Model parameter adjustment methods, devices, storage media, and program products

Abstract

The method, device, storage medium, and program product for adjusting model parameters relate to the field of secure multi-party computing, such as multi-party machine learning, federated learning, and federated modeling. The method comprises the steps of: performing classification processing on input data using a classification model obtained by training based on secure multi-party computing to obtain a classification prediction value of the classification model (S201); performing down processing on the classification prediction value (S202); performing normalization processing on the classification prediction value after down processing to obtain a normalized result of the classification prediction value (S203); and updating the parameters of the classification model based on the normalized result of the classification prediction value (S204). By down processing the classification prediction value to reduce the input value of the normalization method, data overflow can be prevented during the normalization process, and thereby the training precision of the classification model can be improved.

Description

Technology Field

[0001] The present application relates to the field of secure multi-party computing, such as multi-party machine learning, federated learning, and federated modeling, and in particular to a method for adjusting model parameters, a device, a storage medium, and a program product.

[0002] The present application claims priority to a Chinese patent application filed with the Chinese Patent Office on January 22, 2021, with application number 202110088699.1 and titled "Method for adjusting model parameters, apparatus, storage medium and program product," all of which are incorporated into the present application by reference. Background Technology

[0003] In machine learning based on Secure Multi-Party Computing (abbreviated as MPC or SMC), secure multi-party computing technologies are often used to protect the data privacy of each participating party during machine learning.

[0004] In the training process of a classification model based on secure multi-party computing, when normalizing the classification prediction values ​​of the current classification model using a softmax function, the exponential operation on the classification prediction values ​​generates very large numbers, which easily exceeds the data display range of secure multi-party computing and causes a data overflow problem, which degrades the precision of the classification model obtained through training and even causes anomalies in precision, making it unusable. The problem to be solved

[0005] The present application provides a method for adjusting model parameters, a device, a storage medium, and a program product. means of solving the problem

[0006] According to the first aspect of the present application, a method for adjusting model parameters is provided.

[0007] A step of performing classification processing on input data using a classification model obtained by training based on secure multi-party computing, and obtaining a classification prediction value of said classification model;

[0008] A step of performing reduction processing on the above classification prediction values;

[0009] A step of performing normalization processing on the classification prediction value after reduction processing to obtain the normalized result of the classification prediction value;

[0010] It includes a step of updating the parameters of the classification model based on the normalization result of the above classification prediction value.

[0011] According to a second aspect of the present application, a model parameter adjustment device is provided,

[0012] A classification processing module that performs classification processing on input data using a classification model obtained by training based on secure multi-party computing, and obtains a classification prediction value of the classification model;

[0013] A reduction processing module that performs reduction processing on the above classification prediction values;

[0014] A normalization processing module that performs normalization processing on the classification prediction value after reduction processing to obtain the normalization result of the classification prediction value;

[0015] It includes a parameter update module that updates the parameters of the classification model based on the normalization result of the above classification prediction value.

[0016] According to a third aspect of the present application, an electronic device is provided,

[0017] At least one processor; and

[0018] It includes a memory connected to the above-mentioned at least one processor; wherein,

[0019] The memory stores an instruction that can be executed by the at least one processor, and the instruction is executed by the at least one processor so that the at least one processor can execute a method according to the first aspect.

[0020] According to a fourth aspect of the present application, a non-transient computer-readable storage medium is provided in which computer instructions are stored, wherein the computer instructions are intended to enable a computer to execute a method according to a first aspect.

[0021] According to a fifth aspect of the present application, a computer program product is provided, said program product comprises a computer program, said computer program is stored in a readable storage medium, and at least one processor of an electronic device can read said computer program from said readable storage medium, said at least one processor executes said computer program so that the electronic device executes the method according to a first aspect. According to a sixth aspect of the present application, a computer program stored in a computer-readable storage medium is provided, said computer program is executed by a processor, thereby implementing the method according to a first aspect. Effects of the invention

[0022] The technology according to the present application improves the precision of a classification model obtained by training based on secure multi-party computing.

[0023] It should be understood that the contents described in this section are not intended to indicate the core or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application can be easily understood from the specification below. Brief explanation of the drawing

[0024] The attached drawings are provided to facilitate a more complete understanding of the present plan and do not constitute a limitation to the present application. Herein, FIG. 1 is a diagram illustrating a machine learning architecture based on multilateral secure computing according to an embodiment of the present application. FIG. 2 is a flowchart of a method for adjusting model parameters according to a first embodiment of the present application. FIG. 3 is a flowchart of a method for adjusting model parameters according to a second embodiment of the present application. FIG. 4 is a flowchart of a funeral method based on secure multi-party computing according to a second embodiment of the present application. FIG. 5 is a diagram showing the process of a plain text writing method according to the second embodiment of the present application. FIG. 6 is a drawing showing a model parameter adjustment device according to the third embodiment of the present application. FIG. 7 is a drawing showing a model parameter adjustment device according to the fourth embodiment of the present application. FIG. 8 is a block diagram of an electronic device for implementing a model parameter adjustment method of an embodiment of the present application. Specific details for implementing the invention

[0025] Exemplary embodiments of the present application are described below in conjunction with the accompanying drawings. For the sake of understanding, various details of the embodiments of the present application are included and should be considered merely illustrative. Accordingly, those skilled in the art should understand that various changes and modifications may be made to the embodiments described herein, and that such changes and modifications do not depart from the scope and spirit of the present application. Likewise, for the sake of clarity and brevity, descriptions of known functions and structures are omitted from the following description.

[0026] The present application provides a method for adjusting model parameters, a device, a storage medium, and a program product, which can be applied to secure multi-party computing fields such as multi-party machine learning, federated learning, and federated modeling to improve the precision of a classification model obtained by training based on secure multi-party computing.

[0027] The model parameter adjustment method according to the present embodiment can be applied to a machine learning architecture based on multi-party secure computing as illustrated in FIG. 1. In a machine learning architecture based on multi-party secure computing as illustrated in FIG. 1, a plurality of data holders (explained exemplarily using the three data holders, Data Holder 1, Data Holder 2, and Data Holder 3 of FIG. 1) jointly perform training of a classification model based on a multi-party secure computing (MPC) protocol based on the privacy data they possess. Since neither side can know the privacy data of either side, this implements improving the value of data usage while simultaneously ensuring the safety of their own privacy data.

[0028] In the process of federated training of a classification model based on secure multi-party computing, normalization processing must be performed on the classification prediction values ​​of the current classification model using a softmax function. Exponential operations on classification prediction values ​​generate very large numbers and easily exceed the data display range of secure multi-party computing, resulting in a data overflow problem, which degrades the precision of the classification model obtained through training and even makes it unusable due to a lack of precision.

[0029] The model parameter adjustment method according to the present embodiment can improve the precision of a classification model obtained by training based on secure multi-party computing by preventing data overflow during normalization processing based on secure multi-party computing.

[0030] As must be explained, the data processing related to the model parameter adjustment method according to the present application is entirely based on secure multi-party computing, and unless otherwise specified, the related data is entirely ciphertext data, and the processing performed is entirely based on ciphertext.

[0031] FIG. 2 is a flowchart of a method for adjusting model parameters according to a first embodiment of the present application. As shown in FIG. 2, the specific steps of the method are as follows.

[0032] Step (S201), using a classification model obtained by training based on secure multi-party computing, classification processing is performed on the input data to obtain the classification prediction value of the classification model.

[0033] Here, the classification prediction value of the classification model includes the prediction value corresponding to each classification of the classification model.

[0034] In the training process of a classification model based on secure multi-party computing, classification processing is performed on input data using the current classification model obtained by training based on secure multi-party computing, and the classification prediction value of the current classification model is obtained.

[0035] In this embodiment, the input data may be data from a training set or a test set, and may be various types of data such as images, text, and voice, and this embodiment is not specifically limited herein.

[0036] In data processing based on secure multi-party computing, ciphertext is generally represented using integers, while floating-point data types are represented as fixed-point numbers and then expanded to integers. Therefore, the actual range of data that can be represented in data processing based on secure multi-party computing is limited.

[0037] Step (S202), reduce the classification prediction value.

[0038] In data processing based on secure multi-party computing, multiple input values The normalization method (softmax) for is, Defined as, is the input value( Represents the normalization result of ), and is the number of input values, and is a positive integer greater than 1, and is the input value( It indicates performing an exponential operation on ). Here, in a softmax method based on secure multi-party computing, the exponential operation increases the input value by the exponent, and the exponential operation and its accumulation very easily generate data that exceeds the data length that MPC can process, causing data overflow.

[0039] In this embodiment, in order to prevent data overflow caused by the result of the exponential operation being excessively large and to prevent data overflow during the process of performing normalization processing based on safe multi-party computing on the classification prediction value, a reduction process is first performed on the classification prediction value to reduce the classification prediction value to a smaller number, so that even if the exponential operation is performed, the data range that MPC can process is not exceeded.

[0040] For example, one setting value may be subtracted from the classification prediction value, and said setting value may be set and adjusted according to the data range that the MPC can process and the range of the classification prediction value in a specific application scenario, and the present embodiment is not specifically limited herein.

[0041] For example, the classification prediction values ​​can be divided by a single larger value; for instance, each classification prediction value is divided by the maximum value of all classification prediction values ​​to reduce the classification prediction values ​​to within the [0,1] range, and the result obtained by performing an exponential operation on the reduced classification prediction values ​​is still within the [0,1] range.

[0042] Step (S203), normalization processing is performed on the classification prediction values ​​after reduction processing to obtain the normalized result of the classification prediction values.

[0043] By performing normalization on the classification prediction values ​​after reduction processing, data overflow during the normalization process can be prevented, and accurate normalized classification prediction results can be obtained.

[0044] Step (S204), the parameters of the classification model are updated based on the normalization results of the classification prediction values.

[0045] After obtaining the normalization results of the classification prediction values, the normalization results of the classification prediction values ​​can be used as the predicted probability values ​​of the corresponding classifications. Based on the predicted probability values ​​of each classification, the classification results of the classification model are finalized. Based on the classification results, the parameters of the current classification model are updated to implement training for the classification model and improve the training accuracy of the classification model.

[0046] An embodiment of the present application, in a model parameter adjustment process based on secure multi-party computing, performs classification processing on input data using a classification model obtained by training based on secure multi-party computing to obtain a classification prediction value of the classification model; performs downscaling processing on the classification prediction value and performs normalization processing on the classification prediction value after downscaling to obtain a normalization result of the classification prediction value; updates the parameters of the classification model based on the normalization result of the classification prediction value and reduces the input value of the normalization method by downscaling the classification prediction value, thereby preventing data overflow from occurring during the normalization process and improving the training precision of the classification model.

[0047] FIG. 3 is a flowchart of a method for adjusting model parameters according to a second embodiment of the present application. Based on the first embodiment described above, this embodiment determines the maximum value among the classification prediction values ​​and divides each classification prediction value by the maximum value to perform a reduction process on the classification prediction values. Furthermore, dividing each classification prediction value by the maximum value can be calculated using a long division method based on secure multi-party computing to obtain the quotient of each classification prediction value and the maximum value. As illustrated in FIG. 3, the specific steps of the method are as follows.

[0048] Step (S301), using a classification model obtained by training based on secure multi-party computing, classification processing is performed on the input data to obtain the classification prediction value of the classification model.

[0049] Here, the classification prediction value of the classification model includes the prediction value corresponding to each classification of the classification model.

[0050] In the training process of a classification model based on secure multi-party computing, classification processing is performed on input data using the current classification model obtained by training based on secure multi-party computing, and the classification prediction value of the current classification model is obtained.

[0051] In this embodiment, the input data may be data from a training set or a test set, and may be various types of data such as images, text, and voice, and this embodiment is not specifically limited herein.

[0052] In data processing based on secure multi-party computing, ciphertext is generally represented using only integers, while floating-point data types are represented as fixed-point numbers and scaled up to integers. Consequently, the actual range of data that can be represented in secure multi-party computing is limited. In the softmax method based on secure multi-party computing, the input for the exponential operation is the classification prediction value of the classification model; when the input is too large, the result of the exponential operation very easily exceeds the data length that MPC can process, causing a data overflow.

[0053] In this embodiment, through the following steps (S302-S303), each classification prediction value is divided by the maximum value of all classification prediction values ​​to reduce the classification prediction value to within the range [0,1], and since the result obtained by performing an exponential operation on the classification prediction value after reduction is still within the range [0,1], data overflow can be prevented during the normalization process.

[0054] Currently, secure multi-party computing protocols can basically implement all basic arithmetic operations such as addition, subtraction, multiplication, and comparison, as well as Boolean operations such as AND, OR, NOT, and XOR. For example, common secure multi-party computing protocols include protocols such as ABY and ABY3, and the arithmetic and Boolean operations related to this application are all operations based on secure multi-party computing protocols, and can be implemented using Boolean circuits or Garbled circuits based on the basic arithmetic and Boolean operations among secure multi-party computing protocols.

[0055] Step (S302), determine the maximum value among the classification prediction values.

[0056] In this embodiment, calculating the maximum value among the two classification prediction values ​​can be expressed as c = max{a, b}, and c = (a > b) * a + (1 - (a > b))*bThis can be implemented by calculating, where a and b represent the two classification prediction values, and c represents the maximum value between a and b.

[0057] Furthermore, determining the maximum value among multiple classification prediction values ​​can be implemented based on a method for calculating the maximum value among two classification prediction values.

[0058] For example, dog classification prediction values Regarding, The maximum value among them Calculate, class The maximum value among them Calculate, and The maximum value among them Calculate, class The maximum value among them Inferring in this way until calculating, the final maximum value go Dog classification prediction value It is the maximum value among them.

[0059] For example, dog classification prediction values Regarding, Divide the predicted classification values ​​into groups of two, calculate the maximum value among the two predicted classification values ​​in each group, and then use a similar method to calculate the maximum value among the maximum values ​​in each group.

[0060] In addition, determining the maximum value among multiple classification prediction values ​​based on a method for calculating the maximum value among two classification prediction values ​​may also be implemented using other methods, and redundant explanation is omitted here in this embodiment.

[0061] Step (S303), each classification prediction value is divided by the maximum value, and the quotient of each classification prediction value obtained and the maximum value is the classification prediction value after reduction processing.

[0062] After determining the maximum value among all classification prediction values, each classification prediction value is divided by the maximum value to perform a reduction process on each classification prediction value, and the quotient of each classification prediction value and the maximum value obtained is the classification prediction value after the reduction process. At this time, the classification prediction value can be reduced to within the range (0,1), and since the result obtained by exponentially operating the classification prediction value after reduction is still within the range (0,1), it is possible to prevent data overflow problems from occurring during the normalization process due to the input value being too large.

[0063] Step (S304), obtain the parameter threshold.

[0064] exponential function Defined as, and here, is the parameter of the exponential function, is a variable that performs exponential operations, and Is It is the exponential function value.

[0065] In the prior art, to prevent data overflow caused by the result of exponential operations being too large, approximate calculations are generally performed using the Rectified Linear Unit (ReLU) function; however, when performing approximate calculations using the ReLU function, the error is larger.

[0066] In this embodiment, based on the definition of an exponential function, the parameters of the exponential function are set to a larger parameter threshold to approximate infinity, the exponential operation is converted into a multiplication operation, and the exponential operation approximated by the multiplication operation based on multi-party secure computing is implemented, so that the error occurring when the parameter threshold is larger is very small.

[0067] Here, parameter thresholds can be set and adjusted according to actual application scenarios, and this embodiment is not specifically limited thereto.

[0068] Step (S305), function parameter among exponential functions ( Set the value of ) as the parameter threshold and calculate the exponential function value of the classification prediction value after each reduction process.

[0069] After obtaining the parameter threshold of the exponential function parameter, the function parameter among the exponential functions ( The value of ) is set as the corresponding parameter threshold, and the exponential function value of the classification prediction value after each reduction process is calculated. At this time, the exponential operation can be converted to a multiplication operation, and an approximate exponential operation is implemented through a multiplication operation based on multi-party safe computing, so that when the parameter threshold is larger, the error that occurs is very small and no data overflow occurs.

[0070] Specifically, for each classification prediction value after reduction processing, the exponential function value of the classification prediction value after reduction processing is calculated and obtained approximately using Formula 1 below.

[0071] Formula 1

[0072] Here, is a natural constant, and is the classification prediction value after downsizing, and is the exponential function value of the classification prediction after downsizing, and is the parameter threshold, and is the reciprocal of the parameter threshold, and M is a constant.

[0073] Step (S306), based on each exponential function value and the sum of all exponential function values, the normalization result of the classification prediction value after each reduction process is determined.

[0074] In this step, the exponential function value of the classification prediction after each reduction process is divided by the sum of all exponential function values ​​to obtain the normalization result of the classification prediction after each reduction process—in other words, the normalization result corresponding to the classification prediction.

[0075] Combining the steps (S302-S306) described above, the present embodiment is, dog classification prediction values For this, the normalization result of each classification prediction can be calculated using Formula 2 below.

[0076] Formula 2

[0077] Here, is the classification prediction value ( Displays the normalization result corresponding to ), and Is It is the maximum value among them, Is It is the exponential function value.

[0078] Step (S307), the parameters of the classification model are updated based on the normalization results of the classification prediction values.

[0079] In secure multi-party computing, after obtaining the normalization results of the classification predictions, the parameters of the classification model can be updated based on the normalization results of the classification predictions to train the classification model.

[0080] Specifically, based on the normalization results of the classification prediction values, the classification results of the classification model are determined; and based on the classification results, the parameters of the classification model are updated, thereby enabling the training of the classification model and improving the precision of the classification model.

[0081] For example, after obtaining the normalization result of the classification prediction value, the normalization result of the classification prediction value is used as the predicted probability value of the corresponding classification, and the classification result of the classification model can be determined based on the predicted probability value of each classification.

[0082] In this embodiment, through a definition based on an exponential function, the function parameter among the exponential functions ( By setting the value of ) to the corresponding parameter threshold, converting the exponential operation to a multiplication operation, and implementing an approximate exponential operation through a multiplication operation based on multi-party secure computing, when the parameter threshold is larger, the error is very small and no data overflow occurs.

[0083] In data processing based on secure multi-party computing, division is typically calculated using the Newtonian approximation method; this method requires that the initial values ​​of the set algorithm parameters be associated with the size range of the input divisor and divider, and imposes certain limitations on data length. However, in large-scale data machine learning, the range of intermediate results is difficult to predict. When intermediate results are used as input data for division, it is very difficult to perform accurate calculations by setting appropriate initial values, and overflow easily occurs during the division operation.

[0084] In one optional implementation method, as illustrated in FIG. 4, the division operation based on secure multi-party computing related to the present application is implemented by a long division method based on secure multi-party computing, and specifically can be implemented through the following steps.

[0085] Step (S401), the first variable and the second variable are determined based on the number of divisions and the number of divisions, wherein the number of significant digits of the second variable is greater than or equal to the number of significant digits of the first variable.

[0086] Figure 5 illustrates one implementation method of the plaintext long division method. In the plaintext long division method, before performing the iteration (processing process corresponding to numbers 3-12 in Figure 5), a process as indicated by number 2 is performed to align the greatest significant digits of the divider and the divisor, and after alignment, the greatest significant digits of the divisor are required to be greater than or equal to the greatest significant digits of the divider. However, in division based on secure multi-party operations, since both the divider and the divisor are ciphertexts, the greatest significant digits of the divider and the divisor cannot be directly aligned.

[0087] Optionally, the number of significant digits of the divisor is increased to obtain a second variable; the divisor can be directly used as the first variable. At this time, by increasing the number of significant digits of the divisor so that the number of significant digits of the divisor is greater than or equal to the number of significant digits of the divisor, the alignment operation between the divisor and the divisor of the ciphertext is implemented.

[0088] For example, the second variable is obtained by increasing the significant digits of the divisor, and the fact that the significant digits of the second variable are greater than or equal to the significant digits of the first variable can be specifically implemented in the following manner.

[0089] Based on the maximum significant digits of the dividend, the divisor is shifted left by that amount to obtain a second variable; in this case, sorting of the dividend and divisor can be implemented without the need to determine the significant digits of the dividend and divisor. If the difference between the significant digits of the dividend and divisor is known, the divisor is shifted left by that amount to ensure that the divisor's significant digits are not smaller than the dividend's.

[0090] In this embodiment, all shift operations performed on the data are shift operations based on secure multi-party computing, and specifically can be implemented using Boolean circuits or decomposition circuits.

[0091] In addition, the quotients of the divisor and the dividing number can be initialized to 0. In this embodiment, the divisor, the dividing number, and the quotient are all represented in binary.

[0092] After sorting the number of divisions and the number of divisors, the processing of the following steps (S402-S404) is repeated based on the first variable corresponding to the number of divisions and the second variable corresponding to the number of divisors, and in each iteration, the value of one bit of the quotient is determined and added to the end of the quotient; the number of iterations is determined according to the length of the significant digits of the quotient, and through this method, a long division method based on secure multi-party computing for the number of divisions and the number of divisors of the ciphertext can be obtained, and data overflow caused by division operations based on secure multi-party computing can be prevented.

[0093] Step (S402), compare whether the first variable is greater than or equal to the second variable.

[0094] In this embodiment, the comparison operation for comparing the sizes of two data is a comparison operation based on secure multi-party computing, and specifically can be implemented using a Boolean circuit or a composite circuit.

[0095] Step (S403), based on the comparison result, determine the value of the current bit of the quotient and update the first variable and the second variable.

[0096] At that step, based on the comparison result, the value of the current bit of the quotient and the update of the first and second variables must be implemented as follows: specifically, if the first variable is greater than or equal to the second variable, add 1 to the end of the quotient, subtract the second variable from the first variable, and shift the second variable one position to the right; if the first variable is less than the second variable, add 0 to the end of the quotient, shift the first variable one position to the left, and do not change the second variable.

[0097] In this embodiment, determining the value of the current bit of the quotient based on the comparison result can be implemented using the following method.

[0098] Based on the result of the comparison, add 1 or 0 to the end of the quotient, where if the first variable is greater than or equal to the second variable, add 1 to the end of the quotient, and if the first variable is less than the second variable, add 0 to the end of the quotient. Based on secure multi-party computing, based on the result of the comparison, add 1 or 0 to the end of the quotient.

[0099] The comparison result during the step is a comparison operation obtained based on secure multi-party computing, and the comparison result is a ciphertext, so the comparison result of the plaintext cannot be known, and in a situation where the comparison result of the plaintext cannot be known, there is a technical difficulty in determining the value of the current bit of the quotient and updating the first variable and the second variable based on the comparison result.

[0100] In one implementation method, adding 1 or 0 to the end of the quotient based on the comparison result is specifically implemented using the following method: that is, adding 0 to the end of the quotient; and performing an XOR operation with the comparison result on the quotient with 0 added to the end. At this time, based on the ciphertext comparison result, in a situation where the plaintext comparison result is unknown, 1 or 0 is directly added to the end of the quotient, and if the first variable is greater than or equal to the second variable, 1 is added to the end of the quotient, and if the first variable is less than the second variable, 0 is added to the end of the quotient.

[0101] For example, if the first variable is greater than or equal to the second variable, the comparison result is 1 (ciphertext), and if the first variable is less than the second variable, the comparison result is 0 (ciphertext), 0 (ciphertext) is added to the end of the quotient, and the quotient after adding 0 is XORed with the comparison result, and if the current comparison result is 1 (ciphertext), the value after XOR is 1, and the 0 (ciphertext) added to the end of the quotient is changed to 1 (ciphertext), that is, if the first variable is greater than or equal to the second variable, 1 is added to the end of the quotient; if the current comparison result is 0 (ciphertext), the value after XOR is 1, and the 0 (ciphertext) added to the end of the quotient is still 0 (ciphertext), that is, if the first variable is less than the second variable, 0 is added to the end of the quotient.

[0102] In one implementation method, updating the first variable based on the comparison result can be implemented using the following method.

[0103] A first update process is performed on a first variable to obtain a first update result, wherein the first update process is an update process for the first variable when the first variable is greater than or equal to the second variable; a second update process is performed on the first variable to obtain a second update result, wherein the second update process is an update process for the first variable when the first variable is smaller than the second variable; a first multiplication of the comparison result and the first update result, and a second multiplication of the result of the NOT operation of the comparison result and the second update result are calculated; and the sum of the first multiplication and the second multiplication is set as the first variable after the update. At this time, based on the ciphertext comparison result, the first variable can be directly updated in cases where the plaintext comparison result is unknown.

[0104] Here, the first update process is an update process for the first variable when the first variable is greater than or equal to the second variable, and specifically, the first update process obtains the first update result by subtracting the second variable from the first variable through a subtraction operation based on secure multi-party computing.

[0105] The second update process is an update process for the first variable when the first variable is smaller than the second variable, and specifically, the second update process obtains the second update result by shifting the first variable one digit to the left through a shift operation based on secure multi-party computing.

[0106] For example, if the first variable is greater than or equal to the second variable, the comparison result is 1 (ciphertext) and the NOT operation result of the comparison result is 0 (ciphertext); if the first variable is less than the second variable, the comparison result is 0 (ciphertext) and the NOT operation result of the comparison result is 1 (ciphertext). Calculate the first multiplication of the comparison result and the first update result, and the second multiplication of the NOT operation result of the comparison result and the second update result; when the first variable is greater than or equal to the second variable, the first multiplication is equal to the first update result and the second multiplication is 0 (ciphertext), and calculate the sum of the first multiplication and the second multiplication so that the first variable after the confirmed update is equal to the first update result; When the first variable is smaller than the second variable, the first multiplication is 0 (ciphertext), the second multiplication is equal to the second update result, and the first variable after the update can be determined as the second update result by calculating the sum of the first multiplication and the second multiplication.

[0107] In one implementation method, updating the second variable based on the comparison result can be implemented using the following method.

[0108] A third update process is performed on the second variable to obtain the third update result, wherein the third update process is an update process for the second variable when the first variable is greater than or equal to the second variable; the third multiplication of the comparison result and the third update result, and the fourth multiplication of the comparison result and the second variable are calculated; and the sum of the third multiplication and the fourth multiplication is set as the second variable after the update. At this time, based on the ciphertext comparison result, the second variable can be directly updated in a situation where the plaintext comparison result is unknown.

[0109] Here, the third update process is the update of the second variable when the first variable is greater than or equal to the second variable; specifically, the third update process obtains the result of the third update by shifting the second variable one position to the right through a shift operation based on secure multi-party computing. If the first variable is smaller than the second variable, the second variable does not change.

[0110] For example, if the first variable is greater than or equal to the second variable, the comparison result is 1 (ciphertext) and the NOT operation result of the comparison result is 0 (ciphertext); if the first variable is less than the second variable, the comparison result is 0 (ciphertext) and the NOT operation result of the comparison result is 1 (ciphertext). Calculate the third multiplication of the comparison result and the third update result, and the fourth multiplication of the NOT operation result of the comparison result and the second variable; when the first variable is greater than or equal to the second variable, the first multiplication is equal to the third update result and the second multiplication is 0 (ciphertext), and calculate the sum of the third and fourth multiplications, and the second variable after the confirmed update is equal to the third update result; When the first variable is smaller than the second variable, the third multiplication is 0 (ciphertext), the fourth multiplication is equal to the second variable, and by calculating the sum of the third and fourth multiplications, it can be determined that the second variable after the update remains the second variable and does not change.

[0111] Step (S404), determine whether the number of digits of the quotient is equal to the length of the pre-set significant digits.

[0112] After each iteration process is finished, determine whether the number of iterations confirmed in the currently obtained quotient is equal to the length of the significant digits of the pre-set quotient.

[0113] If the number of digits of the quotient is not the same as the preset number of significant digits, the continuation step (S402) is performed, and the next iteration process is performed based on the first variable and the second variable after the update.

[0114] If the number of digits of the quotient is equal to the length of the significant digits, the currently obtained quotient is the quotient of the number of divisions and the number of divisions, and step (S405) is performed.

[0115] Step (S405), the currently obtained share is made into the share of the number to be divided and the number to be divided.

[0116] For example, in the above-described step (S303), dividing each classification prediction value by the maximum value among all classification prediction values ​​can be obtained by calculating the quotient of each classification prediction value and the maximum value using the method shown in FIG. 4, where each classification prediction value is the divisor and the maximum value is the divisor.

[0117] For example, an arbitrary classification prediction value is set as the division number, and the maximum value among the classification prediction values ​​is set as the divisor, satisfying the condition that the significant digits of the divisor are greater than or equal to the significant digits of the division number, and that alignment of the division number and the divisor number is not required, and based on a first variable corresponding to the division number and a second variable corresponding to the divisor number, the processing of the steps described above (S402-S404) is repeated until the number of repetitions is greater than or equal to the length of the significant digits of the quotient, and in each repetition process, the value of one bit of the quotient is determined and added to the end of the quotient; thereby obtaining the quotient of the classification prediction value and the maximum value.

[0118] Through this method, it is possible to implement long division operations based on secure multi-party computing for the predicted classification value of the ciphertext and the maximum value among each predicted classification value, and to effectively prevent data overflow caused by division operations based on secure multi-party computing.

[0119] The embodiments of the present application implement the calculation of the quotient of each classification prediction value and the maximum value using a Boolean circuit or a complex circuit through a long division method based on secure multi-party computing, thereby preventing data overflow and achieving high precision.

[0120] FIG. 6 is a drawing showing a model parameter adjustment device according to a third embodiment of the present application. The model parameter adjustment device according to an embodiment of the present application can perform a processing process according to an embodiment of a model parameter adjustment method. As shown in FIG. 6, the model parameter adjustment device (60) includes a classification processing module (601), a reduction processing module (602), a normalization processing module (603), and a parameter update module (604).

[0121] Specifically, the classification processing module (601) performs classification processing on input data using a classification model obtained by training based on secure multi-party computing, and obtains a classification prediction value of the classification model.

[0122] The reduction processing module (602) performs reduction processing on the classification prediction value.

[0123] The normalization processing module (603) performs normalization processing on the classification prediction value after reduction processing to obtain the normalization result of the classification prediction value.

[0124] The parameter update module (604) updates the parameters of the classification model based on the normalization result of the classification prediction value.

[0125] The device according to the embodiment of the present application can be used to execute the method embodiment according to the first embodiment described above, and specific functions are omitted from redundant description herein.

[0126] An embodiment of the present application performs a classification processing on input data using a classification model obtained by training based on secure multi-party computing in a model parameter adjustment process based on secure multi-party computing to obtain a classification prediction value of the classification model; performs a reduction processing on the classification prediction value and performs a normalization processing on the classification prediction value after the reduction processing to obtain a normalization result of the classification prediction value; updates the parameters of the classification model based on the normalization result of the classification prediction value and reduces the input value of the normalization method by reducing the classification prediction value, thereby preventing data overflow from occurring in the normalization process and improving the training precision of the classification model.

[0127] FIG. 7 is a drawing showing a model parameter adjustment device according to the fourth embodiment of the present application. Based on the third embodiment described above, the present embodiment, as illustrated in FIG. 7, includes a model parameter adjustment device (70) comprising a classification processing module (701), a reduction processing module (702), a normalization processing module (703), and a parameter update module (704).

[0128] Specifically, the classification processing module (701) performs classification processing on input data using a classification model obtained by training based on secure multi-party computing, and obtains a classification prediction value of the classification model.

[0129] The reduction processing module (702) performs reduction processing on the classification prediction value.

[0130] The normalization processing module (703) performs normalization processing on the classification prediction value after reduction processing to obtain the normalization result of the classification prediction value.

[0131] The parameter update module (704) updates the parameters of the classification model based on the normalization result of the classification prediction value.

[0132] In one selectable implementation method, the reduction processing module (702) is,

[0133] A maximum value determination unit (7021) that determines the maximum value among the classification prediction values;

[0134] It includes a division operation unit (7022) that divides each classification prediction value by the maximum value.

[0135] In one selectable implementation method, the division operation unit (7022) is,

[0136] Each classification prediction value and maximum value shall be the first variable and the second variable, respectively, and the alignment subunit having a significant digit of the second variable greater than or equal to the significant digit of the first variable;

[0137] Compare whether the first variable is greater than or equal to the second variable; based on the comparison result, determine the value of the current bit of the quotient and update the first variable and the second variable; compare the first variable and the second variable after the update until the number of digits of the quotient reaches a preset length of significant digits, and determine the value of the next bit of the quotient based on the comparison result, wherein the obtained quotient includes an iterative subunit which is the result of dividing each classification prediction value by the maximum value.

[0138] In one selectable implementation method, the iterative subunit also,

[0139] Based on the result of the comparison, add 1 or 0 to the end of the quotient, where if the first variable is greater than or equal to the second variable, add 1 to the end of the quotient, and if the first variable is less than the second variable, add 0 to the end of the quotient.

[0140] In one selectable implementation method, the iterative subunit also,

[0141] Add 0 to the end of the quotient; perform XOR operation with the comparison result of the quotient after adding 0 to the end.

[0142] In one selectable implementation method, the iterative subunit also,

[0143] A first update process is performed on a first variable to obtain a first update result, wherein the first update process is an update process for the first variable when the first variable is greater than or equal to the second variable; a second update process is performed on the first variable to obtain a second update result, wherein the second update process is an update process for the first variable when the first variable is smaller than the second variable; a first multiplication of the comparison result and the first update result, and a second multiplication of the result of the NOT operation of the comparison result and the second update result are calculated; and the sum of the first multiplication and the second multiplication is set as the first variable after the update.

[0144] In one selectable implementation method, the iterative subunit also,

[0145] A third update process is performed on the second variable to obtain the third update result, and the third update process is an update process for the second variable if the first variable is greater than or equal to the second variable; the third multiplication of the comparison result and the third update result, and the fourth multiplication of the comparison result and the second variable are calculated; and the sum of the third multiplication and the fourth multiplication is set as the second variable after the update.

[0146] In one selectable implementation method, the normalization processing module (703) is,

[0147] A parameter threshold acquisition unit (7031) for acquiring a parameter threshold;

[0148] An exponential operation unit (7032) that calculates the exponential function value of the classification prediction value after each reduction process by setting the value of the function parameter (n) in the exponential function as a parameter threshold;

[0149] It includes a normalization unit (7033) that determines the normalization result of the classification prediction value after each reduction process based on each exponential function value and the sum of all exponential function values.

[0150] In one selectable implementation method, the exponent operation unit (7032) also,

[0151] For each classification prediction value after reduction processing, the exponential function value of the classification prediction value after reduction processing is obtained by calculating using the formula below.

[0152] , here, is a natural constant, and is the classification prediction value after downsizing, and is the exponential function value of the classification prediction after downsizing, and is the parameter threshold, and is the reciprocal of the parameter threshold, and M is a constant.

[0153] The device according to the embodiment of the present application can be used to execute the method embodiment according to the second embodiment described above, and specific functions are omitted from redundant description herein.

[0154] In this embodiment, based on the definition of an exponential function, the function parameter among the exponential functions ( The value of ) is set as the corresponding parameter threshold, exponential operations are converted to multiplication operations, and approximate exponential operations are implemented through multiplication operations based on multi-party safe computing. When the parameter threshold is greater, the error is very small and no data overflow occurs. Furthermore, by implementing the calculation of the quotient between each classification prediction value and the maximum value using a Boolean circuit or a complex circuit through long division based on safe multi-party computing, no data overflow occurs and high precision is achieved.

[0155] In the technical solution of the present application, the processing of the personal information of the relevant user, including collection, storage, use, processing, transmission, provision, and disclosure, all comply with the provisions of relevant laws and regulations and do not violate public order and customs. According to an embodiment of the present application, the present application further provides an electronic device and a readable storage medium.

[0156] According to an embodiment of the present application, the present application further provides a computer program product, wherein the program product includes a computer program, the computer program is stored in a readable storage medium, and at least one processor of an electronic device can read the computer program from the readable storage medium, and the electronic device executes the method according to any embodiment described above by the at least one processor executing the computer program. According to an embodiment of the present application, the present application provides a computer program stored in a computer-readable storage medium, and when the computer program is executed by a processor, the method according to any embodiment described above is implemented.

[0157] FIG. 8 illustrates an exemplary block diagram of an exemplary electronic device (800) capable of carrying out an embodiment of the present application. The electronic device is intended to represent various types of digital computers, e.g., laptop computers, desktop computers, workstations, personal information terminals, servers, blade servers, large computers, and other suitable computers. The electronic device may further represent various types of mobile devices, e.g., personal information terminals, cell phones, smartphones, wearable devices, and other similar computing devices. The components, their connections and relationships, and their functions disclosed herein are merely exemplary and are not intended to limit the implementation of the present application disclosed herein and / or required thereof.

[0158] As illustrated in FIG. 8, the electronic device (800) includes a computing unit (801) and can perform various suitable operations and processing based on a computer program stored in read-only memory (802, ROM) or a computer program loaded into random access memory (803, RAM) from a storage unit (808). Various programs and data required for the operation of the electronic device (800) may be further stored in the RAM (803). The computing unit (801), ROM (802), and RAM (803) are connected to each other via a bus (804). An input / output (I / O) interface (805) is also connected to the bus (804).

[0159] A plurality of components of the electronic device (800) are connected to an I / O interface (805) and include an input unit (806), such as a keyboard, mouse, etc.; an output unit (807), such as various types of displays, speakers, etc.; a storage unit (808), such as a magnetic disk, optical disk, etc.; and a communication unit (809), such as a network card, modem, wireless communication transceiver, etc. The communication unit (809) allows the electronic device (800) to exchange information / data with other devices through a computer network such as the Internet and / or various electronic communication networks.

[0160] The computing unit (801) may be various general-purpose and / or dedicated processing modules equipped with processing and computation capabilities. Some examples of the computing unit (801) include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, a computing unit executing various machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit (801) performs each of the methods and processes described above, such as a model parameter adjustment method. For example, in some embodiments, the model parameter adjustment method may be implemented as a computer software program and explicitly stored on a machine-readable medium such as a storage unit (808). In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device (800) via a ROM (802) and / or a communication unit (809). When the computer program is loaded into RAM (803) and executed by the computing unit (801), one or more steps of the model parameter adjustment method described above may be performed. Optionally, in other embodiments, the computing unit (801) may be configured to perform a model parameter adjustment method through any other suitable method (e.g., via firmware).

[0161] Various embodiments of the systems and technologies described herein may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable arrays (FPGAs), dedicated integrated circuits (ASICs), dedicated standard products (ASSPs), system-on-chip systems (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include being implemented in one or more computer programs, said one or more computer programs may be executed and / or interpreted on a programmable system comprising at least one programmable processor, said programmable processor may be a dedicated or general-purpose programmable processor, may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to said storage system, said at least one input device, and said at least one output device.

[0162] Program code for carrying out the method of the present application may be written in any combination of one or more programming languages. Such program code may be provided to a processor or controller of a general-purpose computer, a dedicated computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations defined in the flowchart and / or block diagram are carried out. The program code may be executed entirely on a machine or partially on a machine, or as an independent software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0163] In the context of this application, a machine-readable medium may be a tangible medium and may contain or store a program used by or in combination with a command execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or apparatus, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), optical fibers, portable compact read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any combination of the foregoing.

[0164] To provide interaction with a user, the system and technology described herein may be implemented on a computer, said computer having a display device for displaying information to a user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and a directional device (e.g., a mouse or a trackball), said user may provide input to the computer through said keyboard and said directional device. Other types of devices may also provide interaction with a user. For example, feedback provided to a user may be any form of sensing feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from a user may be received through any form (sound input, voice input, or tactile input).

[0165] The systems and technologies described herein may be implemented in a computing system including a background component (e.g., as a data server), or a computing system including an intermediate component (e.g., an application server), or a computing system including a front-end component (e.g., a user computer equipped with a graphical user interface or an internet browser, through which the user may interact with embodiments of the systems and technologies described herein), or any combination of such background components, intermediate components, or front-end components. Components of the systems may be connected to each other through digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0166] A computer system may include clients and servers. Clients and servers are generally located far apart from each other and typically interact through a communication network. The relationship between the client and the server is established through computer programs that run on corresponding computers and have a client-server relationship with each other. The server can be a cloud server, also referred to as a cloud computing server or cloud host, and serves as a host product within a cloud computing service system that addresses the shortcomings of existing physical hosts and Virtual Private Server (VPS) services, such as significant management difficulties and limited service scalability. The server may be a server in a distributed system or a server combined with blockchain.

[0167] It should be understood that steps may be rearranged, added, or deleted using the various forms of processes described above. For example, each step described in this application may be performed in parallel, sequentially, or in a different order, and is not limited thereto as long as the technical solution according to this application can obtain the desired result.

[0168] The specific embodiments described above are not a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions may be made based on design needs and other factors. All modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall fall within the scope of protection of this application.

Claims

Claim 1 A method for adjusting model parameters performed by a model parameter adjustment device comprises: a step of performing classification processing on input data using a classification model obtained by training based on secure multi-party computing to obtain a classification prediction value of said classification model; a step of performing reduction processing on said classification prediction value; a step of performing normalization processing on the classification prediction value after reduction processing to obtain a normalized result of said classification prediction value; and a step of updating the parameters of said classification model based on the normalized result of said classification prediction value; wherein the step of performing reduction processing on said classification prediction value includes a step of determining a maximum value among said classification prediction values; and a step of dividing each of said classification prediction values ​​by said maximum value; wherein each of said classification prediction value and said maximum value is respectively a first variable and a second variable, and the number of significant digits of said second variable is greater than or equal to the number of significant digits of said first variable; a step of comparing whether said first variable is greater than or equal to said second variable; and, based on the comparison result, determining the value of the current bit of the quotient, and said first variable and A method for adjusting model parameters comprising: a step of updating a second variable; a step of comparing the first variable and the second variable after the update until the number of digits of the quotient reaches a preset length of significant digits, determining the value of the next bit of the quotient based on the comparison result, wherein the obtained quotient is the result of dividing each of the classification prediction values ​​by the maximum value. Claim 2 delete Claim 3 delete Claim 4 A method for adjusting model parameters according to claim 1, wherein the step of determining the value of the current bit of the quotient based on the comparison result comprises adding 1 or 0 to the end of the quotient based on the comparison result, wherein if the first variable is greater than or equal to the second variable, 1 is added to the end of the quotient, and if the first variable is smaller than the second variable, 0 is added to the end of the quotient. Claim 5 A method for adjusting model parameters according to claim 4, wherein, based on the comparison result, 1 or 0 is added to the end of the quotient, wherein if the first variable is greater than or equal to the second variable, 1 is added to the end of the quotient, and if the first variable is smaller than the second variable, 0 is added to the end of the quotient, the step of adding 0 to the end of the quotient; and the step of performing an XOR operation with the comparison result on the quotient after adding 0 to the end. Claim 6 A method for adjusting model parameters according to claim 1, wherein the step of updating the first variable based on the comparison result comprises: performing a first update process on the first variable to obtain a first update result, wherein the first update process is an update process for the first variable when the first variable is greater than or equal to the second variable; performing a second update process on the first variable to obtain a second update result, wherein the second update process is an update process for the first variable when the first variable is smaller than the second variable; calculating a first multiplication of the comparison result and the first update result, and a second multiplication of the NOT operation result of the comparison result and the second update result; and making the sum of the first multiplication and the second multiplication the first variable after the update. Claim 7 A method for adjusting model parameters according to claim 1, wherein the step of updating the second variable based on the comparison result comprises: performing a third update process on the second variable to obtain a third update result, wherein the third update process is an update process for the second variable when the first variable is greater than or equal to the second variable; calculating a third multiplication of the comparison result and the third update result, and a fourth multiplication of the result of a NOT operation of the comparison result and the second variable; and making the sum of the third multiplication and the fourth multiplication the second variable after the update. Claim 8 In claim 1, the step of performing normalization processing on the classification prediction value after reduction processing to obtain the normalization result of the classification prediction value comprises: a step of obtaining a parameter threshold value; a function parameter among exponential functions ( A method for adjusting model parameters comprising: a step of setting the value of ) as the parameter threshold and calculating the exponential function value of each classification prediction value after the reduction process; and a step of determining the normalization result of each classification prediction value after the reduction process based on each of the exponential function values ​​and the sum of all the exponential function values. Claim 9 In paragraph 8, the function parameter in the exponential function ( The step of setting the value of ) as the parameter threshold and calculating the exponential function value of each classification prediction value after the reduction process includes, for each classification prediction value after the reduction process, calculating using the following formula to obtain the exponential function value of the classification prediction value after the reduction process. , here, is a natural constant, and is the predicted classification value after the above reduction processing, and is the exponential function value of the classification prediction value after the above reduction processing, and is the above parameter threshold, and A model parameter tuning method in which is the reciprocal of the above parameter threshold and M is a constant. Claim 10 A method for adjusting model parameters according to claim 1, wherein the step of updating the parameters of the classification model based on the normalization result of the classification prediction value comprises: the step of determining the classification result of the classification model based on the normalization result of the classification prediction value; and the step of updating the parameters of the classification model based on the classification result. Claim 11 A classification processing module that performs classification processing on input data using a classification model obtained by training based on secure multi-party computing to obtain a classification prediction value of said classification model; a reduction processing module that performs reduction processing on said classification prediction value; a normalization processing module that performs normalization processing on the classification prediction value after reduction processing to obtain a normalized result of said classification prediction value; and a parameter update module that updates the parameters of said classification model based on the normalized result of said classification prediction value; said reduction processing module includes a maximum value determination unit that determines the maximum value among said classification prediction values; and a division operation unit that divides each of said classification prediction values ​​by said maximum value; said division operation unit includes an alignment sub-unit in which each of said classification prediction value and said maximum value is a first variable and a second variable, respectively, and the number of significant digits of said second variable is greater than or equal to the number of significant digits of said first variable; and compares whether said first variable is greater than or equal to said second variable; and based on the comparison result, determines the value of the current bit of the quotient and updates said first variable and said second variable; A model parameter adjustment device comprising an iterative subunit that compares a first variable and a second variable after an update until the number of digits of the above quotient reaches a preset length of significant digits, determines the value of the next bit of the above quotient based on the comparison result, and wherein the obtained quotient is the result of dividing each above-mentioned classification prediction value by the above-mentioned maximum value. Claim 12 delete Claim 13 delete Claim 14 A model parameter adjustment device according to claim 11, wherein the iterative subunit also adds 1 or 0 to the end of the quotient based on the comparison result, wherein if the first variable is greater than or equal to the second variable, 1 is added to the end of the quotient, and if the first variable is smaller than the second variable, 0 is added to the end of the quotient. Claim 15 In paragraph 14, the above-mentioned iteration subunit also adds a zero to the end of the above-mentioned quotient; and a model parameter adjustment device that performs an XOR operation with the comparison result on the quotient after adding a zero to the end. Claim 16 In claim 11, the iterative sub-unit also performs a first update process on the first variable to obtain a first update result, wherein the first update process is an update process for the first variable when the first variable is greater than or equal to the second variable; performs a second update process on the first variable to obtain a second update result, wherein the second update process is an update process for the first variable when the first variable is smaller than the second variable; calculates a first multiplication of the comparison result and the first update result, and a second multiplication of the NOT operation result of the comparison result and the second update result; and the sum of the first multiplication and the second multiplication is the first variable after the update, a model parameter adjustment device. Claim 17 In claim 11, the iterative sub-unit also performs a third update process on the second variable to obtain a third update result, wherein the third update process is an update process for the second variable when the first variable is greater than or equal to the second variable; calculates a third multiplication of the comparison result and the third update result, and a fourth multiplication of the NOT operation result of the comparison result and the second variable; and the sum of the third multiplication and the fourth multiplication is the second variable after the update, a model parameter adjustment device. Claim 18 In paragraph 11, the normalization processing module comprises: a parameter threshold acquisition unit for acquiring a parameter threshold; and a function parameter among an exponential function. A model parameter adjustment device comprising: an exponential operation unit that sets the value of as the parameter threshold and calculates the exponential function value of each classification prediction value after the reduction process; and a normalization unit that determines the normalization result of each classification prediction value after the reduction process based on each of the exponential function values ​​and the sum of all the exponential function values. Claim 19 In paragraph 18, the above-mentioned exponential calculation unit also obtains the exponential function value of the classification prediction value after the reduction processing by calculating for each of the above-mentioned classification prediction values ​​using the following formula; , here, is a natural constant, and is the predicted classification value after the above reduction processing, and is the exponential function value of the classification prediction value after the above reduction processing, and is the above parameter threshold, and A model parameter tuning device in which is the reciprocal of the above parameter threshold and M is a constant. Claim 20 A model parameter adjustment device according to any one of claims 11, 14 to 19, wherein the parameter update module comprises: a classification result determination unit that determines the classification result of the classification model based on the normalization result of the classification prediction value; and a parameter update unit that updates the parameters of the classification model based on the classification result. Claim 21 An electronic device comprising at least one processor; and a memory connected via communication with said at least one processor, wherein the memory stores a command that can be executed by said at least one processor, and said command is executed by said at least one processor so that said at least one processor can execute a model parameter adjustment method according to any one of claims 1, 4 to 10. Claim 22 A non-transient computer-readable storage medium in which computer instructions are stored, wherein the computer instructions cause the computer to execute a model parameter adjustment method according to any one of claims 1, 4 through 10. Claim 23 A computer program stored on a computer-readable storage medium, wherein, when an instruction in the computer program is executed by a processor, the computer program stored on the computer-readable storage medium implements a method for adjusting model parameters according to any one of claims 1, 4 to 10.

Images