a feature map Z∈R along the channel dimension. H×W×C . Divide Z into M×M blocks along the spatial dimension and straighten them into one-dimensional vectors. Assume that the one-dimensional vectors obtained from the i-th and j-th blocks are p i , Then the similarity between blocks is constructed based on the dot product i,j as follows:

[0095]

[0096] Among them, δ is the Sigmoid activation function, sim i,j The value of is between 0 and 1. The closer it is to 1, the higher the similarity between the two blocks, and vice versa. By traversing and calculating the similarity between all blocks and other blocks, a multi-scale local similarity feature F can be constructed. sim .

[0097] Due to the enhanced feature map output by the RGB branch After complementary fusion of different domains and scales, the enhanced feature map of each scale not only highlights the artifact area of ​​concern at the current stage, but also supplements the detailed information from other scales and high-frequency noise. The bilinear interpolation algorithm makes the feature map size consistent, and then it is spliced ​​along the channel and learned through a certain number of convolutional layers to obtain the feature At this time It can be seen as a global feature that covers different artifacts of attention at different scales.

[0098] In order to improve the decision performance of the classifier, this embodiment combines it with the multi-scale local similarity feature F sim After being straightened into a one-dimensional vector, the final prediction probability is obtained through a fully connected layer and a Sigmoid activation function. The cost function uses cross entropy loss as follows:

[0099]

[0100] Here, if the input image is a fake face, y is set to 1, otherwise it is set to 0. During the training phase, backpropagating the loss forces the network to learn the difference between real and fake faces.

[0101] This embodiment introduces a progressive spatial attention submodule, which can not only focus on the features of the image's significant regions, but also force the model to mine other potential features through a weight suppression mechanism, solving the problem of overfitting caused by over-focusing on a single region in the convolutional neural network; in the extraction of high-frequency noise features, an adaptive constrained convolution kernel is used to improve the manually designed SRM filter, solving the problem of the filter's insensitivity to image post-processing due to fixed weights. At the same time, combined with the phase spectrum that is more sensitive to position information, the noise flow's perception of artifact information is improved; a method of joint decision-making of local consistency information and global information is adopted to solve the problem that the current method is only based on local features, resulting in incomplete information expression, and improve the model's discrimination ability.

[0102] Embodiment 2

[0103] Embodiment 2 of the present disclosure introduces a deep fake image detection system integrating noise perception.

[0104] like Figure 4 A deep fake image detection system integrating noise perception is shown, comprising:

[0105] An acquisition module, configured to acquire an image to be detected;

[0106] A progressive spatial attention module is configured to extract features of the RGB stream according to the acquired image to be detected and the dual attention calculation to obtain multi-scale enhanced features of the image;

[0107] The noise perception module is configured to extract noise domain features from the acquired image to be detected, obtain image noise flow features according to the obtained image multi-scale enhancement features and the extracted noise domain features; perform dual-stream feature fusion on the obtained image noise flow features and the extracted RGB flow features to obtain image dual-stream fusion enhancement features;

[0108] The detection module is configured to complete the image forgery detection based on the obtained image dual-stream fusion enhancement features and local consistency calculation.

[0109] The detailed steps are the same as the deep fake image detection method integrating noise perception provided in Example 1, and will not be repeated here.

[0110] Embodiment 3

[0111] A third embodiment of the present disclosure provides a computer-readable storage medium.

[0112] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the deep fake image detection method integrating noise perception as described in the first embodiment of the present disclosure.

[0113] The detailed steps are the same as the deep fake image detection method integrating noise perception provided in Example 1, and will not be repeated here.

[0114] Embodiment 4

[0115] A fourth embodiment of the present disclosure provides an electronic device.

[0116] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the deep fake image detection method integrating noise perception as described in the first embodiment of the present disclosure are implemented.

[0117] The detailed steps are the same as the deep fake image detection method integrating noise perception provided in Example 1, and will not be repeated here.

[0118] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A deep fake image detection method integrating noise perception, It is characterized in that include: Acquire the image to be detected; According to the acquired image to be detected and the dual attention calculation, the features of the RGB stream are extracted to obtain the multi-scale enhanced features of the image; Extracting noise domain features from the acquired image to be detected, and obtaining image noise flow features based on the obtained image multi-scale enhancement features and the extracted noise domain features; The noise stream features of the obtained image and the features of the extracted RGB stream are fused into two stream features to obtain the image two stream fusion enhanced features; According to the obtained image dual-stream fusion enhancement features and local consistency calculation, the image forgery detection is completed; In the process of extracting the features of the RGB stream, a progressive spatial attention module is used to adaptively learn the enhancement weights and suppression weights of the image spatial domain features through dual attention calculation, and the original feature map is weightedly fused with the enhancement weights and suppression weights to obtain an enhanced feature map and a suppressed feature map; the enhanced feature map is used for dual-stream feature fusion, and the suppressed feature map is input again into the dual attention calculation of the next stage after convolution to mine the potential features of the image area outside the attention area; The dual-stream fusion enhancement features are complementarily fused to obtain multi-scale enhancement features, and the multi-scale enhancement features are input into the convolution block to obtain global features containing high-level semantic information. At the same time, the multi-scale enhancement features are divided into multiple local blocks, and the local consistency information is obtained by calculating the similarity between the local blocks. The obtained global features and local consistency information are combined to complete forgery detection.

2. A method for detecting deep fake images by integrating noise perception as claimed in claim 1, It is characterized in that High-frequency noise information is obtained through adaptive convolution kernels. The obtained high-frequency noise information and phase spectrum are used together as the input of the noise stream to realize the extraction of noise domain features, and are complementary fused with the multi-scale enhanced features in the spatial domain obtained from the RGB stream to obtain the image dual-stream fusion enhanced features.

3. A method for detecting deep fake images by integrating noise perception as claimed in claim 1, It is characterized in that The bilinear interpolation algorithm is used to transform the multi-scale enhanced feature maps into the same scale and splice them into feature maps along the channel dimension to obtain local consistency information.

4. A deep fake image detection system integrating noise perception, It is characterized in that include: An acquisition module, configured to acquire an image to be detected; A progressive spatial attention module is configured to extract features of the RGB stream according to the acquired image to be detected and the dual attention calculation to obtain multi-scale enhanced features of the image; A noise perception module is configured to extract noise domain features from the acquired image to be detected, and obtain image noise flow features based on the obtained image multi-scale enhancement features and the extracted noise domain features; The noise stream features of the obtained image and the features of the extracted RGB stream are fused into two stream features to obtain the image two stream fusion enhanced features; A detection module, which is configured to complete the forgery detection of the image according to the obtained image dual-stream fusion enhancement features and local consistency calculation; In the process of extracting the features of the RGB stream, a progressive spatial attention module is used to adaptively learn the enhancement weights and suppression weights of the image spatial domain features through dual attention calculation, and the original feature map is weightedly fused with the enhancement weights and suppression weights to obtain an enhanced feature map and a suppressed feature map; the enhanced feature map is used for dual-stream feature fusion, and the suppressed feature map is input again into the dual attention calculation of the next stage after convolution to mine the potential features of the image area outside the attention area; The dual-stream fusion enhancement features are complementarily fused to obtain multi-scale enhancement features, and the multi-scale enhancement features are input into the convolution block to obtain global features containing high-level semantic information. At the same time, the multi-scale enhancement features are divided into multiple local blocks, and the local consistency information is obtained by calculating the similarity between the local blocks. The obtained global features and local consistency information are combined to complete forgery detection.

5. A computer-readable storage medium having a program stored thereon, It is characterized in that When the program is executed by a processor, the steps in the noise-aware deep fake image detection method as described in any one of claims 1 to 3 are implemented.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, It is characterized in that When the processor executes the program, the steps in the noise-aware deep fake image detection method as described in any one of claims 1 to 3 are implemented.

Images