Network Risk Control Neural Rule Update Method and Device Based on Feature Pattern Constraints

Abstract

Network risk control neural rule update method and device based on feature pattern consistency constraint. The method includes: S1. Determine the neural rules to be updated, mainly including the neural rules in the deep image classifier based on the convolutional neural network; S2. Randomly strengthen the input data of the neural rules within a well-constrained range; S3. Run the neural rules and calculate the loss based on the feature pattern consistency constraint algorithm and the classification result; S4. Optimize the loss through backpropagation to achieve the update of the neural rules. The present invention can complete the update of neural rules for network risk control on the premise of introducing a small additional computational overhead, provides a neural rule update method with better network risk control effect, and can more effectively improve the security and robustness of the target network; provides a flexible neural rule update method for network risk control that can be implemented in a plug-and-play manner at any usage stage of the network, including the initial stage and the ex-post stage.

Description

Technical Field

[0001] The invention relates to the fields of deep neural networks, network risk control, and image classification, and particularly to a method and device for updating neural rules for network risk control based on feature pattern consistency constraints. Background Art

[0002] Deep neural network technology: A model composed of multiple layers of neurons in computer science, which simulates the connection mode of the biological nervous system and learns and processes information through a series of neural rules. A variety of deep neural networks have been constructed in the prior art and are widely used in fields such as image recognition, speech processing, and natural language processing. The image classifier based on the deep neural network is the object of the present invention.

[0003] Neural rules and their updates: In the prior art, the process of signal transmission, update, learning, and creation of input and output mappings between neurons in a deep neural network is abstracted into neural rules, and the deep neural network is regarded as a combination of several neural rules. Some neural rules in the deep neural network are parameterized. In order to make the network adapt to a specific scenario and meet the requirements, it is necessary to update the parameters in the neural rules. The prior art calculates the loss function based on the results and objectives of network operation, and adjusts the network parameters by backpropagating the loss, so as to optimize and update along the gradient direction. However, the present invention aims at network risk control and introduces a new calculation term into the loss function. The neural rules are updated by optimizing on the new loss function, and finally the goal of network risk control is achieved.

[0004] Features and feature patterns: The prior art believes that input samples will be abstracted into features by neural rules in a deep neural network, usually presented in the form of vectors with a certain dimension. Based on this, the present patent proposes a new concept of feature pattern as a supplement to the original concept, and further proposes a feature subset selection operator and a feature pattern constraint operator for updating neural rules for network risk control.

[0005] Image classification and network risk control: The most widely applied task of the prior art using deep neural networks is the image classification task, which is also the task faced by the present invention, and the goal is to classify the input image. The prior art usually hopes to improve the accuracy of image classification by updating neural rules. Network risk control focuses on network security and robustness, aiming to reduce application risks. Deep neural networks are widely used in fields with extremely high security requirements such as medicine, autonomous driving, security, and finance. Therefore, network risk control is crucial in these fields. However, the existing technologies are difficult to balance the accuracy, security, and robustness of the network. One of the goals of the present invention to update neural rules is to improve the network risk control ability on the premise of ensuring network accuracy, and strengthen the security and robustness of the network.

[0006] The prior art has the following deficiencies:

[0007] (1) When implementing network risk control in the prior art, it is usually necessary to introduce attack samples during the network training process so that the network can learn how to deal with attacks. However, this method is prone to overfitting, which significantly reduces the accuracy of the network on normal samples. In contrast, the method of this patent updates the neural rules by learning robust and stable feature patterns on normal samples, effectively alleviating the overfitting phenomenon.

[0008] (2) Another method in the prior art for implementing network risk control is to filter the network input to eliminate harmful information. However, such methods have the problem of unstable effects and cannot effectively deal with diverse harmful samples. The method of this patent avoids these limitations by enhancing the security and robustness of the network itself without relying on filtering operations.

[0009] (3) When implementing network risk control in the prior art, it usually requires consuming a large amount of additional computing resources. The method of this patent only introduces the loss calculated based on the feature pattern consistency algorithm during neural rule update, with relatively small additional computing overhead; and during actual use, no additional computing overhead is required.

[0010] (4) The network risk control methods in the prior art often lack flexibility. The method of this patent has high flexibility and can be introduced in any usage stage of the target neural network in a plug-and-play form, thereby updating the neural rules in the target network to achieve the purpose of network risk control. Summary of the Invention

[0011] The present invention aims to solve the problems of the prior art in implementing network risk control, such as generating a large amount of additional computing overhead, unsatisfactory effects, and lack of flexibility. It provides a method and device for updating neural rules for network risk control based on feature pattern consistency constraints. By introducing this method at any usage stage of the network, the neural rules in the network are updated to achieve the effects of enhancing the security and robustness of the network in network risk control.

[0012] The present invention is implemented through the following technical solutions:

[0013] The first aspect of the present invention relates to a method for updating neural rules for network risk control based on feature pattern consistency constraints, specifically including:

[0014] S1. Determine the neural rules to be updated, including the neural rules in the depth image classifier based on a convolutional neural network;

[0015] S2. Randomly strengthen the input data of the neural rules within a constrained range;

[0016] S3. Run the neural rules and calculate the loss based on the feature pattern consistency constraint algorithm and the classification results ;

[0017] S4. Optimize the loss through backpropagation to update the neural rules

[0018] The determination of the neural rules to be updated in step S1 includes:

[0019] S11. Represent the neural rules as a deep image classifier based on a convolutional neural network. The neural rules are composed of a rule set including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Among them, the input layer is used to represent the input image, and the input image is a matrix , where , represents the height and width of the image, represents the number of channels of the image

[0020] S12. The convolutional layer extracts the features of the input image through convolutional operations; the pooling layer performs dimensionality reduction on the extracted features; the fully connected layer maps the high-dimensional features after dimensionality reduction into a vector of a fixed dimension; the output layer generates the probability distribution of the classification results through an activation function

[0021] S13. The neural rules implement the image classification task in the following way: ensure the correct connection of the input layer, convolutional layer, pooling layer, fully connected layer, and output layer in the rule set, and optimize and update the parameters involved in the rules through a parameter update mechanism, so as to correctly execute the image classification task

[0022] The random reinforcement in step S2 includes: for each sample in the input layer data , add a random noise , to obtain the reinforced sample , and use as the new training sample to replace .

[0023] The loss in step S3 is composed of the classification loss and the loss calculated based on the feature pattern consistency algorithm , and is specifically expressed as

[0024] , (1)

[0025] represents the balance factor

[0026] The classification loss is the cross-entropy between the prediction result and the true result obtained by running the neural rules:

[0027] , (2)

[0028] is the number of all input samples, is the logarithmic function, is the total number of classes, is the natural logarithm is the exponential function with base is the probability distribution vector of the classification result output by the neural rule, and are the -th and -th digits of this vector respectively.

[0029] The loss calculated based on the feature pattern consistency algorithm is jointly calculated by the feature subset selection operator and the feature pattern constraint operator.

[0030] The described feature subset selection operator includes the following method: Among the sample features output by part of the neural rules, select a subset of features, including: For the sample at a certain layer before the -th layer of the network, randomly select a feature subset from the features output by the neural rules, where , and is a binary vector drawn from the Bernoulli distribution . Use to replace as the input of the subsequent neural rules.

[0031] As shown in Figure 4 , the described feature pattern constraint operator includes the following method: Calculate the loss , and its form is:

[0032] , (3)

[0033] , (4)

[0034] where represents the feature pattern of the sample in the -th layer of the network; represents the feature pattern of the sample true class , which is the parameter learned in the neural rule of the -th layer of the network; is the feature pattern constraint operator, which is and The distance quantifies the difference between the characteristic pattern of the sample and the characteristic pattern of its true class; is a set of layers, indicating that only the loss of the network layer is calculated.

[0035] For the input sample , the characteristic pattern of the sample in the layer network of the

[0036] is calculated as follows: (5)

[0037] where represents the feature output after the sample passes through the layer network, which is a vector, denoted as the length of this vector, and are calculated through the following formula.

[0038] is calculated as follows: (6)

[0039] is calculated as follows: (7)

[0040] represents the balance factor.

[0041] The specific method for optimizing the loss through backpropagation as described in step S4 to update the neural rules is as follows: By performing backpropagation on the loss value generated after running the neural rules, the parameters involved in the neural rules are updated based on the gradient descent algorithm to optimize and update the neural rules in a training manner.

[0042] The second aspect of the present invention relates to a network risk control neural rule update device based on feature pattern consistency constraint, including a memory and one or more processors. Executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the network risk control neural rule update method of the present invention.

[0043] The working principle of the present invention is:

[0044] (1) A new loss function is constructed with network risk control as the goal to complete the update of neural rules. Therefore, compared with the normal training process, the introduced additional computational overhead is small;

[0045] (2) The present invention achieves the purpose of network risk control by learning robust feature patterns from normal samples, so it has less impact on the accuracy of normal samples and can more effectively improve the security and robustness of the network compared with the existing technologies.

[0046] (3) The present invention constructs a neural rule update process by inputting random reinforcement, feature subset selection operator and feature pattern constraint operator, and has a low coupling degree with the deep neural network constructed by the existing technology. Therefore, it can be introduced into each usage stage of the network in a plug-and-play manner and has good flexibility.

[0047] The innovation points of the present invention are:

[0048] (1) Based on the concept of network features, the concept of feature patterns is proposed, revealing the essence of deep neural network learning.

[0049] (2) Based on the concept of feature patterns proposed by the present invention, novel feature subset selection operator and feature pattern constraint operator are designed and used for network risk control neural rule update, which is closer to the essence of strengthening the network itself compared with the existing methods.

[0050] (3) The method of the present invention has the advantages of small additional computational overhead, good effect and flexible use compared with the existing methods.

[0051] The advantages of the present invention are:

[0052] (1) The present invention can complete the neural rule update for the purpose of network risk control with less additional computational overhead;

[0053] (2) In terms of the effect of network risk control, the present invention has advantages compared with the existing methods, mainly manifested as being able to more effectively improve the security and robustness of the network while having less impact on the accuracy of the network on normal samples;

[0054] (3) The present invention has better flexibility compared with the existing methods. It can act on each usage stage of the network, including the initial stage that has not been trained and the post-training stage, and can be introduced in a plug-and-play manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the network risk control neural rule update method based on feature pattern consistency constraint of the present invention.

[0056] Figure 2 It is a schematic diagram of randomly reinforcing the input samples in the present invention, including images of 10 categories of the CIFAR-10 image dataset.

[0057] Figure 3Schematic diagram of the feature subset selection operator of the present invention. Each circle in the picture represents a feature. After the feature subset operator, the feature subset selected based on the Bernoulli distribution is retained, and the white dashed circle represents the discarded feature, and the corresponding vector value is set to 0.

[0058] Figure 4 Schematic diagram of the feature pattern constraint operator of the present invention. The feature pattern of the sample and the feature pattern of the corresponding correct category are used to calculate the feature pattern consistency constraint loss, and finally used for neural rule update.

[0059] Figure 5a and Figure 5b Schematic diagram of the neural rule structure of the method of the present invention introduced into two deep neural networks for image classification. The repeated residual block neural rule structure is simply represented by a dashed box, where Figure 5a is the WRN-28-10 deep neural network, Figure 5b is the ResNet-50 deep neural network.

[0060] Figure 6 is the neural rule structure diagram of VGG-16 of the present invention.

[0061] Figure 7 is the schematic diagram of the device of the present invention. Detailed implementation manners

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0063] Example 1

[0064] As Figure 1 shown, this embodiment relates to a method for updating neural rules for network risk control based on feature pattern consistency constraints. The specific method is as follows:

[0065] S1. Determine the neural rules to be updated, including the neural rules in the deep image classifier based on the convolutional neural network;

[0066] S2. Randomly strengthen the input data of the neural rules within a constrained range;

[0067] S3. Run the neural rules and calculate the loss L based on the feature pattern consistency constraint algorithm and the classification result;

[0068] S4. Optimize the loss through backpropagation to achieve the update of the neural rules.

[0069] The determination of the neural rules to be updated includes:

[0070] S11. Represent the neural rule as a deep image classifier based on a convolutional neural network. The neural rule consists of a rule set including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Among them, the input layer is used to represent the input image, and the input image is a matrix , where , represents the height and width of the image, represents the number of channels of the image.

[0071] S12. The convolutional layer extracts the features of the input image through convolutional operations; the pooling layer performs dimensionality reduction on the extracted features; the fully connected layer maps the high-dimensional features after dimensionality reduction into vectors of a fixed dimension; the output layer generates the probability distribution of the classification result through an activation function.

[0072] S13. The neural rule implements the image classification task in the following way: ensuring the correct connection of the input layer, convolutional layer, pooling layer, fully connected layer, and output layer in the rule set, and optimizing and updating the parameters involved in the rule through a parameter update mechanism, so as to correctly execute the image classification task.

[0073] As Figure 2 shown, the random reinforcement described in step S2 includes: for each sample in the input layer data, adding a random noise to obtain the reinforced sample , and using as the new training sample to replace .

[0074] The loss described in step S3 is composed of the classification loss and the loss calculated based on the feature pattern consistency algorithm, and is specifically expressed as

[0075] , (1)

[0076] represents the balance factor.

[0077] The classification loss is the cross-entropy between the prediction result obtained by running the neural rule and the true result:

[0078] , (2)

[0079] is the number of all input samples, is the logarithmic function, is the total number of classes, is the natural logarithm The exponential function with base is the probability distribution vector of the classification result of the neural rule output, and are the -th and

[0080] The loss calculated based on the feature pattern consistency algorithm is jointly calculated by the feature subset selection operator and the feature pattern constraint operator.

[0081] As Figure 3 shown, the feature subset selection operator includes the following method: In the sample features of the partial neural rule output, select a subset of features, including: For the sample at a certain layer before the -th layer of the network, in the features of the neural rule output randomly select a feature subset based on the Bernoulli distribution , where is a binary vector drawn from the Bernoulli distribution . Use to replace as the input of the subsequent neural rule.

[0082] As Figure 4 shown, the feature pattern constraint operator includes the following method: Calculate the loss , whose form is:

[0083] , (3)

[0084] , (4)

[0085] where represents the feature pattern of the sample in the -th layer of the network; represents the feature pattern of the sample true class , which is the parameter learned in the neural rule of the -th layer of the network; is the feature pattern constraint operator, which is the and 's distance, quantifying the difference between the feature pattern of the sample and the feature pattern of its true class; is a set of layers, indicating that only the loss of the network layer is calculated.

[0086] For the input sample , the feature pattern of the sample in the n-th layer network is calculated as follows: That is, (5)

[0087] , (5)

[0088] Among them, represents the feature output by the sample after passing through the n-th layer network, which is a vector, denoted as the length of this vector, and are calculated through the following formula. and That is, (6)

[0089] , (6)

[0090] , (7)

[0091] represents the balance factor.

[0092] The specific method for optimizing the loss through backpropagation and updating the neural rules described in step S4 is as follows: By performing backpropagation on the loss value generated after running the neural rules, the parameters involved in the neural rules are updated based on the gradient descent algorithm to achieve the optimization and update of the neural rules in a training manner.

[0093] The overall network risk control neural rule update method based on feature pattern consistency constraint is as Figure 1 shown.

[0094] The positive effects of the present invention are as follows:

[0095] (1) Complete the update of neural rules for the purpose of network risk control on the premise of introducing a relatively small additional computational overhead;

[0096] (2) Implement a neural rule update method with better network risk control effect, which can more effectively improve the security and robustness of the target network;

[0097] (3) Provide a flexible neural rule update method for network risk control that can be implemented in a plug-and-play manner at any usage stage of the network, including the initial stage and the post-event stage.

[0098] To verify the effectiveness of the present invention and the comparison between this technology and existing methods, the present invention uses CIFAR-10 as the image data set for verification, which contains 10 image categories, as shown in the example Figure 2 shown; Use the deep image classification networks WRN-28-10 and ResNet-50 as verification objects, asFigure 5a and Figure 5b As shown, the method of the present invention is introduced in a plug-and-play manner, with PGDAT, TRADES, and MARGIN as the comparison methods, and PGD, AutoPGD, and EOT-PGD as the deep attack methods. OURS represents the patented method. The comparison results are the accuracy on normal samples and the accuracy under deep attacks, and the latter reflects the security and robustness of the network. The higher it is, the better the network risk control effect of the corresponding method. The method of the present invention is optimal in terms of the accuracy on normal samples and the network risk control effect. Table 1 shows the comparison of the accuracy and robustness of the method of the present invention and other methods on the CIFAR-10 dataset and under deep attacks.

[0099] Table 1

[0100]

[0101] Example 2

[0102] This example relates to a method for enhancing the security of a medical image recognition network by applying the network risk control neural rule update method based on feature pattern consistency constraint of the present invention. The steps include:

[0103] S1. Determine the neural rules to be updated, which are determined to be the neural rules in the VGG-16 medical image recognition network in this example;

[0104] S2. Randomly reinforce the input data of the neural rules within a constrained range;

[0105] S3. Run the neural rules and calculate the loss L based on the feature pattern consistency constraint algorithm and the classification results;

[0106] S4. Optimize the loss through backpropagation to achieve the update of the neural rules.

[0107] S5. Through the update of the neural rules, achieve the enhancement of the security of the VGG-16 medical image recognition network.

[0108] The determination of the neural rules to be updated includes: in this example, it is determined to be the neural rules in the VGG-16 medical image recognition network based on a convolutional neural network, and the neural rule structure is as Figure 6 shown;

[0109] As Figure 2 shown, the random reinforcement described in step S2 includes: for each sample in the input layer data , add a random noise , , to obtain the reinforced sample , and use as the new training sample to replace 。

[0110] The loss described in step S3 is composed of the classification loss and the loss calculated based on the feature pattern consistency algorithm and is specifically expressed as

[0111] , (1)

[0112] represents the balance factor.

[0113] The classification loss is the cross - entropy between the prediction result and the true result obtained by running the neural rule:

[0114] , (2)

[0115] is the number of all input samples, is the logarithmic function, is the total number of classes, is the exponential function with the natural logarithm as the base, is the probability distribution vector of the classification result output by the neural rule, and are the -th and

[0116] The loss calculated based on the feature pattern consistency algorithm is jointly calculated by the feature subset selection operator and the feature pattern constraint operator.

[0117] The loss described in step S3 is composed of the classification loss and the loss calculated based on the feature pattern consistency algorithm and is specifically expressed as

[0118] , (1)

[0119] represents the balance factor.

[0120] The classification loss is the cross - entropy between the prediction result and the true result obtained by running the neural rule:

[0121] , (2)

[0122] is the number of all input samples, is a logarithmic function, is the total number of categories, is the exponential function with the natural logarithm as the base, is the probability distribution vector of the classification result of the neural rule output, and are the -th and

[0123] The loss calculated based on the feature pattern consistency algorithm is obtained by jointly calculating the feature subset selection operator and the feature pattern constraint operator.

[0124] As Figure 3 shown, the feature subset selection operator includes the following method: Among the sample features of the partial neural rule output, select a subset of features, including: For sample at a certain layer before the -th layer of the network, for the features of the neural rule output randomly select a feature subset based on the Bernoulli distribution , where is a binary vector drawn from the Bernoulli distribution . Use to replace as the input of the subsequent neural rule.

[0125] As Figure 4 shown, the feature pattern constraint operator includes the following method: Calculate the loss , whose form is:

[0126] , (3)

[0127] , (4)

[0128] where represents the feature pattern of sample in the -th layer of the network; represents the feature pattern of sample 's true category , which is the parameter learned in the neural rule of the -th layer of the network; is the feature pattern constraint operator, which is the and 's distance, quantifying the difference between the feature pattern of the sample and the feature pattern of its true category; is a set of layers, indicating that only the loss of the network layer is calculated.

[0129] For the input sample , the feature pattern of the sample in the -th layer network is calculated as follows:

[0130] , (5)

[0131] where represents the feature output after the sample passes through the -th layer network, which is a vector, denoted as the length of this vector, and are calculated through the following formula.

[0132] , (6)

[0133] , (7)

[0134] represents the balance factor.

[0135] The specific method for optimizing the loss through backpropagation and updating the neural rules as described in step S4 is as follows: By performing backpropagation on the loss value generated after running the neural rules, the parameters involved in the neural rules are updated based on the gradient descent algorithm to optimize and update the neural rules in a training manner.

[0136] The specific method described in step S5 is as follows: After the neural rules are updated, a medical image recognition network with stronger security can be obtained, which can better defend against attacks on the network.

[0137] To verify the effectiveness of the method for enhancing the security of the medical image recognition network using the network risk control neural rule update method based on feature pattern consistency constraint of the present invention, PGD-White and PGD-Black are used as methods to attack the medical image recognition network VGG-16 before and after enhancing the security using this method. The success rates of the attacks are shown in Table 2. The lower the attack success rate, the safer the network. It can be seen that the method of the present invention greatly improves the security of the network. Table 2 is the comparison of the security of the VGG-16 network before and after using the method for enhancing the security of the medical image recognition network using the network risk control neural rule update method based on feature pattern consistency constraint of the present invention.

[0138] Table 2

[0139]

[0140] Embodiment 3

[0141] As Figure 7 , this embodiment relates to a network risk control neural rule update device based on feature pattern consistency constraints, including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the network risk control neural rule update method based on feature pattern consistency constraints in Embodiment 1 of this embodiment.

[0142] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept of the present invention.

Claims

1. A network risk control neural rule updating method based on characteristic pattern constraints, characterized in that: Specifically include: S1. Determine the neural rules that need to be updated, including a deep image classifier based on a convolutional neural network; S2, randomly strengthen the input data of the neural rule within a constrained range; S3. Run neural rules and calculate losses based on feature pattern consistency constraint algorithm and classification results ; Loss calculated based on the characteristic pattern consistency algorithm Depend on The feature subset selection operator and the feature pattern constraint operator are calculated together; The characteristic mode constraint operator includes the following method: calculating the loss , which is of the form: ,in Indicates Samples in layer network characteristic patterns; Representation sample True category The characteristic mode is The parameters learned in the neural rules of the layer network; is the characteristic mode constraint operator, and of Distance, which quantifies the difference between the characteristic pattern of the sample and the characteristic pattern of its true category; Is a layer set, indicating that only Network layer losses; For the input sample , No. Samples in layer network The characteristic pattern The calculation is performed as follows: ,in, Representation sample After The feature of the layer network output is a vector, is the length of the vector, and Calculated using the following formula: , represents the balance factor; S4. Optimize the loss through back-propagation to update the neural rules.

2. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 1 is characterized in that: The neural rules that need to be updated in step S1 include: S11. The neural rule is represented as a deep image classifier based on a convolutional neural network, wherein the neural rule is composed of a rule set including an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer; wherein the input layer is used to represent an input image, and the input image is a matrix ,in Indicates the height and width of the image, Indicates the number of channels of the image; S12, the convolution layer extracts the features of the input image through convolution operations; the pooling layer performs dimensionality reduction on the extracted features; the fully connected layer maps the reduced high-dimensional features into vectors of fixed dimensions; the output layer generates the probability distribution of the classification results through the activation function; S13. The neural rule implements the image classification task in the following manner: ensuring the correct connection of the input layer, convolution layer, pooling layer, fully connected layer and output layer in the rule set, and optimizing and updating the parameters involved in the rule through a parameter update mechanism, thereby achieving the correct execution of the image classification task.

3. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 1 is characterized in that: The random reinforcement described in step S2 includes: for each sample in the input layer data , add a random noise , get the modified feature sample ,use As a new training sample replacement .

4. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 1 is characterized in that: The loss in step S3 By classification loss and the loss calculated based on the characteristic pattern consistency algorithm The composition is specifically expressed as , Represents the balance factor.

5. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 4 is characterized in that: Classification Loss The cross entropy between the predicted and true results obtained by running the neural rule: , is the number of all input samples, is a logarithmic function, is the total number of categories, The natural logarithm The exponential function with base , is the probability distribution vector of the classification results output by the neural rule, and are the first and Number of digits.

6. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 4 is characterized in that: The feature subset selection operator includes the following method: selecting a subset of features from the sample features output by some neural rules, including: for the sample In the network Features of the neural rule output of a layer before the layer In the example, a subset of features is randomly selected based on the Bernoulli distribution. ,in , is a Bernoulli distribution The binary vector extracted from ; use Alternative As input to subsequent neural rules.

7. The network risk control neural rule updating method based on characteristic pattern constraints according to claim 1 is characterized in that: The specific method of optimizing the loss by back propagation and updating the neural rules as described in step S4 is as follows: by back propagating the loss value generated after running the neural rules, the parameters involved in the neural rules are updated based on the gradient descent algorithm, and the neural rules are optimized and updated in a training manner.

8. A network risk control neural rule updating device based on characteristic pattern constraints, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the network risk control neural rule updating method based on characteristic pattern constraints as described in any one of claims 1-7.

Images