A 5G machine room data security sharing method based on network security

Abstract

The application relates to the technical field of data security, in particular to a 5G machine room data security sharing method based on network security. The method is applied to a federal model system, and the method comprises the following steps: a participant carries out model training based on a model architecture file and original model parameter information issued by a center party, obtains initial model parameter information, carries out parameter iteration based on the initial model parameter information, until an iteration termination instruction issued by the center party is obtained, and noiseless parameter information obtained through the last iteration is determined as target parameter information; in the iteration process, the participant adjusts a noise intensity value of the previous iteration based on a convergence degree of corresponding model parameters in the federal model system in the current iteration, so as to obtain a noise intensity value of the current iteration, and carries out parameter noise encryption according to the noise intensity value, and meanwhile, model loss of the current iteration is reported. The application can improve the data protection effect of the 5G machine room in the data sharing process.

Description

Technical Field

[0001] This invention relates to the technical field of data security, and specifically to a method for secure data sharing in 5G data centers based on network security. Background Technology

[0002] As a data hub, the secure sharing of data in 5G data centers is crucial for unlocking data value. However, due to practical challenges such as data isolation, privacy protection, and resource constraints, traditional sharing models are difficult to implement. Federated learning provides a security foundation for such scenarios, achieving the security requirement of data not leaving the domain through local training and parameter sharing.

[0003] Applications have revealed that attackers can steal the parameters shared by the federated learning model to invert and obtain relevant features of the original business data, such as sample distribution and abnormal sample behavior. Therefore, during the parameter sharing process, it is necessary to adaptively encrypt the shared parameters to prevent attackers from inferring relevant information of the original business data through the parameters.

[0004] However, at present, most encryption mechanisms are static encryption schemes, which are still at risk of being stolen by attackers. This makes it difficult to ensure data security in 5G data centers during the data sharing process. Summary of the Invention

[0005] The purpose of this invention is to provide a 5G data security sharing method based on network security, which solves the technical problem of poor data protection effect in the data sharing process of 5G data centers.

[0006] In a first aspect, one embodiment of the present invention provides a method for secure data sharing in a 5G data center based on network security, applied to a target participant, wherein the target participant is one of multiple participants included in a federated model system, and the federated model system further includes a central party; the method includes: The model is trained based on the model architecture file and the original model parameter information issued by the central party to obtain the initial model parameter information; Based on the initial model parameter information, parameter iteration is performed until the iteration termination instruction issued by the center is obtained, and the noiseless parameter information obtained in the last iteration is determined as the target parameter information. Wherein, when n is greater than 1, the nth iteration in the parameter iteration includes: The noise intensity value of the (n-1)th iteration is adjusted based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration. The iterative characteristic values ​​are used to indicate the degree of convergence of the corresponding model parameters in the federated model system in the corresponding iteration. Noise is added to the noise-free parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration. The noise-free parameter information is obtained based on the joint parameter information issued by the central party in the corresponding iteration, and the joint parameter information is obtained based on the noise parameter information reported by multiple participating parties received by the central party in the corresponding iteration. The noise parameter information and parameter loss value of the nth iteration are reported to the center, and the joint parameter information or iteration termination instruction issued by the center are received accordingly. The parameter loss value indicates the model loss of the noise-free parameter information obtained in the corresponding iteration.

[0007] In some embodiments, when n is 1, the nth iteration in the parameter iteration includes: The preset initial noise intensity value is adjusted according to the sensitive feature value of each model parameter to obtain the noise intensity value of the nth iteration. The sensitive feature value is used to indicate the noise sensitivity of the corresponding model parameter in the business scenario corresponding to the target participant. Noise is added to the initial model parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration; The noise parameter information of the nth iteration is reported to the central authority, and the joint parameter information issued by the central authority is received accordingly.

[0008] In some embodiments, when n is 1, the step of obtaining the sensitive feature values ​​of each model parameter includes: Based on the numerical changes of each model parameter during the initial training phase, the numerical distribution curve of each model parameter is determined. The initial training phase refers to the process of training the model based on the model architecture file and the original model parameter information. Based on the stage center value of the initial training phase, the numerical distribution curves of each model parameter are divided into a first curve segment and a second curve segment. The first curve segment is the part of the corresponding numerical distribution curve that is before the stage center value, and the second curve segment is the part of the corresponding numerical distribution curve that is after the stage center value. The stage center value indicates the center time between the start and end times of the initial training phase. Analyze the area difference between the first and second curve segments of each model parameter to determine the first sensitive component value of each model parameter; Analyze the degree of numerical fluctuation in the second curve segment of each model parameter to determine the value of the second sensitive component of each model parameter; Based on the first and second sensitive component values ​​of each model parameter, the sensitive feature values ​​of each model parameter are determined.

[0009] In some embodiments, when n is greater than 1, the step of obtaining the iterative feature values ​​of each model parameter in the nth iteration includes: The receiving center sends out joint parameter information and federated feature values ​​of each model parameter in the nth iteration. The federated feature values ​​are used to indicate the degree of convergence of the corresponding model parameters in the nth iteration through the combined performance of multiple participants. Model training is performed based on the model architecture file and the joint parameter information of the nth iteration to obtain the noiseless parameter information of the nth iteration. Analyze the numerical changes of each model parameter during the model training phase in the nth iteration, determine the local feature value of each model parameter in the nth iteration, and use the local feature value to indicate the degree of convergence of the corresponding model parameter in the business scenario corresponding to the target participant in the nth iteration. Based on the federated and local feature values ​​of each model parameter in the nth iteration, determine the iterative feature values ​​of each model parameter in the nth iteration.

[0010] In some embodiments, the step of analyzing the numerical changes of each model parameter during the model training phase in the nth iteration and determining the local feature values ​​of each model parameter in the nth iteration includes: Based on the numerical changes of the target model parameter during the model training phase in the nth iteration, the amplitude rate of the target model parameter in the nth iteration is determined. The target model parameter is one of multiple model parameters. The amplitude rate is the ratio of the numerical amplitude of the corresponding model parameter during the model training phase in the corresponding iteration to the maximum parameter value. The numerical amplitude is the difference between the maximum parameter value and the minimum parameter value of the corresponding model parameter during the model training phase in the corresponding iteration. If the amplitude rate of the target model parameter in the nth iteration is less than the first threshold, the local feature value of the target model parameter in the nth iteration is determined based on the amplitude rate of the target model parameter in the nth iteration. If the amplitude rate of the target model parameter in the nth iteration is greater than or equal to the first threshold, the model training phase in the nth iteration is divided into multiple sub-phases, and the changes in the numerical dispersion of the target model parameter in multiple sub-phases are analyzed to obtain the discrete convergence value of the target model parameter in the nth iteration. Based on the discrete convergence value of the target model parameter in the nth iteration, its local feature value in the nth iteration is determined.

[0011] In some embodiments, when n is greater than 1, the step of adjusting the noise intensity value of the (n-1)th iteration based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration includes: Calculate the difference between the iterative eigenvalues ​​and the eigenth thresholds of each model parameter in the nth iteration to obtain the iterative adjustment values ​​of each model parameter in the nth iteration; Based on the iterative adjustment values ​​of each model parameter in the nth iteration, determine the noise adjustment coefficient of each model parameter in the nth iteration; Calculate the product of the noise adjustment coefficient of each model parameter in the nth iteration and the noise intensity value in the (n-1)th iteration to obtain the noise intensity value in the nth iteration.

[0012] Secondly, embodiments of the present invention also provide a method for secure data sharing in a 5G data center based on network security, applied to a central party, wherein the central party and multiple participating parties constitute a federated model system, and the method includes: Broadcast the model architecture file and initial model parameter information to multiple participants so that they can train the model and obtain the corresponding initial model parameter information. It coordinates multiple participants to iterate parameters, and broadcasts an iteration termination command to multiple participants when the parameter iteration terminates; The nth iteration in the parameter iteration includes: In the nth iteration, individual parameter information and parameter loss values ​​reported by multiple participants are received. When n is 1, the individual parameter information is the initial model parameter information. When n is greater than 1, the individual parameter information is the noise parameter information. The parameter loss value indicates the model loss corresponding to the individual parameter information reported by the corresponding participant in the corresponding iteration. In the nth iteration, the individual parameter information reported by multiple participants is fused to obtain joint parameter information; In the nth iteration, the differences in individual parameter information reported by multiple participants are analyzed to determine the federated feature value of each model parameter. The federated feature value is used to indicate the degree of convergence of the corresponding model parameter in the nth iteration through the combined performance of multiple participants. In the nth iteration, the iteration termination decision is made based on the parameter loss values ​​reported by multiple participants, and the decision result is generated. If the result of the nth iteration indicates that the iteration has terminated, the parameter iteration process ends. If the result of the nth iteration indicates that the iteration has not terminated, the joint parameter information and the federated feature values ​​of each model parameter under the nth iteration are broadcast to multiple participants.

[0013] In some embodiments, the step of analyzing the differences in individual parameter information reported by multiple participants in the nth iteration and determining the federated feature values ​​of each model parameter includes: In the nth iteration, the discrete coefficients of each model parameter are determined based on the individual parameter information reported by multiple participants. In the nth iteration, the discrete coefficients of each model parameter are normalized to obtain the discrete normalized values ​​of each model parameter. In the nth iteration, the federated feature value of each model parameter is determined based on the discrete normalized value of each model parameter, wherein the sum of the discrete normalized value and the federated feature value corresponding to the same model parameter is 1.

[0014] In some embodiments, the step of determining the termination of the iteration based on the parameter loss values ​​reported by multiple participants in the nth iteration and generating a determination result includes: In the nth iteration, the average of the parameter loss values ​​reported by multiple participants is calculated to obtain the federated loss value. If the federal loss value is less than the first loss threshold, a decision result indicating the termination of the iteration is generated; If the federal loss value is greater than or equal to the first loss threshold, a determination result indicating that the iteration has not terminated is generated.

[0015] In some embodiments, the step of determining the termination of the iteration based on the parameter loss values ​​reported by multiple participants in the nth iteration and generating a determination result includes: In the nth iteration, the parameter loss values ​​reported by multiple participants are compared with the second loss threshold to identify the convergent participant among the multiple participants, wherein the parameter loss value reported by the convergent participant is less than the second loss threshold. The convergence ratio is obtained by statistically analyzing the proportion of convergent participants among multiple participants. If the convergence ratio is greater than the ratio threshold, a determination result indicating the termination of the iteration is generated. If the convergence ratio is less than or equal to the ratio threshold, a result indicating that the iteration has not terminated is generated.

[0016] Thirdly, in another embodiment of the present invention, an electronic device is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method described in the first or second aspect above.

[0017] Fourthly, in another embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in the first or second aspect above.

[0018] The present invention has the following beneficial effects: In a federated model system, this invention uses noise addition to encrypt the model parameters reported by each participant to the central authority. Based on the real-time convergence of each model parameter within the federated model system, the noise intensity is dynamically adjusted for each parameter report. When the model parameters are converging, the noise intensity is adaptively increased to enhance the robustness of the model parameters, thus improving the parameter encryption effect. Conversely, when the model parameters are not converging, the noise intensity is adaptively reduced to suppress noise interference with the central authority's fusion of reported parameters from each participant, improving the accuracy of the joint parameter information obtained by the central authority. This dynamic noise intensity adjustment measure enhances the data encryption effect during parameter reporting while ensuring the parameter training effect of the federated model system, providing better data protection for 5G data centers during data sharing. Attached Figure Description

[0019] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for secure data sharing in a 5G data center based on network security, provided by an embodiment of the present invention. Figure 2 This is a flowchart illustrating another method for secure data sharing in a 5G data center based on network security, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a 5G data security sharing slave device based on network security provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of another 5G data security sharing master device based on network security provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a 5G data center data security sharing method based on network security proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] The following description, in conjunction with the accompanying drawings, details a specific scheme for a network security-based 5G data center secure sharing method provided by the present invention.

[0024] This invention provides a method for secure data sharing in 5G data centers based on network security. Specifically, it is applied to a target participant, which is one of multiple participants in a federated model system. The federated model system also includes a central party, such as... Figure 1 As shown, the method includes: Step S11: Train the model based on the model architecture file and original model parameter information issued by the central party to obtain the initial model parameter information.

[0025] In this embodiment, one participant indicates a 5G data center, and the central party indicates a federated learning aggregation server located in the central cloud or core network of multiple participants. The central party and multiple participants together constitute a federated model system.

[0026] It should be noted that different 5G data centers instructed by different participants may be for the same type of business scenario (such as mobile communication scenario, industrial production scenario, etc.). The original format of the business data processed or collected by different 5G data centers instructed by different participants may be the same or different. However, after different 5G data centers instructed by different participants convert the corresponding business data into standard data for model training, the standard data used by different 5G data centers instructed by different participants shall maintain the same format requirements. In addition, different 5G data centers instructed by different participants may belong to the same company or different companies.

[0027] The aforementioned model architecture file can be understood as a file containing a network model used to handle the business requirements of the corresponding business scenario. The aforementioned network model can be a neural network model (e.g., a model used for predicting communication services, a model used for assessing production risks, etc.). The original model parameter information carries the initial values ​​of multiple model parameters included in the model architecture file.

[0028] In this embodiment of the invention, the termination condition for model training performed by the participating party is that the number of training rounds or the training duration of the model training performed by the participating party reaches a set value. The training data used by the participating party for model training is all business data recorded locally.

[0029] The initial model parameter information can be understood as: the parameter values ​​of each model parameter in the last training round, based on the model architecture file and the original model parameter information issued by the central party, until the model training ends.

[0030] Step S21: Perform parameter iteration based on the initial model parameter information until the iteration termination instruction issued by the center is obtained, and determine the noise-free parameter information obtained in the last iteration as the target parameter information.

[0031] Wherein, when n is greater than 1, the nth iteration in the parameter iteration includes: The noise intensity value of the (n-1)th iteration is adjusted based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration. The iterative characteristic values ​​are used to indicate the degree of convergence of the corresponding model parameters in the federated model system in the corresponding iteration. Noise is added to the noise-free parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration. The noise-free parameter information is obtained based on the joint parameter information issued by the central party in the corresponding iteration, and the joint parameter information is obtained based on the noise parameter information reported by multiple participating parties received by the central party in the corresponding iteration. The noise parameter information and parameter loss value of the nth iteration are reported to the center, and the joint parameter information or iteration termination instruction issued by the center are received accordingly. The parameter loss value indicates the model loss of the noise-free parameter information obtained in the corresponding iteration.

[0032] In the above settings, the process of obtaining noiseless parameter information is as follows: the model is trained based on the model architecture file and the joint parameter information sent by the center in the corresponding iteration, and the parameter values ​​indicating the parameters of each model are obtained.

[0033] It should be noted that the larger the iterative eigenvalue, the higher the convergence of the corresponding model parameters in the federated model system during the corresponding iteration.

[0034] The parameter loss value can be obtained based on the preset model loss function, the preset test dataset, and the noiseless parameter information obtained under the corresponding iteration. The smaller the parameter loss value, the better the noiseless parameter information obtained under the corresponding iteration processes the preset test dataset.

[0035] When n is 1, the nth iteration in the parameter iteration includes: The preset initial noise intensity value is adjusted according to the sensitive feature value of each model parameter to obtain the noise intensity value of the nth iteration. The sensitive feature value is used to indicate the noise sensitivity of the corresponding model parameter in the business scenario corresponding to the target participant. Noise is added to the initial model parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration; The noise parameter information of the nth iteration is reported to the central authority, and the joint parameter information issued by the central authority is received accordingly.

[0036] In this embodiment of the invention, Gaussian noise with a normal distribution (other noise models can also be used in the application) is used to add parameter noise, and the standard deviation index in the Gaussian noise model is used as a quantitative representation of the noise intensity. The initial value of the standard deviation index in the Gaussian noise model (which can be understood as the aforementioned initial noise intensity value) is set to 1.

[0037] The larger the sensitivity value, the higher the noise sensitivity of the corresponding model parameters in the business scenario of the target participant, and the lower the intensity of the noise added in the initial noise processing.

[0038] For example, the noise intensity value of the model parameters in the first iteration It can be represented as: in, This represents the preset initial noise intensity value. These represent the sensitive feature values ​​of the model parameters.

[0039] In this embodiment of the invention, noise is added to encrypt the model parameters reported by each participant to the central authority in a federated model system. Based on the real-time convergence of each model parameter within the federated model system, the noise intensity is dynamically adjusted for each parameter report. When the model parameters tend to converge, the noise intensity is adaptively increased to ensure effective parameter encryption. Conversely, when the model parameters do not converge, the noise intensity is adaptively decreased to suppress noise interference with the central authority's fusion of reported parameters from each participant. This improves the accuracy of the joint parameter information obtained by the central authority through fusing the reported parameters. This dynamic noise intensity adjustment measure enhances data encryption during parameter reporting while ensuring the parameter training effect of the federated model system, enabling 5G data centers to achieve better data protection during data sharing.

[0040] Specifically, when n is 1, the steps for obtaining the sensitive feature values ​​of each model parameter include: Based on the numerical changes of each model parameter during the initial training phase, the numerical distribution curve of each model parameter is determined. The initial training phase refers to the process of training the model based on the model architecture file and the original model parameter information. Based on the stage center value of the initial training phase, the numerical distribution curves of each model parameter are divided into a first curve segment and a second curve segment. The first curve segment is the part of the corresponding numerical distribution curve that is before the stage center value, and the second curve segment is the part of the corresponding numerical distribution curve that is after the stage center value. The stage center value indicates the center time between the start and end times of the initial training phase. Analyze the area difference between the first and second curve segments of each model parameter to determine the first sensitive component value of each model parameter; Analyze the degree of numerical fluctuation in the second curve segment of each model parameter to determine the value of the second sensitive component of each model parameter; Based on the first and second sensitive component values ​​of each model parameter, the sensitive feature values ​​of each model parameter are determined.

[0041] In this embodiment, in the rectangular coordinate system corresponding to the numerical distribution curve, the horizontal axis represents the corresponding training round, the vertical axis represents the specific parameter value of the corresponding model parameter under the corresponding training round, and the stage center value can be understood as half of the maximum number of training rounds (such as 50 or 100 times).

[0042] It should be noted that the first curve segment does not include the parameter value indicated by the stage center value, while the second curve segment does include the parameter value indicated by the stage center value.

[0043] The area of ​​a curve segment should be understood as the sum of the multiple parameter values ​​included in the curve segment.

[0044] In this embodiment, the second curve segment of each model parameter includes the ratio of the extreme value difference to the mean among multiple parameter values ​​to quantify the degree of numerical fluctuation. The extreme value difference is the difference between the maximum and minimum values ​​among the corresponding multiple parameter values.

[0045] For example, the sensitive eigenvalues ​​of model parameters It can be represented as: in, This represents the area of ​​the second curve segment corresponding to the model parameters. This represents the area of ​​the first curve segment corresponding to the model parameters. This represents the value of the first sensitive component of the model parameters. This represents the quantized value indicating the degree of numerical fluctuation of the model parameters corresponding to the second curve segment. This represents the value of the second sensitive component of the model parameters.

[0046] It should be noted that, in order to ensure the availability of sensitive feature values, the range of sensitive feature values ​​calculated for each model parameter is limited, so that the minimum value of the sensitive feature value is no less than 0.3 and the maximum value is no more than 4.

[0047] Specifically, when n is greater than 1, the steps for obtaining the iterative feature values ​​of each model parameter in the nth iteration include: The receiving center sends out joint parameter information and federated feature values ​​of each model parameter in the nth iteration. The federated feature values ​​are used to indicate the degree of convergence of the corresponding model parameters in the nth iteration through the combined performance of multiple participants. Model training is performed based on the model architecture file and the joint parameter information of the nth iteration to obtain the noiseless parameter information of the nth iteration. Analyze the numerical changes of each model parameter during the model training phase in the nth iteration, determine the local feature value of each model parameter in the nth iteration, and use the local feature value to indicate the degree of convergence of the corresponding model parameter in the business scenario corresponding to the target participant in the nth iteration. Based on the federated and local feature values ​​of each model parameter in the nth iteration, determine the iterative feature values ​​of each model parameter in the nth iteration.

[0048] In the above setup, the central party quantifies the convergence of the corresponding model parameters in the nth iteration across different business scenarios of the same type from a global perspective, while the participating parties quantify the convergence of the corresponding model parameters in the nth iteration across specific business scenarios from a local perspective. The two parties then combine their findings to determine the convergence of each model parameter in the corresponding iteration, thereby improving the accuracy of the determined iterative feature values.

[0049] In the nth iteration, the larger the federated feature value, the higher the convergence of the corresponding model parameters in the nth iteration through the combined performance of multiple participants; similarly, in the nth iteration, the larger the local feature value, the higher the convergence of the corresponding model parameters in the business scenario corresponding to the target participant in the nth iteration.

[0050] It should be understood that iterative eigenvalues ​​are positively correlated with federated eigenvalues, and iterative eigenvalues ​​are also positively correlated with local eigenvalues.

[0051] For example, when the local feature value and the federated feature value have the same range, the weighted average of the two can be determined as the corresponding iterative feature value. The calculation weights of the two can be set adaptively as needed (e.g., both are set to 0.5, or one is set to 0.4 and the other to 0.6).

[0052] Furthermore, the steps of analyzing the numerical changes of each model parameter during the model training phase in the nth iteration and determining the local feature values ​​of each model parameter in the nth iteration include: Based on the numerical changes of the target model parameter during the model training phase in the nth iteration, the amplitude rate of the target model parameter in the nth iteration is determined. The target model parameter is one of multiple model parameters. The amplitude rate is the ratio of the numerical amplitude of the corresponding model parameter during the model training phase in the corresponding iteration to the maximum parameter value. The numerical amplitude is the difference between the maximum parameter value and the minimum parameter value of the corresponding model parameter during the model training phase in the corresponding iteration. If the amplitude rate of the target model parameter in the nth iteration is less than the first threshold, the local feature value of the target model parameter in the nth iteration is determined based on the amplitude rate of the target model parameter in the nth iteration. If the amplitude rate of the target model parameter in the nth iteration is greater than or equal to the first threshold, the model training phase in the nth iteration is divided into multiple sub-phases, and the changes in the numerical dispersion of the target model parameter in multiple sub-phases are analyzed to obtain the discrete convergence value of the target model parameter in the nth iteration. Based on the discrete convergence value of the target model parameter in the nth iteration, its local feature value in the nth iteration is determined.

[0053] In the above settings, amplitude rate is used to quickly identify situations where parameter values ​​tend to stabilize. However, when the amplitude rate indicates that the parameter values ​​of the model parameters still have large numerical changes, a phased progressive evaluation method is further used to effectively identify situations where the parameter values ​​are in a rapid convergence state. The two methods work together to comprehensively and accurately quantify the local feature values ​​of the corresponding model parameters.

[0054] For example, the amplitude rate in the nth iteration of the model parameters It can be represented as: in, This represents the numerical amplitude of the model parameters in the nth iteration. This represents the maximum absolute value of the parameters in the nth iteration of the model parameters. To prevent constants with a denominator of 0 from being included, their values ​​are set to... .

[0055] In this example, the first threshold can be set to 0.28 based on experience, and if the amplitude rate of the target model parameter in the nth iteration is less than the first threshold, the difference between the value 1 and the amplitude rate of the target model parameter in the nth iteration can be directly determined as its local feature value in the nth iteration.

[0056] In another example, if the amplitude rate of the target model parameters in the nth iteration is greater than or equal to the first threshold, the model training phase in the nth iteration is divided into multiple sub-phases. The discrete convergence value of the target model parameters in the nth iteration can be obtained by calculating the slope of the standard deviation of the target model parameters in the multiple sub-phases (which can be obtained by linear fitting using the least squares method).

[0057] To further explain: when the slope of the standard deviation of the target model parameters in multiple sub-stages is negative, it indicates that the model training stage in the corresponding iteration of the target model parameters is converging, and the larger the absolute value of the slope, the faster the convergence. On the other hand, when the slope of the standard deviation of the target model parameters in multiple sub-stages is non-negative, it indicates that the model training stage in the corresponding iteration of the target model parameters does not show a convergence trend.

[0058] Therefore, when the slope of the standard deviation of the target model parameter in multiple sub-stages is non-negative, the discrete convergence value of the target model parameter in the nth iteration can be directly set to the maximum default value (which can be understood as the maximum value of the discrete convergence value); while when the slope of the standard deviation of the target model parameter in multiple sub-stages is negative, the slope of the standard deviation of the target model parameter in multiple sub-stages is used as the exponent, and the natural constant e is used as the base. The value obtained by exponentiation is determined as the discrete convergence value of the target model parameter in the nth iteration.

[0059] The above sub-stage division can be adapted based on actual needs. For example, if the model training corresponds to 50 training rounds, 10 consecutive training rounds can be divided into one sub-stage, resulting in 5 sub-stages.

[0060] Specifically, when n is greater than 1, the step of adjusting the noise intensity value of the (n-1)th iteration based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration includes: Calculate the difference between the iterative eigenvalues ​​and the eigenth thresholds of each model parameter in the nth iteration to obtain the iterative adjustment values ​​of each model parameter in the nth iteration; Based on the iterative adjustment values ​​of each model parameter in the nth iteration, determine the noise adjustment coefficient of each model parameter in the nth iteration; Calculate the product of the noise adjustment coefficient of each model parameter in the nth iteration and the noise intensity value in the (n-1)th iteration to obtain the noise intensity value in the nth iteration.

[0061] For example, the noise adjustment coefficient of the model parameters in the nth iteration It can be represented as: in, This represents the iterative eigenvalues ​​of the model parameters in the nth iteration. Indicates the feature threshold. It is a natural constant. This represents the iterative adjustment value of each model parameter in the nth iteration.

[0062] In applications, the noise intensity value is set to no higher than 3 to avoid noise intensity exceeding the limit in extreme scenarios and to ensure the effective operation of the federated model system.

[0063] This invention also provides another method for secure data sharing in 5G data centers based on network security, specifically applied to a central party, where the central party and multiple participating parties constitute a federated model system, such as... Figure 2 As shown, the method includes: Step S21: Broadcast the model architecture file and initial model parameter information to multiple participants so that they can train the model and obtain the corresponding initial model parameter information.

[0064] Step S22: Collaborate with multiple participants to iterate parameters, and broadcast an iteration termination instruction to multiple participants when the parameter iteration terminates.

[0065] The nth iteration in the parameter iteration includes: In the nth iteration, individual parameter information and parameter loss values ​​reported by multiple participants are received. When n is 1, the individual parameter information is the initial model parameter information. When n is greater than 1, the individual parameter information is the noise parameter information. The parameter loss value indicates the model loss corresponding to the individual parameter information reported by the corresponding participant in the corresponding iteration. In the nth iteration, the individual parameter information reported by multiple participants is fused to obtain joint parameter information; In the nth iteration, the differences in individual parameter information reported by multiple participants are analyzed to determine the federated feature value of each model parameter. The federated feature value is used to indicate the degree of convergence of the corresponding model parameter in the nth iteration through the combined performance of multiple participants. In the nth iteration, the iteration termination decision is made based on the parameter loss values ​​reported by multiple participants, and the decision result is generated. If the result of the nth iteration indicates that the iteration has terminated, the parameter iteration process ends. If the result of the nth iteration indicates that the iteration has not terminated, the joint parameter information and the federated feature values ​​of each model parameter under the nth iteration are broadcast to multiple participants.

[0066] In the nth iteration, the process of fusing the individual parameter information reported by multiple participants to obtain joint parameter information is as follows: In the nth iteration, the average of the individual parameter information reported by multiple participants is calculated to obtain the joint parameter information. That is, the parameter value of each model parameter in the joint parameter information is the average of the multiple parameter values ​​reported by multiple participants in the corresponding iteration.

[0067] Specifically, in the nth iteration, the steps of analyzing the differences in individual parameter information reported by multiple participants and determining the federated feature values ​​of each model parameter include: In the nth iteration, the discrete coefficients of each model parameter are determined based on the individual parameter information reported by multiple participants. In the nth iteration, the discrete coefficients of each model parameter are normalized to obtain the discrete normalized values ​​of each model parameter. In the nth iteration, the federated feature value of each model parameter is determined based on the discrete normalized value of each model parameter, wherein the sum of the discrete normalized value and the federated feature value corresponding to the same model parameter is 1.

[0068] In the above settings, the coefficient of variation is used to quantify the degree of disorder in the numerical distribution of each model parameter under the corresponding iteration. It should be understood that the lower the degree of disorder, the more similar the numerical trends of the corresponding model parameters are in the model training process of different participants. This also means that the corresponding model parameters show a better convergence trend.

[0069] In one example, in the nth iteration, the discrete normalized value of the model parameter can be the ratio of its corresponding discrete coefficient to its corresponding normalized coefficient, where the normalized coefficient is the sum of the corresponding discrete coefficient and the value 1.

[0070] In some implementations, the step of determining the iteration termination based on the parameter loss values ​​reported by multiple participants in the nth iteration and generating a determination result includes: In the nth iteration, the average of the parameter loss values ​​reported by multiple participants is calculated to obtain the federated loss value. If the federal loss value is less than the first loss threshold, a decision result indicating the termination of the iteration is generated; If the federal loss value is greater than or equal to the first loss threshold, a determination result indicating that the iteration has not terminated is generated.

[0071] In other implementations, the step of determining the iteration termination based on the parameter loss values ​​reported by multiple participants in the nth iteration and generating a determination result includes: In the nth iteration, the parameter loss values ​​reported by multiple participants are compared with the second loss threshold to identify the convergent participant among the multiple participants, wherein the parameter loss value reported by the convergent participant is less than the second loss threshold. The convergence ratio is obtained by statistically analyzing the proportion of convergent participants among multiple participants. If the convergence ratio is greater than the ratio threshold, a determination result indicating the termination of the iteration is generated. If the convergence ratio is less than or equal to the ratio threshold, a result indicating that the iteration has not terminated is generated.

[0072] In application, either of the two determination schemes mentioned above can be flexibly selected to determine the termination condition of parameter iteration according to actual needs.

[0073] It should be understood that the settings of the first loss threshold, the second loss threshold, and the ratio threshold can all be adaptively set according to actual needs, and the present invention does not limit them.

[0074] In one embodiment, the present invention also provides a network security-based 5G data center data security sharing device, applied to a target participant, wherein the target participant is one of multiple participants included in a federated model system, and the federated model system further includes a central party, such as... Figure 3 As shown, the 5G data security sharing slave device 300 based on network security includes: The initial module 301 is used to train the model based on the model architecture file and the original model parameter information issued by the central party, and to obtain the initial model parameter information. The parameter iteration module 302 is used to perform parameter iteration based on the initial model parameter information until the iteration termination instruction issued by the center is obtained, and to determine the noiseless parameter information obtained in the last iteration as the target parameter information. Wherein, when n is greater than 1, the nth iteration in the parameter iteration includes: The noise intensity value of the (n-1)th iteration is adjusted based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration. The iterative characteristic values ​​are used to indicate the degree of convergence of the corresponding model parameters in the federated model system in the corresponding iteration. Noise is added to the noise-free parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration. The noise-free parameter information is obtained based on the joint parameter information issued by the central party in the corresponding iteration, and the joint parameter information is obtained based on the noise parameter information reported by multiple participating parties received by the central party in the corresponding iteration. The noise parameter information and parameter loss value of the nth iteration are reported to the center, and the joint parameter information or iteration termination instruction issued by the center are received accordingly. The parameter loss value indicates the model loss of the noise-free parameter information obtained in the corresponding iteration.

[0075] It should be noted that the 5G data security sharing slave device for data centers based on network security provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the 5G data security sharing slave device for data centers based on network security provided in the above embodiments and... Figure 1 The embodiment of the 5G data security sharing method based on network security shown belongs to the same concept, and its specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0076] In one embodiment, the present invention also provides a 5G data security sharing master device based on network security, applied to a central party, wherein the central party and multiple participating parties constitute a federated model system, such as... Figure 4 As shown, the 5G data security sharing main device 400 based on network security includes: The initial startup module 401 is used to broadcast the model architecture file and the original model parameter information to multiple participants, so that the participants can train the model and obtain the corresponding initial model parameter information. The collaborative iteration module 402 is used to coordinate multiple participants to perform parameter iteration and broadcast an iteration termination command to multiple participants when the parameter iteration terminates. The nth iteration in the parameter iteration includes: In the nth iteration, individual parameter information and parameter loss values ​​reported by multiple participants are received. When n is 1, the individual parameter information is the initial model parameter information. When n is greater than 1, the individual parameter information is the noise parameter information. The parameter loss value indicates the model loss corresponding to the individual parameter information reported by the corresponding participant in the corresponding iteration. In the nth iteration, the individual parameter information reported by multiple participants is fused to obtain joint parameter information; In the nth iteration, the differences in individual parameter information reported by multiple participants are analyzed to determine the federated feature value of each model parameter. The federated feature value is used to indicate the degree of convergence of the corresponding model parameter in the nth iteration through the combined performance of multiple participants. In the nth iteration, the iteration termination decision is made based on the parameter loss values ​​reported by multiple participants, and the decision result is generated. If the result of the nth iteration indicates that the iteration has terminated, the parameter iteration process ends. If the result of the nth iteration indicates that the iteration has not terminated, the joint parameter information and the federated feature values ​​of each model parameter under the nth iteration are broadcast to multiple participants.

[0077] It should be noted that the 5G data security sharing main device for data centers based on network security provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the 5G data security sharing main device for data centers based on network security provided in the above embodiments and... Figure 2 The embodiment of the 5G data security sharing method based on network security shown belongs to the same concept, and its specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0078] This invention also provides an electronic device. Please refer to [link to relevant documentation]. Figure 5 The electronic device may include a processor 501, a memory 502, and a program 5021 stored in the memory 502 and capable of running on the processor 501.

[0079] When program 5021 is executed by processor 501, it can achieve the following: Figure 1 or Figure 2 Any steps in the corresponding method embodiments and the achievement of the same beneficial effects will not be repeated here.

[0080] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.

[0081] This invention also provides a readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described functions. Figure 1 or Figure 2 Any step in the corresponding method embodiment can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0082] The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0083] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0084] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0085] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or terminal. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0086] This invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to achieve the network security-based 5G data center secure sharing method provided in the above embodiments.

[0087] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0088] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for secure data sharing in a 5G data center based on network security, applied to a target participant, wherein the target participant is one of multiple participants in a federated model system, and the federated model system further includes a central party, characterized in that, The method includes: The model is trained based on the model architecture file and the original model parameter information issued by the central party to obtain the initial model parameter information; Based on the initial model parameter information, parameter iteration is performed until the iteration termination instruction issued by the center is obtained, and the noiseless parameter information obtained in the last iteration is determined as the target parameter information. Wherein, when n is greater than 1, the nth iteration in the parameter iteration includes: The noise intensity value of the (n-1)th iteration is adjusted based on the iterative characteristic values ​​of each model parameter in the nth iteration to obtain the noise intensity value of the nth iteration. This includes: calculating the difference between the iterative characteristic value and the characteristic threshold of each model parameter in the nth iteration to obtain the iterative adjustment value of each model parameter in the nth iteration; determining the noise adjustment coefficient of each model parameter in the nth iteration based on the iterative adjustment value of each model parameter in the nth iteration; calculating the product of the noise adjustment coefficient of each model parameter in the nth iteration and the noise intensity value of the (n-1)th iteration to obtain the noise intensity value of the nth iteration; the iterative characteristic value is used to indicate the convergence degree of the corresponding model parameter in the federated model system in the corresponding iteration; based on the real-time convergence degree of each model parameter in the federated model system, the noise addition intensity at each parameter reporting is dynamically adjusted to adaptively increase the intensity of the added noise when the model parameters tend to converge, taking into account the stronger robustness of the model parameters; Noise is added to the noise-free parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration. The noise-free parameter information is obtained based on the joint parameter information issued by the central party in the corresponding iteration, and the joint parameter information is obtained based on the noise parameter information reported by multiple participating parties received by the central party in the corresponding iteration. The noise parameter information and parameter loss value of the nth iteration are reported to the center, and the joint parameter information or iteration termination instruction issued by the center are received accordingly. The parameter loss value indicates the model loss of the noise-free parameter information obtained in the corresponding iteration. When n is greater than 1, the steps for obtaining the iterative feature values ​​of each model parameter in the nth iteration include: The receiving center sends out joint parameter information and federated feature values ​​of each model parameter in the nth iteration. The federated feature values ​​are used to indicate the degree of convergence of the corresponding model parameters in the nth iteration through the combined performance of multiple participants. Model training is performed based on the model architecture file and the joint parameter information of the nth iteration to obtain the noiseless parameter information of the nth iteration. Analyze the numerical changes of each model parameter during the model training phase in the nth iteration, determine the local feature value of each model parameter in the nth iteration, and use the local feature value to indicate the degree of convergence of the corresponding model parameter in the business scenario corresponding to the target participant in the nth iteration. Based on the federated and local feature values ​​of each model parameter in the nth iteration, determine the iterative feature values ​​of each model parameter in the nth iteration.

2. The method for secure data sharing in a 5G data center based on network security according to claim 1, characterized in that, When n is 1, the nth iteration in the parameter iteration includes: The preset initial noise intensity value is adjusted according to the sensitive feature value of each model parameter to obtain the noise intensity value of the nth iteration. The sensitive feature value is used to indicate the noise sensitivity of the corresponding model parameter in the business scenario corresponding to the target participant. Noise is added to the initial model parameter information of the nth iteration based on the noise intensity value of the nth iteration to obtain the noise parameter information of the nth iteration; The noise parameter information of the nth iteration is reported to the central authority, and the joint parameter information issued by the central authority is received accordingly.

3. The method for secure data sharing in a 5G data center based on network security according to claim 2, characterized in that, When n is 1, the steps for obtaining the sensitive feature values ​​of each model parameter include: Based on the numerical changes of each model parameter during the initial training phase, the numerical distribution curve of each model parameter is determined. The initial training phase refers to the process of training the model based on the model architecture file and the original model parameter information. Based on the phase center value of the initial training phase, the numerical distribution curves of each model parameter are divided into a first curve segment and a second curve segment. The first curve segment is the part of the corresponding numerical distribution curve that is before the phase center value, and the second curve segment is the part of the corresponding numerical distribution curve that is after the phase center value. The phase center value indicates the center time between the start and end times of the initial training phase. Analyze the area difference between the first and second curve segments of each model parameter to determine the first sensitive component value of each model parameter; Analyze the degree of numerical fluctuation in the second curve segment of each model parameter to determine the value of the second sensitive component of each model parameter; Based on the first and second sensitive component values ​​of each model parameter, the sensitive feature values ​​of each model parameter are determined.

4. The method for secure data sharing in a 5G data center based on network security according to claim 1, characterized in that, The steps for analyzing the numerical changes of each model parameter during the model training phase in the nth iteration and determining the local feature values ​​of each model parameter in the nth iteration include: Based on the numerical changes of the target model parameter during the model training phase in the nth iteration, the amplitude rate of the target model parameter in the nth iteration is determined. The target model parameter is one of multiple model parameters. The amplitude rate is the ratio of the numerical amplitude of the corresponding model parameter during the model training phase in the corresponding iteration to the maximum parameter value. The numerical amplitude is the difference between the maximum parameter value and the minimum parameter value of the corresponding model parameter during the model training phase in the corresponding iteration. If the amplitude rate of the target model parameter in the nth iteration is less than the first threshold, the local feature value of the target model parameter in the nth iteration is determined based on the amplitude rate of the target model parameter in the nth iteration. If the amplitude rate of the target model parameter in the nth iteration is greater than or equal to the first threshold, the model training phase in the nth iteration is divided into multiple sub-phases, and the changes in the numerical dispersion of the target model parameter in multiple sub-phases are analyzed to obtain the discrete convergence value of the target model parameter in the nth iteration. Based on the discrete convergence value of the target model parameter in the nth iteration, its local feature value in the nth iteration is determined.

5. A method for secure data sharing in a 5G data center based on network security, applied to a central party, wherein the central party and multiple participating parties constitute a federated model system, and the participating parties are configured to execute the method of claim 1, characterized in that... The method includes: Broadcast the model architecture file and initial model parameter information to multiple participants so that they can train the model and obtain the corresponding initial model parameter information. It coordinates multiple participants to iterate parameters, and broadcasts an iteration termination command to multiple participants when the parameter iteration terminates; The nth iteration in the parameter iteration includes: In the nth iteration, individual parameter information and parameter loss values ​​reported by multiple participants are received. When n is 1, the individual parameter information is the initial model parameter information. When n is greater than 1, the individual parameter information is the noise parameter information. The parameter loss value indicates the model loss corresponding to the individual parameter information reported by the corresponding participant in the corresponding iteration. In the nth iteration, the individual parameter information reported by multiple participants is fused to obtain joint parameter information; In the nth iteration, the differences in individual parameter information reported by multiple participants are analyzed to determine the federated feature values ​​of each model parameter. This includes: determining the discrete coefficients of each model parameter based on the individual parameter information reported by multiple participants in the nth iteration; normalizing the discrete coefficients of each model parameter in the nth iteration to obtain the discrete normalized value of each model parameter; and determining the federated feature values ​​of each model parameter based on the discrete normalized value of each model parameter in the nth iteration. The sum of the discrete normalized value and the federated feature value for the same model parameter is 1. The federated feature value indicates the degree of convergence of the corresponding model parameter in the nth iteration, as demonstrated by the combined performance of multiple participants. In the nth iteration, the iteration termination decision is made based on the parameter loss values ​​reported by multiple participants, and the decision result is generated. If the result of the nth iteration indicates that the iteration has terminated, the parameter iteration process ends. If the result of the nth iteration indicates that the iteration has not terminated, the joint parameter information and the federated feature values ​​of each model parameter under the nth iteration are broadcast to multiple participants.

6. The method for secure data sharing in a 5G data center based on network security according to claim 5, characterized in that, In the nth iteration, the steps of determining the termination of the iteration based on the parameter loss values ​​reported by multiple participants and generating the determination result include: In the nth iteration, the average of the parameter loss values ​​reported by multiple participants is calculated to obtain the federated loss value. If the federal loss value is less than the first loss threshold, a decision result indicating the termination of the iteration is generated; If the federal loss value is greater than or equal to the first loss threshold, a determination result indicating that the iteration has not terminated is generated.

7. The method for secure data sharing in a 5G data center based on network security according to claim 6, characterized in that, In the nth iteration, the steps of determining the termination of the iteration based on the parameter loss values ​​reported by multiple participants and generating the determination result include: In the nth iteration, the parameter loss values ​​reported by multiple participants are compared with the second loss threshold to identify the convergent participant among the multiple participants, wherein the parameter loss value reported by the convergent participant is less than the second loss threshold. The convergence ratio is obtained by statistically analyzing the proportion of convergent participants among multiple participants. If the convergence ratio is greater than the ratio threshold, a determination result indicating the termination of the iteration is generated. If the convergence ratio is less than or equal to the ratio threshold, a result indicating that the iteration has not terminated is generated.

Images