Image anti-fraud method and device based on deep learning

Abstract

The invention discloses an image anti-fraud method and device based on deep learning, belongs to the technical field of computers, and particularly relates to an image anti-fraud method based on deep learning, which comprises the following steps: acquiring image data, and preprocessing the acquired image data; constructing a deep learning model; a deep learning model is adopted to perform feature extraction and learning on the preprocessed image data, and deep learning model training is performed; and inputting a to-be-detected image into the trained deep learning model, and carrying out fraud detection and classification. According to the method, the problem that a traditional image fraud detection technology depends on a manually designed feature extraction method can be avoided, the technical barrier that enough useful information is difficult to capture in the face of complex and changeable fraud means is avoided, and then the detection accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to an image anti-fraud method and device based on deep learning. Background Art

[0002] With the rapid development of the Internet and digital media, images have become an indispensable part of people's daily lives. Images are not only used for entertainment and social interaction, but also play an important role in many fields such as finance, medicine, and law. The widespread use of images also poses a series of security issues, especially image fraud. Image fraud refers to the act of tampering with or forging image content through technical means to mislead the audience or achieve some improper purpose. These fraudulent acts may include but are not limited to identity impersonation, false advertising, forged documents, etc., posing a serious threat to social order and personal property security.

[0003] In order to meet this challenge, the prior art has proposed a variety of image fraud detection technologies. The inventors have found through research that, in the prior art, these methods can identify and prevent image fraud to a certain extent, but traditional image fraud detection technologies often rely on manually designed feature extraction methods, which are difficult to capture sufficient useful information when faced with complex and changeable fraud methods, resulting in low detection accuracy. Summary of the invention

[0004] In order to at least solve the above technical problems, the present invention provides an image anti-fraud method, device, equipment and readable storage medium based on deep learning, which effectively improves the accuracy and efficiency of image fraud detection by using the powerful feature learning and generalization capabilities of the deep learning model, and has strong robustness and adaptability. Through continuous learning and optimization, it adapts to the ever-changing fraud methods and provides a more reliable guarantee for image security.

[0005] According to a first aspect of the present invention, a deep learning-based image anti-fraud method is provided, comprising: Collecting image data and preprocessing the collected image data; Build deep learning models; Use deep learning models to extract features and learn from preprocessed image data, and perform deep learning model training; The images to be tested are fed into the trained deep learning model for fraud detection and classification.

[0006] Further, The collecting of image data and preprocessing of the collected image data include: Use automated data collection tools to obtain image data from several sources; The acquired image data is normalized, resized, denoised and enhanced.

[0007] Further, The obtained image data is normalized, resized, denoised and enhanced using the formula: ,in is the original pixel value, is the normalized pixel value; Resize all images to the size required by the model input; Use a filter to remove image noise, using Ismooth=G∗I, where G is the Gaussian kernel, I is the original image, and Ismooth is the denoised image.

[0008] Further, The building of the deep learning model includes: Build an innovative deep hybrid model that combines convolutional neural network, recurrent neural network and neural network architectures, where The convolution layer uses convolution kernels of different sizes to extract features at different levels of the image, using the formula: F=W*I+b, where W is the convolution kernel, I is the input image, b is the bias term, and F is the output feature map. The pooling layer reduces the spatial dimension of the feature and the number of parameters. The maximum pooling P=max(F), where P is the feature after pooling. Recurrent Neural Networks use the formula: ,in is the hidden state at the current moment, is the current input, is the weight matrix, is the bias term, is the activation function; The self-attention mechanism is used to calculate the relationship between different regions in the image, so as to achieve the modeling of global information. The calculation formula of the self-attention mechanism is: , where Q, K, and V are query vectors, key vectors, and value vectors, respectively. is the dimension of the key vector.

[0009] Further, The deep learning model training includes: Adopt a multi-stage training strategy; Pre-train the fused convolutional neural network, recurrent neural network, and neural network architecture separately, using a large number of labeled positive and negative sample images; During pre-training, unsupervised learning methods, autoencoders or contrastive learning are used to learn the common features of images; The three parts are trained jointly, and the model parameters are continuously adjusted through the back propagation algorithm and the optimization algorithm to minimize the loss function. The loss function adopts a multi-task loss function, combining image classification loss, image reconstruction loss and adversarial loss. Among them, the image classification loss adopts the formula: cross entropy loss ,in is the true label, is the prediction probability; the image reconstruction loss adopts the formula: square error loss ,in is the original image, Generate images; use adversarial loss to train GAN, so that the images generated by the generator are more realistic, and the discriminator can distinguish between real images and generated images more accurately; The optimization algorithm uses the adaptive moment estimation (Adam) optimizer, and the formula is ,in is the current parameter, is the learning rate, is the first-order moment estimate, is the second-order moment estimate, is a preset constant.

[0010] Further, The image to be detected is input into the trained deep learning model to perform fraud detection and classification, including: Input the image to be detected into the model and calculate the fraud probability using the formula , where Z is the output of the fully connected layer, is a computational function, The probability of fraud is determined by setting different thresholds and making judgments based on the probability of fraud and specific application requirements; Multi-level classification is adopted, and several classifiers are added to the output layer of the model, corresponding to different fraud types. By training these classifiers, specific fraud types can be identified.

[0011] Further, The deep learning model also includes: Regularly use new annotated data to train the model using online learning and incremental learning techniques, so that the model can continuously learn and adapt to new fraud methods in a changing environment; Introducing model distillation technology to compress complex deep hybrid models so that they can be deployed on resource-constrained devices; A feedback mechanism is introduced to further adjust and optimize the model based on the detection results.

[0012] According to a second aspect of the present invention, an image anti-fraud device based on deep learning comprises: An acquisition module, used for acquiring image data and preprocessing the acquired image data; Building blocks for building deep learning models; A training module is used to extract features and learn from the preprocessed image data using a deep learning model, and to perform deep learning model training; The detection module is used to input the image to be detected into the trained deep learning model for fraud detection and classification.

[0013] According to a third aspect of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of any method described in the first aspect of the present invention are implemented.

[0014] According to a fourth aspect of the present invention, a computer-readable storage medium stores a program, and when the program is executed, it can implement any method as described in the first aspect of the present invention.

[0015] Beneficial effects of the present invention: The present invention combines the advantages of fusion convolutional neural network, recursive neural network and neural network architecture through a deep hybrid model, which can automatically extract and learn complex features of images more efficiently, greatly improving the efficiency of fraud detection.

[0016] The use of a multi-stage training strategy and a multi-task loss function enables the model to learn more accurate fraud image features, significantly improving the detection accuracy.

[0017] Through the introduction of online learning, incremental learning and model distillation technology, the model can quickly and effectively self-optimize and update according to new data to adapt to the ever-changing fraud methods.

[0018] The image enhancement technology in the preprocessing module and the structural design of the deep hybrid model make the model more robust to noise, quality changes and special processing in the image, and can maintain stable detection performance under a variety of complex conditions.

[0019] The multimodal feature extraction capability and multi-level classification strategy of the deep hybrid model enable the model to process and identify more types of fraudulent images with stronger generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 A flow chart of an image anti-fraud method based on deep learning provided by the present invention; Figure 2 A schematic diagram of a data preprocessing process provided by the present invention; Figure 3 A schematic diagram of a deep learning model structure provided by the present invention; Figure 4 A flow chart of a model training method provided by the present invention. DETAILED DESCRIPTION

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0022] In order to more clearly illustrate the present invention, the present invention is further described below in conjunction with preferred embodiments and accompanying drawings. Similar components in the accompanying drawings are represented by the same reference numerals. It should be understood by those skilled in the art that the content specifically described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.

[0023] In a first aspect of the present invention, a deep learning-based image anti-fraud method is provided. Figure 1 As shown, including: Step 101: collecting image data and preprocessing the collected image data; In one embodiment of the present invention, image data is collected, including normal images and known fraudulent images. Specifically, an automated data collection tool is used to obtain image data from several channels to ensure the diversity and representativeness of the data. These images are preprocessed, and the specific operations are as follows Figure 2 As shown: including image screening, resizing, normalization, denoising and GAN image enhancement. Further, for normalization, the image pixel values ​​are usually scaled to between 0 and 1 or standardized to reduce the impact of lighting and contrast changes. The formula is used: ,in is the original pixel value, is the normalized pixel value.

[0024] Resizing, specifically resizing all images to the size required by the model input, such as H×W, to ensure input consistency.

[0025] Denoising specifically involves using filters to remove image noise, such as a Gaussian filter, which can be expressed as Ismooth=G∗I, where G is the Gaussian kernel, I is the original image, and Ismooth is the denoised image.

[0026] Step 102: Build a deep learning model; In one embodiment of the present invention, an innovative deep hybrid model is constructed, integrating convolutional neural networks, recurrent neural networks and neural network architectures. The convolutional neural network is used to extract local features of the image, the recurrent neural network is used to capture the sequence information of the image, and the architecture realizes the modeling of global information through the self-attention mechanism.

[0027] The convolutional neural network part contains multiple convolutional layers, pooling layers, and activation functions. The convolutional layer uses convolution kernels of different sizes to extract features at different levels of the image, which can be expressed as F=W*I+b, where W is the convolution kernel, I is the input image, b is the bias term, and F is the output feature map. The pooling layer reduces the spatial dimension of the feature and the number of parameters, such as the maximum pooling P=max(F), where P is the feature after pooling. The activation function introduces nonlinearity, and commonly used ones include ReLU, LeakyReLU, etc.

[0028] The recurrent neural network part uses a long short-term memory network (LSTM) or a gated recurrent unit (GRU), which can process the sequence information of images, such as image fraud detection in a video frame sequence. The formula of LSTM or GRU is: ,in is the hidden state at the current moment, is the current input, is the weight matrix, is the bias term, is the activation function; In the neural network architecture, the self-attention mechanism is used to calculate the relationship between different regions in the image, thereby realizing the modeling of global information. The calculation formula of the self-attention mechanism is , where Q, K, and V are query vectors, key vectors, and value vectors, respectively. is the dimension of the key vector.

[0029] Furthermore, in another embodiment of the present invention, the deep learning model includes an input layer, a CNN part, an RNN part, a Transformer part, an activation function, a fully connected layer, and an output layer. The structure of the specific deep learning model, the arrangement and connection of each part, and the function and role of each part are as follows: Figure 3 shown.

[0030] Step 103: extracting features and learning from the preprocessed image data using a deep learning model, and performing deep learning model training; In one embodiment of the present invention, a deep learning model is used to extract and learn features of preprocessed image data.

[0031] In the training phase, a multi-stage training strategy is adopted.

[0032] First, the fused convolutional neural network, recurrent neural network, and neural network architecture are pre-trained separately, using a large number of labeled positive and negative sample images. During the pre-training process, unsupervised learning methods such as autoencoders or contrastive learning can be used to learn the common features of images.

[0033] Then, the three parts are jointly trained, and the model parameters are continuously adjusted through the back propagation algorithm and the optimization algorithm to minimize the loss function. The loss function adopts a multi-task loss function, combining image classification loss, image reconstruction loss and adversarial loss. Image classification loss is used to distinguish normal images from fraudulent images, which can be expressed as cross entropy loss. ,in is the true label, is the prediction probability; the image reconstruction loss adopts the formula: square error loss ,in is the original image, Generate images; use adversarial loss to train GAN, so that the images generated by the generator are more realistic, and the discriminator can distinguish between real images and generated images more accurately; The optimization algorithm uses the adaptive moment estimation (Adam) optimizer, and the formula is ,in is the current parameter, is the learning rate, is the first-order moment estimate, is the second-order moment estimate, is a very small constant.

[0034] In another embodiment of the present invention, the deep learning model is trained using the preprocessed data set, specifically as follows: Figure 4 As shown in the figure. A multi-stage training strategy is adopted. The convolutional neural network, recursive neural network and neural network architecture are pre-trained separately, and then jointly trained. The model parameters are continuously adjusted through the back propagation algorithm and the optimization algorithm to minimize the multi-task loss function. The model training process includes pre-training, joint training, forward propagation, loss function calculation, back propagation, parameter update and judging whether the stop condition is met.

[0035] Step 104: Input the image to be detected into the trained deep learning model for fraud detection and classification.

[0036] In one embodiment of the present invention, after training is completed, the image to be detected is input into the model, and the model will output the fraud probability or classification result of each image. Using a multi-level classification strategy, not only the binary classification result of whether the image is a fraudulent image is output, but also the specific fraud type can be output, such as identity impersonation, false advertising, forged documents, etc.

[0037] For the calculation of fraud probability, the output of the model can be converted into a probability distribution through the softmax function. The formula is , where Z is the output of the fully connected layer, is a computational function, It is the probability of fraud. Different thresholds are set to make judgments based on the probability of fraud and specific application requirements. For example, for high-risk application scenarios, a lower threshold can be set to increase the sensitivity of detection; for low-risk application scenarios, a higher threshold can be set to reduce the false alarm rate.

[0038] For multi-class classification, multiple classifiers can be added to the output layer of the model, corresponding to different fraud types. By training these classifiers, specific fraud types can be identified.

[0039] In another embodiment of the present invention, the model is further optimized and updated, and further, the model is optimized and updated regularly using new annotated data. Online learning and incremental learning techniques are used to enable the model to continuously learn and adapt to new fraud methods in a constantly changing environment. Online learning can receive new data and update it in real time, while incremental learning can learn new knowledge without forgetting old knowledge.

[0040] Model distillation is introduced to extract the knowledge of complex deep hybrid models and compress them into a smaller model for deployment on resource-constrained devices. Model distillation achieves knowledge transfer by training a small model to imitate the output of a large model.

[0041] At the same time, by introducing a feedback mechanism, the model can be further adjusted and optimized based on the detection results. For example, if the same new fraud method is detected multiple times in a row, the system can automatically adjust the model parameters to enhance the detection ability of this type of fraud.

[0042] In another embodiment of the present invention, model evaluation and optimization specifically include: Use the test set to evaluate, calculate the performance index, and determine whether optimization is needed. If optimization is not needed, end the operation. If optimization is needed, perform online learning, incremental learning, model distillation, feedback mechanism adjustment, and then return to the step of using the test set for evaluation.

[0043] In yet another embodiment of the present invention, the operation method of model deployment and real-time detection includes: The application system includes image input, model integration, fraud detection, real-time feedback, and returning to the application system for image input steps.

[0044] The method of the present invention has the following advantages: Extremely high accuracy: The innovative deep hybrid model and multi-stage training strategy achieve extremely high accuracy classification and recognition in image recognition and fraud detection tasks.

[0045] Powerful feature learning capabilities: The model can automatically learn rich and useful feature representations from large amounts of data without much human intervention. It can capture extremely subtle image features that are difficult to describe with traditional methods, and more effectively identify fraudulent images.

[0046] Excellent generalization ability: The multimodal feature extraction capability and multi-level classification strategy of the deep hybrid model enable the model to learn common features, easily process new images that have not appeared in the training data, and effectively respond to emerging fraud methods.

[0047] Strong adaptive and continuous learning capabilities: The application of online learning, incremental learning, and model distillation technology enables the model to quickly adapt to new fraud patterns, continuously evolve, and maintain its detection capabilities.

[0048] Fast processing speed and high scalability: The structural design of the deep hybrid model and the selection of optimization algorithms enable the model to efficiently process large amounts of data in parallel and quickly process large-scale image data sets. With the advancement of hardware technology, the processing speed and scalability of the model are further improved.

[0049] High robustness: The GAN image enhancement technology in the preprocessing module and the structural design of the deep hybrid model make the model more robust to common noise and changes in the image (such as lighting changes, occlusion, compression, etc.), more reliable in practical applications, and can maintain high detection performance even under harsh image acquisition conditions.

[0050] Powerful real-time detection capabilities: Due to the model’s efficient processing capabilities and the selection of optimized algorithms, it can be used in real-time or near-real-time fraud detection systems to provide users with immediate feedback and warnings.

[0051] Easy to integrate and deploy: It can be more easily integrated into existing image processing and security systems to provide a stronger security layer. With the cloudification of models and services, these models can be easily deployed in the cloud to provide users with flexible fraud detection services.

[0052] Significantly reduce labor costs: Compared with traditional methods, deep learning-based methods can automatically perform most detection tasks, significantly reduce labor costs and improve efficiency.

[0053] Highly adaptable: It can be more accurately customized and optimized for specific fields or applications, making it better suited to specific fraud detection needs, such as ID card fraud detection in the financial field or content fraud detection on social media.

[0054] In a second aspect of the present invention, a deep learning-based image anti-fraud device is provided, comprising: An acquisition module, used for acquiring image data and preprocessing the acquired image data; In one embodiment of the present invention, the acquisition module is specifically used to acquire image data from several channels using an automated data collection tool; and perform image normalization, size adjustment, denoising and image enhancement on the acquired image data.

[0055] Furthermore, the acquisition module adopts the formula: ,in is the original pixel value, is the normalized pixel value; Resize all images to the size required by the model input; Use a filter to remove image noise, using Ismooth=G∗I, where G is the Gaussian kernel, I is the original image, and Ismooth is the denoised image.

[0056] Building blocks for building deep learning models; Furthermore, the building block is specifically used to build an innovative deep hybrid model that integrates convolutional neural networks, recurrent neural networks, and neural network architectures, where: The convolution layer uses convolution kernels of different sizes to extract features at different levels of the image, using the formula: F=W*I+b, where W is the convolution kernel, I is the input image, b is the bias term, and F is the output feature map. The pooling layer reduces the spatial dimension of the feature and the number of parameters. The maximum pooling P=max(F), where P is the feature after pooling. Recurrent Neural Networks use the formula: ,in is the hidden state at the current moment, is the current input, is the weight matrix, is the bias term, is the activation function; In the neural network architecture, the self-attention mechanism is used to calculate the relationship between different regions in the image, thereby realizing the modeling of global information. The calculation formula of the self-attention mechanism is , where Q, K, and V are query vectors, key vectors, and value vectors, respectively. is the dimension of the key vector.

[0057] A training module is used to extract features and learn from the preprocessed image data using a deep learning model, and to perform deep learning model training; Furthermore, the training module is specifically used to adopt a multi-stage training strategy; Pre-train the fused convolutional neural network, recurrent neural network, and neural network architecture separately, using a large number of labeled positive and negative sample images; During pre-training, unsupervised learning methods, autoencoders or contrastive learning are used to learn the common features of images; The three parts are trained jointly, and the model parameters are continuously adjusted through the back propagation algorithm and the optimization algorithm to minimize the loss function. The loss function adopts a multi-task loss function, combining image classification loss, image reconstruction loss and adversarial loss. Among them, the image classification cross entropy loss ,in is the true label, is the prediction probability; the image reconstruction loss adopts the formula: square error loss ,in is the original image, Generate images; use adversarial loss to train GAN, so that the images generated by the generator are more realistic, and the discriminator can distinguish between real images and generated images more accurately; The optimization algorithm uses the adaptive moment estimation (Adam) optimizer, and the formula is ,in is the current parameter, is the learning rate, is the first-order moment estimate, is the second-order moment estimate, is a preset constant.

[0058] The detection module is used to input the image to be detected into the trained deep learning model for fraud detection and classification.

[0059] In one embodiment of the present invention, the detection module is further used to input the image to be detected into the model to calculate the fraud probability, the formula is , where Z is the output of the fully connected layer, is a computational function, The probability of fraud is determined by setting different thresholds and making judgments based on the probability of fraud and specific application requirements; Multi-level classification is adopted, and several classifiers are added to the output layer of the model, corresponding to different fraud types. By training these classifiers, specific fraud types can be identified.

[0060] In another embodiment of the present invention, the device further comprises an updating module for regularly using new annotated data to update the model using online learning and incremental learning techniques, so that the model can continuously learn and adapt to new fraud methods in a constantly changing environment; Introducing model distillation technology to compress complex deep hybrid models so that they can be deployed on resource-constrained devices; A feedback mechanism is introduced to further adjust and optimize the model based on the detection results.

[0061] In the third aspect of the present invention, an electronic device is provided. The terminal device in the embodiment of the present disclosure may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), etc., and fixed terminals such as digital TVs, desktop computers, etc.

[0062] An electronic device may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). In the RAM, various programs and data required for the operation of the electronic device are also stored. The processing device, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0063] Typically, the following devices may be connected to the I / O interface: input devices including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices including, for example, a magnetic tape, a hard disk, etc.; and communication devices. The communication devices may allow the electronic device to communicate with other devices wirelessly or by wire to exchange data.

[0064] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0065] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0066] The above-mentioned computer-readable medium carries one or more programs. When the above-mentioned one or more programs are executed by the electronic device, the electronic device: receives a voice signal from a first device; parses the voice signal to obtain a second device identifier and a second device control instruction; sends the second device control instruction to the second device according to the second device identifier; receives the execution result of the second device control instruction from the second device; and sends the execution result of the second device control instruction to the first device.

[0067] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0068] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in the embodiments, or replace some of the technical features by equivalents, based on reading the specification of the present invention; and 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 deep learning-based image anti-fraud method, characterized in that: include: Collecting image data and preprocessing the collected image data; Build deep learning models; Use deep learning models to extract features and learn from preprocessed image data, and perform deep learning model training; The images to be tested are fed into the trained deep learning model for fraud detection and classification.

2. The method according to claim 1, characterized in that The collecting of image data and preprocessing of the collected image data include: Use automated data collection tools to obtain image data from several sources; The acquired image data is normalized, resized, denoised and enhanced.

3. The method according to claim 2, characterized in that The step of normalizing, resizing, denoising and enhancing the acquired image data includes: Using the formula: ,in is the original pixel value, is the normalized pixel value; Resize all images to the size required by the model input; Use a filter to remove image noise, using Ismooth=G∗I, where G is the Gaussian kernel, I is the original image, and Ismooth is the denoised image.

4. The method according to claim 3, characterized in that The deep learning model is constructed, comprising: Build an innovative deep hybrid model that combines convolutional neural network, recurrent neural network and neural network architectures, where The convolution layer uses convolution kernels of different sizes to extract features at different levels of the image, using the formula: F=W*I+b, where W is the convolution kernel, I is the input image, b is the bias term, and F is the output feature map. The pooling layer reduces the spatial dimension of the feature and the number of parameters. The maximum pooling P=max(F), where P is the feature after pooling. Recurrent Neural Networks use the formula: ,in is the hidden state at the current moment, is the current input, is the weight matrix, is the bias term, is the activation function; The self-attention mechanism is used to calculate the relationship between different regions in the image, so as to achieve the modeling of global information. The calculation formula of the self-attention mechanism is: , where Q, K, and V are query vectors, key vectors, and value vectors, respectively. is the dimension of the key vector.

5. The method according to claim 4, characterized in that The deep learning model training includes: Adopt a multi-stage training strategy; Pre-train the fused convolutional neural network, recurrent neural network, and neural network architecture separately, using a large number of labeled positive and negative sample images; During pre-training, unsupervised learning methods, autoencoders or contrastive learning are used to learn the common features of images; The three parts are trained jointly, and the model parameters are continuously adjusted through the back propagation algorithm and the optimization algorithm to minimize the loss function. The loss function adopts a multi-task loss function, combining image classification loss, image reconstruction loss and adversarial loss. Among them, the image classification loss adopts the formula: cross entropy loss ,in is the true label, is the prediction probability; the image reconstruction loss adopts the formula: mean square error loss ,in is the original image, Generate images; use adversarial loss to train GAN, so that the images generated by the generator are more realistic, and the discriminator can distinguish between real images and generated images more accurately; The optimization algorithm uses an adaptive moment estimation optimizer, and the formula is ,in is the current parameter, is the learning rate, is the first-order moment estimate, is the second-order moment estimate, is a preset constant.

6. The method according to claim 5, characterized in that The image to be detected is input into the trained deep learning model to perform fraud detection and classification, including: Input the image to be detected into the model and calculate the fraud probability using the formula , where Z is the output of the fully connected layer, is a computational function, The probability of fraud is determined by setting different thresholds and making judgments based on the probability of fraud and specific application requirements; Multi-level classification is adopted, and several classifiers are added to the output layer of the model, corresponding to different fraud types. By training these classifiers, specific fraud types can be identified.

7. The method according to claim 6, characterized in that The deep learning model also includes: Regularly use new annotated data to train the model using online learning and incremental learning techniques, so that the model can continuously learn and adapt to new fraud methods in a changing environment; Introducing model distillation technology to compress complex deep hybrid models so that they can be deployed on resource-constrained devices; A feedback mechanism is introduced to further adjust and optimize the model based on the detection results.

8. An image anti-fraud device based on deep learning, characterized in that: include: An acquisition module, used for acquiring image data and preprocessing the acquired image data; Building blocks for building deep learning models; A training module is used to extract features and learn from the preprocessed image data using a deep learning model, and to perform deep learning model training; The detection module is used to input the image to be detected into the trained deep learning model for fraud detection and classification.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, and when the program is executed, the method according to any one of claims 1 to 7 can be implemented.

Images