A Plagiarism Detection Method for Image Design Works Based on Adversarial Networks

Abstract

This invention relates to a method for detecting plagiarism in image design works based on adversarial networks. Addressing the problem of a lack of existing samples of altered image design works, which makes training traditional deep neural networks difficult, this invention employs a novel logical strategy. The generator network generates a model of altered mask images based on the original plagiarized image. This model is then combined with manually labeled altered mask images. A discriminant network determines which mask image is manually labeled. Based on the accuracy of the judgment result, corresponding feedback training operations are performed to continuously improve the accuracy of both networks. Through continuous adversarial iteration, the accuracy of the two networks reaches a final equilibrium. The resulting generator network serves as the image detection model, possessing superior plagiarism detection performance. Applying this model, the altered mask images of the image to be judged can be obtained, efficiently achieving the detection of alterations to image design works.

Description

Technical Field

[0001] This invention relates to a method for detecting plagiarism in image design works based on adversarial networks, belonging to the field of image work tampering detection technology. Background Technology

[0002] Plagiarism detection and identification in design has always been an unavoidable issue in the design and academic communities. Plagiarism identification involves detecting similarities between works and determining the content of those similarities. From the perspective of plagiarism forms, there is the act of copying others' works verbatim or almost verbatim, as well as the act of appropriating others' copyrighted original elements after modification. The intellectual property rights of graphic design works are mainly reflected in the use of color, composition, and the artistry of their expression. However, current methods for defining plagiarism in design works primarily rely on manual identification, which is time-consuming, labor-intensive, highly subjective, and prone to disputes.

[0003] Digital image forensics, as an important means of analyzing, identifying, and confirming digital image tampering and forgery, has developed rapidly in recent years. Low-level plagiarism can be addressed using digital image stitching techniques. The design community has also begun to explore plagiarism detection in design works through low-level image feature similarity calculations. Although digital image forensics has achieved significant results in many fields, few scholars have addressed plagiarism detection in graphic design works. This is because advanced plagiarism rarely involves direct copying from the original image, or the copied image is often artistically processed to conceal traces of plagiarism. Moreover, design plagiarism often heavily borrows from others' design ideas (such as layout, form, and creative concepts), and frequently uses methods such as hand-drawn imitation. This makes it difficult to accurately describe the similarity between plagiarized works, and therefore difficult to generate quantitative indicators. Therefore, determining advanced plagiarism in design works requires a higher level of understanding, at which point most image forensics methods will struggle to identify this type of plagiarism.

[0004] Scale-Invariant Feature Transform (SIFT) exhibits good robustness against attacks such as rotation and brightness changes in the tampered region, and maintains a certain degree of stability against viewpoint changes, affine transformations, and noise. Amerini et al. first proposed a copy-paste forensics method based on SIFT, and subsequently, numerous forensics schemes based on image feature points were proposed, among which SURF features, Markov features, and DAISY features were relatively successful. In recent years, many scholars have proposed using deep neural network models to solve the problem of image clone / stitching detection. However, deep neural networks are generally better at learning image content features than tampering features. To overcome this problem, Bayar and Stamm proposed a new form of convolutional layer specifically designed to suppress image content and adaptively learn tampering features. Filtering, as a commonly used denoising and image enhancement tool, has been widely applied in the field of blur and noise reduction. Resampling is another important feature of tampered images. Bunk et al. proposed an image forensics and localization method based on a combination of resampling features and deep learning. They calculated the Radon transform of resampling features on overlapping image patches and then applied a Long Short-Term Memory (LSTM) network for classification and localization. Liu et al. proposed a segmented keypoint distribution strategy to generate uniformly distributed keypoints, and then used a CNN to perform copy-paste evidence collection on suspicious images. Furthermore, existing technologies have made some targeted designs for CNNs, such as proposing to apply 30 high-pass filters from the Spatial Rich Model (SRM) to initialize the first layer of the network, effectively suppressing the influence of complex image content and accelerating network convergence. Moreover, existing technologies have designed different deep network structures, including multi-scale feature analysis and hierarchical feature matching, to better adapt to different scenarios. However, although CNN features are more representative than traditional features, they still cannot avoid problems such as feature homogeneity and poor semantic information relevance. In supervised learning, the training set requires a large amount of manually labeled data, a process that is costly and inefficient. Therefore, existing technologies have also proposed using Generative Adversarial Networks (GANs) for high-level semantic understanding of images, achieving excellent image similarity calculation results. However, existing GAN-based image similarity calculation methods are mostly aimed at image retrieval, object tracking, and image generation. For image design works, due to the limitation of limited samples, existing technologies cannot yet detect plagiarism in image design works. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for detecting plagiarism of image design works based on adversarial networks. It adopts a novel design logic strategy and applies the concept of adversarial networks to efficiently detect tampering of image design works.

[0006] To solve the above-mentioned technical problems, this invention adopts the following technical solution: This invention designs a method for detecting plagiarism in image design works based on adversarial networks, including an image detection model construction method, and applying the image detection model to detect specified image features of the image to be judged, thereby obtaining the tampered mask image corresponding to the image to be judged; the image detection model construction method includes the following steps:

[0007] Step A. For a preset number of training sample images containing tampered image regions, take the specified image features in the training sample images as input, and take the different regions corresponding to different image feature data with differences greater than the preset corresponding image feature thresholds in the training sample images as outputs, train a preset deep neural network to obtain a generator network, and then proceed to step B; wherein, the specified image features include image noise level features or image texture features.

[0008] Step B. For each training sample image, apply the generative network to obtain the tampered mask image corresponding to the training sample image, and then obtain the tampered mask image corresponding to each training sample image, and then proceed to step C;

[0009] Step C. For each training sample image, form a training sample image group by combining the training sample image with the model-modified mask image from the generative network, and also form a training sample image group by combining the training sample image with the corresponding manually modified mask image of the pre-defined manually marked modified image area; thereby obtaining each training sample image group, and then proceed to step D.

[0010] Step D. For each training sample image group, take the training sample image group as input and determine whether the tampered mask image in the training sample image group is a manually tampered mask image as output. Train a preset convolutional neural network to obtain a discrimination network, and then proceed to step E.

[0011] Step E. For a preset number of verification sample images that each contain a tampered image area, obtain each verification sample image group according to the methods in steps B to C, and then proceed to step F;

[0012] Step F. Apply the discriminant network to discriminate each group of verification sample images in turn, obtain the discrimination result corresponding to each verification sample image, and obtain the accuracy of the discrimination results corresponding to all groups of verification sample images, and then proceed to step G;

[0013] Step G. If the accuracy rate is greater than or equal to the preset accuracy rate threshold, proceed to step H; if the accuracy rate is less than the preset accuracy rate threshold, proceed to step J.

[0014] Step H. Following the method in Step A, continue training the generator network, update the trained generator network, and apply the generator network to obtain the accuracy of the discrimination results corresponding to all validation sample images for all validation sample images, following the methods in Steps E to F. Then proceed to Step I.

[0015] Step I. Determine whether the difference between the current accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the generator network is stabilizing, and proceed to step L; otherwise, return to step H.

[0016] Step J. Following the method in step D, continue training the discriminant network, update the trained discriminant network, and apply the discriminant network to obtain the accuracy of the discrimination results corresponding to all validation sample images, following the method in step F. Then proceed to step K.

[0017] Step K. Determine whether the difference between the current accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the discrimination network is trending towards stability, and proceed to step L; otherwise, return to step J.

[0018] Step L. If only the generator network tends to stabilize, proceed to step J; if only the discriminator network tends to stabilize, proceed to step H; if both the generator network and the discriminator network tend to stabilize, then the generator network is an image detection model that takes the specified image features in the image as input and the corresponding tampered mask image as output.

[0019] As a preferred technical solution of the present invention: In step A, for a preset number of training sample images that each contain tampered image areas, the image noise level features in the training sample images are used as input, and the different regions corresponding to different image noise levels with a difference greater than a preset noise threshold in the training sample images are respectively marked as output. The preset deep neural network is trained to obtain the generator network, and then the process proceeds to step B.

[0020] As a preferred technical solution of the present invention: In step A, for a preset number of training sample images that each contain tampered image areas, the image texture features in the training sample images are used as input, and the different regions corresponding to different image texture data with a difference greater than a preset texture threshold in the training sample images are respectively marked as output. The preset deep neural network is trained to obtain the generator network, and then the process proceeds to step B.

[0021] As a preferred technical solution of the present invention: the preset deep neural network in step A is a convolutional neural network, which includes a fully connected layer, a classification layer, and at least one deep feature learning module.

[0022] Each deep feature learning module includes a convolutional layer and a noisy linear rectified layer. The input of the convolutional layer constitutes the input of the deep feature learning module, the output of the convolutional layer is connected to the input of the noisy linear rectified layer, and the output of the noisy linear rectified layer constitutes the output of the deep feature learning module.

[0023] Each deep feature learning module is connected in series. The input of the first-level deep feature learning module forms the input of the convolutional neural network. The output of the last-level deep feature learning module is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the input of the classification layer. The output of the classification layer forms the output of the convolutional neural network.

[0024] As a preferred technical solution of the present invention: the number of deep feature learning modules in the convolutional neural network in step A is four, and the convolutional layers in each deep feature learning module apply 64 convolutional kernels of size 3*3.

[0025] As a preferred technical solution of the present invention: in step A, the noise linear rectified layer in each deep feature learning module of the convolutional neural network is a noise linear rectified function including Gaussian noise. For the input value of the neuron, the noise linear rectified function adds random uncertainty that satisfies the normal distribution.

[0026] As a preferred embodiment of the present invention: the fully connected layer in the convolutional neural network in step A is a 1024 fully connected network.

[0027] As a preferred technical solution of the present invention: the preset convolutional neural network in step D includes, in sequence from the input end to the output end of the convolutional neural network, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first pooling layer, a fourth convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer, and an output layer.

[0028] As a preferred technical solution of the present invention: in step D, each convolutional layer of the convolutional neural network uses 32 convolutional kernels of size 5*5, the pooling kernels in each pooling layer are all 2*2 in size, and the fully connected layer is a 1024 fully connected network.

[0029] As a preferred embodiment of the present invention, the preset accuracy threshold in step G is 45% to 65%.

[0030] The image design work plagiarism detection method based on adversarial networks described in this invention has the following technical advantages compared with existing technologies:

[0031] This invention presents an image design plagiarism detection method based on adversarial networks. Addressing the problem of insufficient existing samples of altered image design works, which hinders the training of traditional deep neural networks, this method employs a novel logical strategy. The generator network generates a model of altered mask images based on the original plagiarized image. This model is then combined with manually labeled altered mask images. A discriminant network determines which mask image is manually labeled. Based on the accuracy of the judgment result, representing the accuracy of the generator and discriminant networks, corresponding feedback training operations are performed to continuously improve the accuracy of both networks. Through continuous adversarial iteration, the accuracy of the two networks reaches a final equilibrium. The resulting generator network becomes the image detection model, possessing superior plagiarism detection performance. Applying this model, the altered mask images of the image to be judged can be obtained, efficiently achieving the detection of altered image design works. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the architecture of the image design work plagiarism detection method based on adversarial networks designed in this invention;

[0033] Figure 2 This is a schematic diagram of the convolutional neural network architecture corresponding to the generator network in this invention.

[0034] Figure 3 This is a schematic diagram of the convolutional neural network architecture corresponding to the discriminant network in the design of this invention. Detailed Implementation

[0035] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0036] This invention presents a method for detecting plagiarism in image design works based on adversarial networks. The method includes an image detection model construction method, and the application of the image detection model to detect specified image features of the image to be judged, thereby obtaining the tampered mask image corresponding to the image to be judged. In practical applications, such as... Figure 1 As shown, the image detection model construction method specifically executes the following steps A to L.

[0037] Step A. For a preset number of training sample images I each containing tampered image regions, take the specified image features in the training sample images I as input, and take the different regions corresponding to different image feature data with differences greater than the preset corresponding image feature thresholds in the training sample images I as outputs, train a preset deep neural network to obtain a generator network G, and then proceed to step B; wherein, the specified image features include image noise level features or image texture features.

[0038] For the samples used in step A to train the preset deep neural network, two specific schemes can be designed and applied in practical applications. First, generally speaking, in a real image, the noise levels contained in each region are roughly the same. However, for stitched images, the noise levels in the stitched regions and the real regions are usually different. The deep neural network can learn this noise difference. Therefore, this invention proposes to use a deep noise inconsistency learning network as the generator network G. In practical applications, for a preset number of training sample images I that each contain tampered image regions, the noise level features in the training sample images I are used as input, and the labels of different regions corresponding to different noise levels with a difference greater than a preset noise threshold in the training sample images I are used as output. The preset deep neural network is then trained to obtain the generator network G.

[0039] Secondly, for a preset number of training sample images I that each contain tampered image regions, the image texture features in the training sample images I are used as input, and the different regions corresponding to different image texture data with a difference greater than a preset texture threshold in the training sample images I are used as output. The preset deep neural network is trained to obtain the generator network G.

[0040] In practical applications, the preset deep neural network in step A is a convolutional neural network, and as follows: Figure 2 As shown, the specific design of the convolutional neural network includes a fully connected layer, a classification layer, and at least one deep feature learning module.

[0041] Each deep feature learning module includes convolutional layers and noisy linear rectified layers. The input of the convolutional layer forms the input of the deep feature learning module, and the output of the convolutional layer connects to the input of the noisy linear rectified layer. The output of the noisy linear rectified layer forms the output of the deep feature learning module. In practical applications, four deep feature learning modules can be designed. The convolutional layers in each deep feature learning module use 64 convolutional kernels of size 3*3. For the noisy linear rectified layers in each deep feature learning module, a noisy linear rectified function (NoisyReLU) including Gaussian noise is designed. For the input value of the neuron, the noisy linear rectified function (NoisyReLU) adds random uncertainty that satisfies a normal distribution, i.e., f(x) = max(0,x+Y), where the random variable Y ~ N(0,σ(x)).

[0042] Each deep feature learning module is connected in series. The input of the first-level deep feature learning module forms the input of the convolutional neural network. The output of the last-level deep feature learning module is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the input of the classification layer. The output of the classification layer forms the output of the convolutional neural network. In practical applications, the fully connected layer is a 1024-layer fully connected network.

[0043] Step B. For each training sample image I, apply the generative network G to obtain the tampered mask image I corresponding to the training sample image I. g This allows us to obtain the tampered mask image I corresponding to each training sample image. g Then proceed to step C.

[0044] Step C. For each training sample image I, compare the training sample image I with the mask image I modified by the model from the generative network G. g The training sample image group P is formed. g (I,I g Simultaneously, the training sample image I and its corresponding pre-defined artificially labeled and tampered image area artificially tampered mask image I are used. m The training sample image group P(I,I) is formed. m Then, obtain each training sample image group, and then proceed to step D.

[0045] Step D. For each training sample image group, take the training sample image group as input and determine whether the tampered mask image in the training sample image group is a manually tampered mask image as output. Train a preset convolutional neural network to obtain the discrimination network D, and then proceed to step E.

[0046] The preset convolutional neural network involved in step D above, in practical applications, such as... Figure 3 As shown, the specific design, from the input end to the output end of the convolutional neural network, includes the following layers in sequence: first convolutional layer, second convolutional layer, third convolutional layer, first pooling layer, fourth convolutional layer, second pooling layer, first fully connected layer, second fully connected layer, and output layer. Each convolutional layer uses 32 convolutional kernels of size 5*5, and the pooling kernels in each pooling layer are all 2*2. The fully connected layer is a 1024 fully connected network.

[0047] Step E. For a preset number of verification sample images that each contain a tampered image area, obtain each verification sample image group according to the methods in steps B to C, and then proceed to step F.

[0048] Step F. Apply the discriminant network D to discriminate each set of verification sample images in turn, obtain the discrimination result corresponding to each verification sample image, and obtain the accuracy of the discrimination result corresponding to all sets of verification sample images, and then proceed to step G.

[0049] Step G. If the accuracy rate is greater than or equal to the preset accuracy rate threshold, proceed to step H; if the accuracy rate is less than the preset accuracy rate threshold, proceed to step J. The preset accuracy rate threshold ranges from 45% to 65%, and in practical applications, 60% is often chosen.

[0050] Step H. Following the method in Step A, continue training the generator network G, update and obtain the trained generator network G, and apply the generator network G to obtain the accuracy of the discrimination results corresponding to all validation sample images for all validation sample images, following the methods in Steps E to F, and then proceed to Step I.

[0051] Step I. Determine whether the difference between the accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the generator network G is stabilizing, and proceed to step L; otherwise, return to step H.

[0052] Step J. Following the method in step D, continue training the discriminant network D, update and obtain the trained discriminant network D, and apply the discriminant network D to obtain the accuracy of the discrimination results corresponding to all validation sample images for all validation sample images, following the method in step F, and then proceed to step K.

[0053] Step K. Determine whether the difference between the accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the discrimination network D is approaching stability, and proceed to step L; otherwise, return to step J.

[0054] Step L. If only the generator network G tends to be stable, proceed to step J; if only the discriminator network D tends to be stable, proceed to step H; if both the generator network G and the discriminator network D tend to be stable, then the generator network D is an image detection model that takes the specified image features in the image as input and the tampered mask image corresponding to the image as output.

[0055] In the adversarial network training process between the generative network G and the discriminative network D designed above, it is assumed that there are n images in the training image library, as well as their tampered mask images, denoted as P. n (I n ,I mn Its loss function is:

[0056]

[0057] Where z is the random noise input to the generation model; D(x) is the probability that the discriminator model will identify the input sample as a real sample; G(z) is the sample generated by the generation model after receiving the random noise; Pr(x) is the distribution of the real data, and Pz(z) is the distribution of the generated data.

[0058] When training a generative network, the network model aims to maximize this error and attempts to minimize it for use with a discriminative network, i.e.:

[0059] min G max D E(G,D)

[0060] After obtaining the image detection model, the image detection model is applied to detect the specified image features of the image to be judged, and the tampered mask image corresponding to the image to be judged is obtained.

[0061] The plagiarism detection method for image design works based on adversarial networks designed in this patent can be applied in practice, such as as follows.

[0062] 1. Manually mark the plagiarized areas of multiple plagiarized image works. The marking format is a binary image (called a manually manipulated mask image). Due to the high cost of manual annotation, the number of such initial training samples is approximately 50-100.

[0063] 2. First, use 50% of the images I1 and their mask images I from the initial training sample set. m1 Train the generator network G.

[0064] 3. Then, the remaining 50% of the images I2 from the initial training sample set are input into the generator network G to obtain the model-modified mask image I. g2 .

[0065] 4. Combine the original training image I2 with the manually modified mask image I. m2 As a true sample (1 label); the original training image I2+ generated by the network is used to modify the mask image I. g2 These samples are used as fake samples (0-label). The application uses these samples to train the discriminant network.

[0066] 5. Continue iterating until equilibrium is reached. At this point, the accuracy of the discrimination network D reaches the expected level, such as 90%, and tends to stabilize. The resulting generator network G can then be used to detect new plagiarized or tampered image works.

[0067] 6. Input a suspicious image into the generator network G, and it will output the tampered mask image of that image.

[0068] The aforementioned technical solution, based on adversarial networks, addresses the problem of insufficient existing samples of altered image designs, which hinders the training of traditional deep neural networks. It employs a novel logical strategy: the generator network generates a model of altered mask images based on the original plagiarized image, and then combines this with manually labeled altered mask images. A discriminant network determines which mask image is manually labeled. Based on the accuracy of the judgment result, representing the accuracy of the generator and discriminant networks, corresponding feedback training operations are performed to continuously improve the accuracy of both networks. Through continuous adversarial iteration, the accuracy of the two networks reaches a final equilibrium. The resulting generator network becomes the image detection model, possessing superior plagiarism detection performance. Applying this model, the altered mask images of the image to be judged can be obtained, efficiently achieving the detection of altered image designs.

Claims

1. A method for detecting plagiarism in image design works based on adversarial networks, characterized in that: This includes methods for constructing image detection models, and applying these models to detect specified image features in an image to be judged, thereby obtaining the tampered mask image corresponding to the image to be judged. The image detection model construction method includes the following steps: Step A. For a preset number of training sample images containing tampered image regions, take the specified image features in the training sample images as input, and take the labels of different regions corresponding to different image feature data with differences greater than the preset corresponding image feature thresholds in the training sample images as outputs. Train a preset deep neural network to obtain a generator network, and then proceed to step B; wherein, the specified image features include image noise level features or image texture features. The preset deep neural network in step A above is a convolutional neural network. The convolutional neural network includes a 1024 fully connected network, a classification layer, and four deep feature learning modules. The convolutional layers in each deep feature learning module apply 64 convolutional kernels of size 3*3. Each deep feature learning module includes a convolutional layer and a noisy linear rectified layer. The input of the convolutional layer constitutes the input of the deep feature learning module, the output of the convolutional layer is connected to the input of the noisy linear rectified layer, and the output of the noisy linear rectified layer constitutes the output of the deep feature learning module. Each deep feature learning module is connected in series. The input of the first-level deep feature learning module forms the input of the convolutional neural network. The output of the last-level deep feature learning module is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the input of the classification layer. The output of the classification layer forms the output of the convolutional neural network. In step A, the noisy linear rectified layers in each deep feature learning module of the convolutional neural network are all noisy linear rectified functions that include Gaussian noise. For the input value of the neuron, random uncertainty that satisfies the normal distribution is added to the noisy linear rectified function. Step B. For each training sample image, apply the generative network to obtain the tampered mask image corresponding to the training sample image, and then obtain the tampered mask image corresponding to each training sample image, and then proceed to step C; Step C. For each training sample image... to train sample images The model from the generative network G tampered with the mask image. , constitute the training sample image group Simultaneously using training sample images The corresponding preset manually marked image area is a manually modified mask image. , constitute the training sample image group Then, obtain each training sample image group, and proceed to step D; Step D. For each training sample image group, take the training sample image group as input and determine whether the tampered mask image in the training sample image group is a manually tampered mask image as output. Train a preset convolutional neural network to obtain a discrimination network, and then proceed to step E. The preset convolutional neural network in step D above, from the input end to the output end, includes, in sequence, a first convolutional layer, a second convolutional layer, a third convolutional layer, a first pooling layer, a fourth convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer, and an output layer; wherein, each convolutional layer of the convolutional neural network uses 32 convolutional kernels of size 5*5, each pooling layer has a pooling kernel of size 2*2, and the fully connected layer is a 1024 fully connected network; Step E. For a preset number of verification sample images that each contain a tampered image area, obtain each verification sample image group according to the methods in steps B to C, and then proceed to step F; Step F. Apply the discriminant network to discriminate each group of verification sample images in turn, obtain the discrimination result corresponding to each verification sample image, and obtain the accuracy of the discrimination results corresponding to all groups of verification sample images, and then proceed to step G; Step G. If the accuracy rate is greater than or equal to the preset accuracy rate threshold, proceed to step H; if the accuracy rate is less than the preset accuracy rate threshold, proceed to step J; the preset accuracy rate threshold is 45% to 65%. Step H. Following the method in Step A, continue training the generator network, update the trained generator network, and apply the generator network to obtain the accuracy of the discrimination results corresponding to all validation sample images for all validation sample images, following the methods in Steps E to F. Then proceed to Step I. Step I. Determine whether the difference between the current accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the generator network is stabilizing, and proceed to step L; otherwise, return to step H. Step J. Following the method in step D, continue training the discriminant network, update the trained discriminant network, and apply the discriminant network to obtain the accuracy of the discrimination results corresponding to all validation sample images for all validation sample images, following the method in step F, and then proceed to step K; Step K. Determine whether the difference between the current accuracy rate and the previous accuracy rate falls within the preset fluctuation range. If yes, it indicates that the discrimination network is trending towards stability, and proceed to step L; otherwise, return to step J. Step L. If only the generator network tends to stabilize, proceed to step J; if only the discriminator network tends to stabilize, proceed to step H; if both the generator and discriminator networks tend to stabilize, then the generator network is an image detection model that takes the specified image features in the image as input and the corresponding tampered mask image as output.

2. The method for detecting plagiarism in image design works based on adversarial networks according to claim 1, characterized in that: In step A, for a preset number of training sample images that each contain tampered image regions, the noise level features in the training sample images are used as input, and the different regions corresponding to different noise levels with a difference greater than a preset noise threshold in the training sample images are marked as output. The preset deep neural network is trained to obtain the generator network, and then the process proceeds to step B.

3. The method for detecting plagiarism in image design works based on adversarial networks according to claim 1, characterized in that: In step A, for a preset number of training sample images that each contain tampered image areas, the image texture features in the training sample images are used as input, and the different regions corresponding to different image texture data with a difference greater than a preset texture threshold in the training sample images are used as output. The preset deep neural network is trained to obtain a generative network, and then the process proceeds to step B.

Images