Method and device for identifying fraud-related visual resources

Abstract

The invention provides a fraud-related visual resource identification method and device, and the method comprises the steps: converting a to-be-predicted visual resource and a visual resource in an existing fraud-related visual resource library from an RGB color space to an opposite color space, and judging whether the to-be-predicted visual resource relates to a fraud or not based on the cartoon part of the visual resource. According to the method, before visual resource retrieval, effective information in the visual resources is screened in advance by extracting cartoon parts of the visual resources, so that the efficiency and the quality of subsequent feature extraction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a method and device for identifying fraud-related visual resources. Background Art

[0002] With the rapid development of information technology, telecom network fraud crimes are new, diverse and deceptive. Among them, visual resources are one of the main sources of information. Criminals use artificial intelligence (AI) and other technical means to update and iterate images in fraud-related resources, which brings new challenges to anti-fraud identification work.

[0003] For the identification of fraud-related visual resources, existing technical solutions include: Solution 1: Text-based visual resource retrieval technology. This solution uses traditional text retrieval technology, avoiding the analysis of visual resource visualization elements. It generally queries visual resources in the form of keyword questions, and uses data crawlers to parse text information such as the uniform resource locator (URL) of images in web resources. Based on these text analysis results, the characteristics of the visual resources are inferred to determine whether the visual resources to be predicted are fraudulent.

[0004] Solution 2: Content-based visual resource retrieval technology method. This solution is based on supervised learning of machine learning. Its core is to use the visual features of visual resources to retrieve visual resources. Its general framework can be roughly divided into two steps: extracting certain features and calculating similarity. In order to avoid human subjectivity, feature extraction is generally completed automatically by computers. The computer will extract the features of the entire visual resource to be predicted and calculate the similarity with the visual resources in the fraud-related resource library to determine whether the visual resource to be predicted is fraudulent.

[0005] However, the existing solution has the following disadvantages: it is subject to the limitations of text vocabulary, is prone to ambiguity, is slow to update, and is not easy to cope with the ever-changing various fraud-related visual resources on the Internet. The features extracted by text-based retrieval technology cannot represent the content characteristics of visual resources (structure, color, texture, etc.).

[0006] The existing solution 2 has the following disadvantages: When retrieving visual resources, the computer will extract the features of the entire visual resource and calculate the similarity, which is computationally intensive and costly, resulting in low accuracy of visual resource similarity retrieval results. Traditional visual resource feature extraction methods cannot accurately characterize the color and global structure of visual resources. They cannot capture fine-grained visual differences, including but not limited to noise, spectrum, etc.

[0007] In summary, the existing solutions have problems such as slow update, high computational cost, and low recognition accuracy in the identification of fraud-related visual resources. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a method and device for identifying fraud-related visual resources.

[0009] The present invention provides a method for identifying fraud-related visual resources, the method comprising: Determine a visual resource to be predicted and a plurality of visual resources involved in fraud, and convert the visual resource to be predicted and the plurality of visual resources involved in fraud from an RGB color space to an opposing color space to obtain a converted visual resource to be predicted and a plurality of converted visual resources involved in fraud; Respectively obtaining the to-be-predicted conversion visual resources and the cartoon parts corresponding to each of the fraud-related conversion visual resources; Based on the similarity between the to-be-predicted converted visual resource and the cartoon parts corresponding to each of the fraudulent converted visual resources, it is determined whether the to-be-predicted visual resource is fraudulent.

[0010] Optionally, the obtaining of the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources includes: Based on the decorrelation vector total variation DTV-L1 model, the cartoon parts corresponding to the to-be-predicted conversion visual resources and each of the fraud-related conversion visual resources are obtained.

[0011] Optionally, the acquiring of the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources based on the decorrelation vector total variation DTV-L1 model includes: The DTV-L1 model iterates the color channel and the brightness channel of the predicted conversion visual resource and each fraud-related conversion visual resource respectively to obtain the corresponding cartoon parts, and the color channel includes two channels.

[0012] Optionally, in the DTV-L1 model, based on two color channels Information, obtain O1,2 norm; based on the brightness channel Information, obtain the O3 norm; The distances O1, 2 and O3 in the opponent color space are calculated, where x and y are the pixel coordinates of the pixel in the image, and u is a function of each pixel, representing the grayscale value of the channel.

[0013] Optionally, the step of determining the similarity between the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources is as follows: Constructing a visual resource retrieval model based on a twin neural network; The to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources are input into the visual resource retrieval model to obtain the similarity between the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources output by the visual resource retrieval model.

[0014] Optionally, the method further comprises: Obtain a training data set, wherein the training data set includes positive samples and negative samples; The visual resource retrieval model is trained based on the positive samples and the negative samples.

[0015] Optionally, the determining whether the visual resource to be predicted is fraudulent based on the similarity between the visual resource to be predicted and the cartoon parts corresponding to each of the fraudulent conversion visual resources comprises: Determine a target fraud-related visual resource based on the similarity between the to-be-predicted conversion visual resource and the cartoon part corresponding to each of the fraud-related conversion visual resources, and judge whether the to-be-predicted visual resource is fraud-related based on the target fraud-related visual resource; When the similarity reaches a similarity threshold, it is determined that the target fraud-related visual resource and the visual resource to be predicted are the same type of fraud-related visual resources.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned fraud-related visual resource identification methods.

[0017] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned fraud-related visual resource identification methods.

[0018] In the fraud-related visual resource identification method and device provided by the present invention, the visual resource to be predicted and the visual resources in the existing fraud-related visual resource library are converted from the RGB color space to the opposing color space, and whether the visual resource to be predicted is fraud-related is judged based on the cartoon part of the visual resource. Before the visual resource is retrieved, the present invention screens the effective information in the visual resource in advance by extracting the cartoon part of the visual resource, which significantly improves the efficiency and quality of subsequent feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 One of the flow charts of the method for identifying fraud-related visual resources provided in an embodiment of the present invention.

[0021] Figure 2 The second flow chart of the method for identifying fraud-related visual resources provided in an embodiment of the present invention.

[0022] Figure 3 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In actual business, the various embodiments of the present invention can be mainly used in telecommunications anti-fraud systems, and their main function is to identify fraud-related visual resources through technical means, and provide clues for the identification of fraud-related resources. Among them, the fraud-related visual resources may refer to the visual resources in the interface of fraud-related resources (websites, APPs). In the fraud-related visual resource identification scenario of telecommunications network fraud, based on the behavioral characteristics of resource forgery and tampering in fraud-related websites, the various embodiments of the present invention innovatively introduce a color image decomposition model to extract the cartoon part of the visual resource, eliminate the interference of noise and texture, and further perform visual resource retrieval on the cartoon part, thereby realizing the early screening of effective information.

[0025] The following is an introduction to the relevant knowledge. Visual resource decomposition is an important problem in visual resource processing and computer vision, where natural visual resources are viewed as a superposition of cartoon parts and texture parts. The cartoon part models the segmented continuous regions with clear boundaries and contains the global structural information of the visual resources, while the texture part models the locally patterned oscillating parts and contains some repeated patterns. The goal of visual resource decomposition is to find the cartoon part u and the texture part v so that their sum is the target visual resource g: g=u+v.

[0026] In the case of grayscale visual resources (low-level), the visual resource decomposition framework Meyer's Models is generally used: Find the space X1, X2, and the corresponding F1, F2 so that u is the cartoon part, v is the texture part, and F1(u)<<F2(u),F1(v)> >F2(v). Texture parts must be penalized by F1 instead of F2, whereas cartoon parts must be penalized by F2. F1 can be selected as the total variation of u, which makes u contain homogeneous areas and protects sharp edges.

[0027] The most widely used total variation (TV) model proposed by Rudin, Osher and Fatemin, also known as the ROF model, removes noise from visual resources while maintaining sharp edges by minimizing the TV norm and L2 norm. It has been widely used in many visual resource processing problems, such as denoising, visual resource restoration, painting, compressed sensing, etc. Vincent LeGuen proposed that the L1 norm is particularly suitable for cartoon texture decomposition because it preserves geometric features better than the L2 norm.

[0028] However, in the high-order case of color visual resources, the addition of color information makes the high-order extension of the visual resource decomposition model very important. Visual resources are mainly represented by the red, green, and blue (RGB) color model. The simplest method is to define the total variation (TV) in RGB separately channel by channel. In this method, each channel is considered independently, and the TV norm of the color visual resource is the sum of the TV norms of each channel. One disadvantage of this method is that color edges may be lost and color smearing may occur between channel edges. Compared with the RGB model, the opponent color model is closer to human perception. Different from the traditional RGB color space, the opponent color space family is a set of physiological color spaces inspired by the physiology of the human visual system. The opponent color space selects the three color features that are most effective for color visual resource segmentation as three channels, namely , and the three color features are squared and normalized.

[0029] Figure 1 A flowchart of a method for identifying fraud-related visual resources provided by an embodiment of the present invention is shown in the figure. The method includes the following steps: Step 100: determine a visual resource to be predicted and a plurality of fraud-related visual resources, and convert the visual resource to be predicted and the plurality of fraud-related visual resources from the RGB color space to an opposing color space to obtain a converted visual resource to be predicted and a plurality of converted fraud-related converted visual resources; First, obtain a visual resource to be predicted, such as an image resource in a fraudulent website interface. The embodiment of the present invention is to determine whether the image resource is fraudulent. Multiple fraudulent visual resources can be used at the same time. The fraudulent visual resources can be previously available visual resources that have been determined to belong to fraudulent images. These fraudulent visual resources can be stored in a fraudulent visual resource library. Through subsequent processing, try to find one or more fraudulent visual resources that are most similar to the visual resource to be predicted.

[0030] After obtaining the visual resources to be predicted and multiple fraud-related visual resources, the color space conversion module can be used to convert the visual resources to be predicted and the multiple fraud-related visual resources from RGB color space to opposing color space to obtain the converted converted visual resources to be predicted and the converted multiple fraud-related converted visual resources.

[0031] Step 101, respectively obtaining the to-be-predicted conversion visual resources and the cartoon parts corresponding to each of the fraud-related conversion visual resources; After obtaining the converted visual resources to be predicted and multiple converted fraud-related conversion visual resources, a visual resource decomposition model in a pre-constructed opposing color space is statistically analyzed to extract the cartoon parts corresponding to the visual resources to be predicted and the cartoon parts corresponding to each of the fraud-related conversion visual resources.

[0032] In the embodiment of the present invention, the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources may be obtained based on a decorrelated vector total variation (Decorrected Total Variation, DTV) DTV-L1 model.

[0033] Specifically, in the embodiment of the present invention, the DTV-L1 model iterates the color channel and the brightness channel of the predicted conversion visual resource and each fraud-related conversion visual resource respectively to obtain the corresponding cartoon parts, and the color channel includes two channels.

[0034] Step 102: Determine whether the visual resource to be predicted is fraudulent based on the similarity between the visual resource to be predicted and the cartoon parts corresponding to each of the fraudulent conversion visual resources.

[0035] In the embodiment of the present invention, determining the similarity between the predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources may include the following steps: Constructing a visual resource retrieval model based on a twin neural network; The to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources are input into the visual resource retrieval model to obtain the similarity between the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources output by the visual resource retrieval model.

[0036] Furthermore, it also includes a process of training a visual resource retrieval model, specifically obtaining a training data set, wherein the training data set includes positive samples and negative samples; and training the visual resource retrieval model based on the positive samples and negative samples.

[0037] Further, judging whether the visual resource to be predicted is fraudulent based on the target fraudulent visual resource includes: when the similarity reaches a similarity threshold, determining that the target fraudulent visual resource and the visual resource to be predicted are the same type of fraudulent visual resources.

[0038] In the embodiment of the present invention, a color space conversion module is used to convert the visual resources to be predicted and the visual resources in the own fraud-related visual resource library from the RGB color space to the opposing color space; a visual resource decomposition model in the opposing color space is established through DTV-L1, and the representation of the cartoon part of the visual resource in the opposing color space is extracted; further, the features of the cartoon part can be extracted through the visual resource retrieval module and the similarity matching is performed with the own cartoon fraud-related visual resource library to output the most similar fraud-related visual resource; finally, a fraud-related judgment module is used to judge whether the visual resource to be predicted is fraudulent, and the same type of fraud-related visual resources that are similar to the known fraud-related visual resources or updated and iterated are identified.

[0039] In the fraud-related visual resource identification method and device provided by the present invention, the visual resource to be predicted and the visual resources in the existing fraud-related visual resource library are converted from the RGB color space to the opposing color space, and similarity matching is performed based on the cartoon part of the visual resource to output the most similar fraud-related visual resource, and determine whether the visual resource to be predicted is fraud-related. Before the visual resource is retrieved, the present invention extracts the cartoon part of the visual resource to screen the effective information in the visual resource in advance, which significantly improves the efficiency and quality of subsequent feature extraction.

[0040] Figure 2 The second flow chart of the fraud-related visual resource identification method provided by the embodiment of the present invention is as follows: Figure 2 As shown: Color space conversion module: Convert visual resources from RGB color space to opposing color space, and then perform subsequent modeling and recognition of visual resources.

[0041] The family of opponent color spaces is a set of physiological color spaces inspired by the physiology of the human visual system, assuming is a visual resource in RGB space, for The corresponding visual resources in the opponent color space. Convert the RGB color model to the opponent color model through matrix transformation:

[0042] In order to effectively separate the color and brightness information, the embodiment of the present invention selects the three most effective color features for color vision resource segmentation as the three channels of the opposing color space according to the regional segmentation experiment of the color vision resource of the system, which are respectively The square sum of the coefficients of the three color features is normalized to obtain the conversion matrix of the orthogonal color vision resource , the first, second, and third rows represent the three normalized color feature weight coefficients arranged in sequence :

[0043] In the opposing color space, are color channels, they represent color information. It is an achromatic channel, which represents lightness (or brightness). In the opponent color space, the color channel is separated from the lightness channel, which effectively reduces the correlation between the RGB channels.

[0044] DTV-L1 model decomposition module: Based on the channel characteristics of the above opposing color space, the distance in the opposing color space is defined, and the decorrelation vector total variation (DTV) is constructed. Then, a visual resource decomposition model based on DTV is constructed to effectively separate color and brightness, retain effective color and structural information in the cartoon part, and separate texture noise. Finally, the primal-dual algorithm is solved for high-order color expansion to obtain the cartoon part of the color visual resource.

[0045] In the DTV-L1 model decomposition module: Distance definition submodule: Based on the characteristics of the three channels of the opposing color space, the two color channels are The information fusion of defines the O1,2 norm; separately, for the brightness channel Define the O3 norm:

[0046] Then define the distance in the opponent color space, assuming , are visual resources in RGB space, and the distance between them is expressed as:

[0047] In the formula, x and y are the pixel coordinates of the pixel in the image, and u is the function of each pixel, representing the grayscale value of each color channel and the grayscale value of the brightness channel.

[0048] DTV construction submodule: Based on the above definitions of distance and norm, the embodiment of the present invention constructs a decorrelation vector total variation (DTV) in the opposing color space:

[0049] in is a positive parameter that controls the weight of brightness in DTV. And:

[0050] Visual resource decomposition modeling submodule: Based on the above definition and the construction of decorrelation vector total variation (DTV), the embodiment of the present invention further establishes a color vision resource high-order decomposition model under the Meyer's Models decomposition framework.

[0051] Assumptions For original visual resources, This is the cartoon part of the model that is being solved. For the visual resource area, , Using DTV as a regular term and L1 norm as a fidelity term, we construct a cartoon and texture decomposition model of visual resources, namely the DTV-L1 model:

[0052] in, is the regularization parameter, The smaller it is, the greater the degree of regularization, the more features the cartoon part retains, and the finer the granularity of the decomposed texture part.

[0053] DTV-L1 high-order solution submodule: In order to iteratively calculate the optimal solution of the model, the visual resources are represented in discrete form, assuming that the visual resources in the opposing color space (The pixel size of the visual resource is ) is the discrete gradient operator ,in is a discrete first-order gradient operator for a single channel, and its discrete differential operators in two directions are: remember , so the discrete form of the DTV-L1 model established above can be expressed as:

[0054] In order to solve the above minimization problem, the embodiment of the present invention performs a high-order color expansion on the first-order primal-dual algorithm (the method is shown in steps 2 and 3 below), and uses fixed point iteration to alternately minimize two variables.

[0055] In summary, the DTV-L1 model decomposition module can be summarized into the following steps: Step 1: Enter your visual assets , initialize the variable :

[0056] Step 2: Use approximation operator Compute iterative updates :

[0057] in, .

[0058] Specifically, Each sub-dimension of is iteratively updated separately, that is, the gradients in different directions are iteratively updated in the c-th channel (c=1, 2, 3) of the visual resource:

[0059] For the color channel (c=1, 2) and the brightness channel (c=3), the color channel fusion factor is used respectively. Luminance channel factor And perform gradient calculation:

[0060] Step 3: Use approximation operator Compute iterative visual resources :

[0061] in, .

[0062] Specifically, for visual resources The color channel (c=1, 2) and brightness channel (c=3) are iteratively updated respectively:

[0063] in,

[0064] as well as, .

[0065] Step 4: Calculate iterations : .

[0066] Step 5: Set the maximum number of steps , repeat steps 3 to 5 until the maximum number of steps is reached. The optimal solution for the cartoon part of the visual resource in the opposing color space is obtained. .

[0067] Visual resource retrieval module: The Siamese neural network is used to build a visual resource retrieval model. The same convolutional neural network is used to extract features from two input visual resources, and then the distance between their feature vectors is calculated to compare the similarity.

[0068] The visual resource retrieval model of the embodiment of the present invention needs to be trained, and the specific training steps are: Step 1: Prepare the training data set, which is divided into positive samples and negative samples. Positive sample sampling is to randomly select two visual resource combinations from a similar visual resource library. Label Recorded as 1; negative sample sampling randomly extracts two visual resource combinations from the dissimilar visual resource library, label Recorded as 0.

[0069] Step 2: Decompose the visual resource into two cartoon parts of the visual resource Input into the same convolutional neural network to extract the feature vectors of the two visual resources, recorded as .

[0070] Step 3: Subtract the two eigenvectors and take their absolute values ​​to get the Euclidean distance between them ,Right now:

[0071] Step 4: Input to a fully connected layer and get a scalar ,Right now:

[0072] Step 5: Call the activation function Sigmoid function and output a similarity prediction value between (0, 1), that is:

[0073] Step 6: Take the difference between the sample label and the predicted value as the loss function. Here we use the Contrastive Loss function:

[0074] in, Represents the set threshold value, is the sample logarithm.

[0075] Step 7: Calculate the gradient of the loss function with respect to the fully connected layer parameters and the convolutional neural network parameters through back propagation.

[0076] Step 8: Use gradient descent to update the parameters of the fully connected layer and convolutional neural network.

[0077] Through the above steps, the visual resource retrieval model is trained to achieve a recall rate and precision rate that can be applied in practice.

[0078] This module inputs the decomposed cartoon visual resources to be predicted and the resources in the cartoon fraud-related visual resource library into the visual resource retrieval model in pairs, and outputs the fraud-related visual resources with the greatest similarity to the visual resources to be predicted and their similarity.

[0079] Fraud judgment module: By comparing the similarity of the cartoon parts of two visual resources output by the visual resource retrieval module, when the similarity reaches the set similarity threshold, they are determined to be the same type of fraud-related visual resources.

[0080] Compared with the existing visual resource retrieval model, which directly extracts features from the entire input visual resource and calculates similarity, the calculation amount is large and the accuracy is low. The fraud-related visual resource identification method provided by the embodiment of the present invention combines the visual resource decomposition model with the visual resource retrieval model, and proposes to use the visual resource decomposition framework to extract the cartoon part of the visual resource before visual resource retrieval, and screen the effective information in the visual resource in advance, which significantly improves the efficiency and quality of subsequent feature extraction.

[0081] Compared with existing recognition methods, feature extraction is generally performed in RGB color space, which fuses color and brightness information in three channels, hindering the extraction of color, structure and other features in visual resources. This proposal converts the RGB color space into an opposing color space that is closer to the perception of the naked eye, and uses O 1，2 The definition of the distance from O3 effectively separates color and brightness, which helps to obtain the color and global structure information of visual resources as important features.

[0082] The existing neural network-based visual resource retrieval model is a black box fitting, which can only learn parameters through training data sets. It is difficult to find the relationship between parameters and model output vectors, has no physical meaning, and cannot optimize parameters in a targeted manner. However, in the visual resource decomposition step, the embodiment of the present invention can adjust the regularization parameters of the DTV-L1 visual resource decomposition model according to the needs of different scenarios to achieve control of the degree of regularization of visual resources. For recognition in fine-grained scenarios (more fine-grained features need to be retained), increase the regularization parameter to identify noise, spectrum, etc.; for coarse-grained scenarios, reduce the regularization parameter to improve retrieval efficiency.

[0083] Figure 3 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 520 and the memory 530 communicate with each other through the communication bus 340. The processor 310 may call the logic instructions in the memory 330 to execute the fraud-related visual resource identification method provided in the above embodiments.

[0084] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0085] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the fraud-related visual resource identification method provided in the above-mentioned embodiments.

[0086] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program is used by a processor to execute the fraud-related visual resource identification method provided in the above-mentioned embodiments.

[0087] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0088] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying fraudulent visual resources, characterized in that: The method comprises: Determine a visual resource to be predicted and a plurality of visual resources involved in fraud, and convert the visual resource to be predicted and the plurality of visual resources involved in fraud from an RGB color space to an opposing color space to obtain a converted visual resource to be predicted and a plurality of converted visual resources involved in fraud; Respectively obtaining the to-be-predicted conversion visual resources and the cartoon parts corresponding to each of the fraud-related conversion visual resources; Based on the similarity between the to-be-predicted converted visual resource and the cartoon parts corresponding to each of the fraudulent converted visual resources, it is determined whether the to-be-predicted visual resource is fraudulent.

2. The fraud-related visual resource identification method according to claim 1 is characterized in that: The obtaining of the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources includes: Based on the decorrelation vector total variation DTV-L1 model, the cartoon parts corresponding to the to-be-predicted conversion visual resources and each of the fraud-related conversion visual resources are obtained.

3. The fraud-related visual resource identification method according to claim 2 is characterized in that: The method of obtaining the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources based on the decorrelation vector total variation DTV-L1 model includes: The DTV-L1 model iterates the color channel and the brightness channel of the predicted conversion visual resource and each fraud-related conversion visual resource respectively to obtain the corresponding cartoon parts, and the color channel includes two channels.

4. The fraud-related visual resource identification method according to claim 3 is characterized in that: The DTV-L1 model is based on two color channels Information, obtain O1,2 norm; based on the brightness channel Information, obtain the O3 norm; 5. Calculate the O1, 2 and O3 distances in the opposing color space, where x and y are the pixel coordinates of the pixel in the image, and u is a function of each pixel, representing the grayscale value of the channel.

6. The fraud-related visual resource identification method according to any one of claims 1 to 4, characterized in that: The step of determining the similarity between the predicted conversion visual resource and the cartoon parts corresponding to each of the fraud-related conversion visual resources is as follows: Constructing a visual resource retrieval model based on a twin neural network; The to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources are input into the visual resource retrieval model to obtain the similarity between the to-be-predicted conversion visual resource and the cartoon parts corresponding to each of the fraudulent conversion visual resources output by the visual resource retrieval model.

7. The fraud-related visual resource identification method according to claim 5, characterized in that: The method further comprises: Obtain a training data set, wherein the training data set includes positive samples and negative samples; The visual resource retrieval model is trained based on the positive samples and the negative samples.

8. The fraud-related visual resource identification method according to claim 5 or 6, characterized in that: The determining whether the visual resource to be predicted is fraudulent based on the similarity between the to-be-predicted converted visual resource and the cartoon parts corresponding to each of the fraudulent converted visual resources comprises: Determine a target fraud-related visual resource based on the similarity between the to-be-predicted conversion visual resource and the cartoon part corresponding to each of the fraud-related conversion visual resources, and judge whether the to-be-predicted visual resource is fraud-related based on the target fraud-related visual resource; When the similarity reaches a similarity threshold, it is determined that the target fraud-related visual resource and the visual resource to be predicted are the same type of fraud-related visual resources.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for identifying fraud-related visual resources as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying fraud-related visual resources as described in any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for identifying fraud-related visual resources as described in any one of claims 1 to 7 is implemented.

Images