CN115664720A - A secure power information transmission method using data representation desensitization - Google Patents

Abstract

The invention discloses a safe transmission method for electric power information desensitized by data representation, which comprises the following steps: inputting the non-sensitive input data into a variational automatic encoder for conversion to obtain mean value and variance data, constructing simulated Gaussian distribution, randomly sampling from the Gaussian distribution to obtain characterization features, dividing the characterization features into a zx part and a za part, fitting the transmission information by using the characterization features, fitting sensitive input data by using the za part data, training by generating an antagonistic network, thereby improving the independence of the zx part and the za part, inputting the zx part output after training into a regression network to obtain an individual evaluation value, and carrying out efficiency investigation on the individual evaluation value to judge whether the preset requirements are met. The electric power information safety transmission method provided by the invention processes the information data characteristics to be transmitted by adopting a data representation method in deep learning, so that an attacker cannot reversely deduce the key sensitive information in the original data through the processed data.

Description

A secure power information transmission method using data representation desensitization

technical field

The invention relates to the technical field of information security, in particular to a power information security transmission method using data representation desensitization.

Background technique

With the rapid development of information technology, the relationship between people and computer technology is getting closer. However, in the process of computer network operation, there are potential security risks of information network being attacked, and once the user's information security is not guaranteed, they will face huge risks such as privacy and property. In the power distribution network system, after the user information in the system is stolen by the attacker, the attacker can obtain sensitive information such as power demand, user personal information, and home address through the relevant information of the user, causing potential harm. And by intercepting the feedback information from the user terminal to the system, the attacker can send wrong information about the user's power usage to the power distributor, causing data contradictions and even power distribution confusion. Improving network information security can ensure the normal operation of the organization or institution's business and avoid problems such as leakage of important information on the Internet. Information attack specifically refers to the process of information interaction in different network environments. Attackers use technologies such as data interception and related illegal software to obtain important information in the transmitted data, so that the core information of enterprises and units is stolen. The core of data transmission security is to encrypt data in the links of sending, transmitting, and receiving, so that information attackers cannot easily obtain the content of information and achieve the purpose of data security.

At present, the most common way to improve information security is to encrypt data. Even if the information is seized by an illegal attacker, the attacker still cannot effectively decipher the inherent meaning of the data, and the data of the protection institution or organization will not be further violated. However, attackers can indirectly decipher encrypted information through differential analysis, data mining and other means. The sensitive information of the system needs to be protected the most. Once the sensitive data of an institution or organization is fully obtained and understood by an attacker, it will cause greater losses. Even if the non-sensitive information is stolen by the attacker, the sensitive information of the data owner can be reversed by using other data through technical means, which poses a great security risk. At the same time, encrypting all the transmitted data requires a lot of processing time, which increases the overall time of data transmission and reduces the overall efficiency, which cannot meet the requirements of power data transmission efficiency. And some data in the power information data is non-sensitive data, and sensitive data and non-sensitive data have different security requirements. Therefore, it is urgent to design a method that not only ensures the security of data transmission, but also fully meets the requirements of high efficiency and quasi-real-time power transmission.

The disclosure of the above background technical content is only used to assist in understanding the inventive concepts and technical solutions of the present invention, and it does not necessarily belong to the prior art of this patent application, nor does it necessarily give technical teaching; if there is no clear evidence to show the above content In the case that the patent application has been published before the filing date, the above background technology should not be used to evaluate the novelty and inventiveness of the application.

Contents of the invention

In order to overcome the deficiencies in the prior art, the present invention provides a power information security transmission method using data representation desensitization, and the specific technical scheme is as follows:

On the one hand, a method for secure transmission of power information using data representation desensitization is provided, including the following steps:

的定义，将传输信息分为非敏感输入数据和敏感输入数据，x为非敏感输入数据，a为敏感输入数据，x和a通过用户隐私预定义的信息敏感类别进行区分，y为待预测的任务标签，

为企业通过所述深度学习模型给出的个体评定值；Establish a deep learning model for the transmission information in the form of machine learning, and give the corresponding model input data (x, a, y) and model output data

The definition of the transmission information is divided into non-sensitive input data and sensitive input data, x is non-sensitive input data, a is sensitive input data, x and a are distinguished by information sensitive categories predefined by user privacy, y is to be predicted task label,

is the individual evaluation value given by the enterprise through the deep learning model;

对

进行效率考察以判断是否满足预设要求，若不满足预设要求，则需进一步优化所述深度学习模型，直至满足预设要求，将优化后输出的zx部分作为加密信息进行传输。In the deep learning model, the non-sensitive input data is input into a variational autoencoder for conversion to obtain mean and variance data, and the obtained mean and variance are used to construct a simulated Gaussian distribution, and random samples are taken from the Gaussian distribution to obtain representative features , denoted as z, divide the characteristic features into two parts, respectively denoted as zx part and za part, use the characterization features to fit the transmission information, use the za part data to fit the sensitive input data, and generate an adversarial network Perform training to improve the independence of the zx part and the za part, and use the zx part output after training as the encrypted feature that has removed the sensitive input to enter the regression network, and then get

right

Efficiency inspection is carried out to determine whether the preset requirements are met. If the preset requirements are not met, the deep learning model needs to be further optimized until the preset requirements are met, and the optimized output zx part is transmitted as encrypted information.

Further, the efficiency inspection includes the inspection of the availability of encrypted information and the inspection of data containing sensitive information indicators, wherein the inspection of the availability of encrypted information is carried out using the following formula,

是企业通过深度学习模型给出的加密信息预测的个体评定值；In the formula, A _t is the efficiency of encrypted information to restore individual t, y _t is the individual evaluation value predicted by the enterprise through actual information,

It is the individual evaluation value predicted by the encrypted information given by the enterprise through the deep learning model;

The data contains sensitive information index inspection using the following formula,

是压缩数据消除敏感属性能力的指标；

是企业依据深度学习模型给出加密信息表征做出的评估结果，

var(x)是x的方差统计值；T是地域总数；k是敏感输入类型；k_t表示个体t的敏感属性k数值分布；

是x的平均值，m是规定的敏感信息数量。In the formula,

is an indicator of the ability to compress data to eliminate sensitive attributes;

It is the evaluation result of the encrypted information representation given by the enterprise based on the deep learning model.

var(x) is the variance statistical value of x; T is the total number of regions; k is the sensitive input type; k _t represents the numerical distribution of the sensitive attribute k of individual t;

is the average value of x, and m is the specified amount of sensitive information.

Further, the variational autoencoder includes a mean encoder and a variance encoder, and both the mean encoder and the variance encoder define a 2-layer neural network, both use the relu function as the interlayer activation function, and both use the sigmoid function as the Tail layer activation function,

Wherein, the conversion formula corresponding to the mean encoder is as follows:

u＝sigmoid(W _u2 ×relu(W _u1 ×x+b _u1 )+b _u2 )

The conversion formula corresponding to the variance encoder is as follows:

v＝sigmoid(W _v2 ×relu(W _v1 ×x+b _v1 )+b _v2 )

In the formula, u and v represent the mean and variance of the data respectively, W _*i represents the weight of the i-th network layer that processes *, where b _*i represents the offset of the i-th network layer that processes *, and * is u or v.

Further, using the characteristic features to fit the transmission information, the corresponding fitting formula is as follows:

Use the Za part data to fit the sensitive input data, and the corresponding fitting formula is as follows:

为拟合后的传输信息，

为拟合后的敏感输入数据，W_*i表示处理*的第i个网络层权重，其中b_*i表示处理*的第i个网络层的偏移量，*为x或a。In the formula,

is the fitted transmission information,

is the sensitive input data after fitting, W _*i represents the weight of the i-th network layer that processes *, where b _*i represents the offset of the i-th network layer that processes *, and * is x or a.

Further, using the characteristic features to fit the transmission information corresponds to a neural network whose output is restored without sensitive input information, and the loss function of the neural network whose output is restored without sensitive input information is calculated as follows:

Use the za part of the data to fit the sensitive input data corresponding to a neural network that outputs sensitive input information restoration, and the neural network loss function of the output sensitive input information restoration is calculated as follows:

The loss function that makes the zx part and the za part independent of each other is calculated as follows:

where p(x|z) is the decoding distribution, q(z|x) is the encoding distribution, p(z) is the characteristic prior, p(a|za) is the sensitive information decoding distribution, za _k is the kth A prediction-sensitive input representation dimension, the relationship between z, zx, za is z=[zx, za];

Through the above three loss functions, the overall loss function of encryption representation is obtained as:

Loss＝Loss _VAE + Loss _a +λ×Loss _z

In the formula, λ is a parameter controlling the degree of removal of sensitive information.

Further, if the availability of encrypted information meets the requirements but the data contains sensitive information indicators do not meet the requirements, then the value of λ needs to be increased for learning optimization again; if the data contains sensitive information indicators meet the requirements but the availability of encrypted information does not meet the requirements, then it is necessary to lower λ is optimized for learning again, where the value range of λ is [0,1].

Further, the regression network is denoted as f, and its corresponding loss function is as follows:

是回归阈值。in,

is the regression threshold.

Further, the transmission information is communication data between the power supply company and the user, the non-sensitive input data includes user power information, power consumption time and account number, and the sensitive input data includes user name, user ID number, User gender, user address, user contact information.

In yet another aspect, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and operable on the processor, when the processor executes the computer program, the power information secure transmission method is realized .

In another aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the power information secure transmission method described in any one is implemented.

Compared with the prior art, the present invention has the following advantages: the method of data representation in deep learning is used to process the characteristics of the information data to be transmitted, so that the attacker cannot deduce the key sensitive information in the original data through the processed data.

Description of drawings

Fig. 1 is a schematic diagram of the data processing process of variational automatic coding in the power information security transmission method using data representation desensitization provided by the embodiment of the present invention;

Fig. 2 is an overall training flow chart in the power information security transmission method using data representation desensitization provided by the embodiment of the present invention;

Fig. 3 is a flow chart of desensitizing information in the power information security transmission method using data representation desensitization provided by an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solutions of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only It is an embodiment of a part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

It should be noted that the terms "first" and "second" in the description and claims of the present invention and the above drawings are used to distinguish similar objects, but not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", as well as any variations thereof, are intended to cover a non-exclusive inclusion.

Deep learning is a machine learning method that uses multi-layer neural networks for data fitting, prediction, and analysis. Representation learning is a data processing method for encoding features in deep learning. By converting the mapping space of data, the potential information of data is mined. A variational autoencoder is an unsupervised data generation model that encodes input data into a representation space and then restores data from the representation space. The encoding process not only encrypts the data, but also obtains more independent characterization features between columns.

The purpose of the present invention is to improve the data encryption method, which can eliminate the sensitive factors in the information while encrypting the information, so that neither the data owner nor the data attacker can deduce the sensitive information of the data owner through the encrypted information. It will greatly enhance the security of the information system.

In one embodiment of the present invention, a method for secure transmission of power information using data representation desensitization is provided, including the following steps:

Step 1. Mathematical Modeling

的定义，将传输信息分为非敏感输入数据和敏感输入数据，x为非敏感输入数据，a为敏感输入数据，x和a通过用户隐私预定义的信息敏感类别进行区分，y为待预测的任务标签，

为企业通过所述深度学习模型给出的个体评定值；Establish a deep learning model for the transmission information in the form of machine learning, and give the corresponding model input data (x, a, y) and model output data

The definition of the transmission information is divided into non-sensitive input data and sensitive input data, x is non-sensitive input data, a is sensitive input data, x and a are distinguished by information sensitive categories predefined by user privacy, y is to be predicted task label,

is the individual evaluation value given by the enterprise through the deep learning model;

Among them, the power distribution management system can be taken as an example to give the corresponding model input data;

Non-sensitive input information x: Describe the owner's information and its electricity consumption information. It does not include defined identity-sensitive input information, it mainly includes related data such as user power information, power consumption time, account number, etc.;

Sensitive input information a: Sensitive information of users in the power distribution system, which includes user name, user ID number, user gender, user address, user contact information, user regional information (Northeast China, North China, South China, etc.), Types of electricity consumption (industrial electricity, commercial electricity, residential electricity, rice field drainage and irrigation electricity), user discounts, etc.;

必须与实际y相符；其次是加密信息的安全性，它不能还原出原始数据带有的敏感信息，或者说，原始信息本身不带有敏感性，但可以通过原始信息进行差分攻击，推断出敏感性，这是不希望看到的，为此，加密信息必须做到敏感信息的保密。Task label y to be predicted: This label is not necessary, and the availability of encrypted information can be checked through this label. Only when encrypted information can be used normally, the encryption process is meaningful. For example, in the actual electricity consumption scenario, according to the encrypted information of the user's historical electricity consumption data, the user's electricity consumption in the next month is predicted, and the predicted amount is compared with the actual user's electricity consumption in the future to evaluate the ability to use the encrypted information. The evaluation can be done without predicting tasks. The model can provide the corresponding decoder to restore the encrypted information, and compare the matching degree of the data before encryption and after restoration, so as to evaluate the usability of encryption. This embodiment uses the label prediction process as an example to evaluate model availability. There are two main indicators to investigate the deep learning information encryption model. The first is the availability of encrypted information. The encrypted information must be able to be restored to a form that the enterprise can understand. Through encrypted information, the evaluation of the individual by the enterprise

must be consistent with the actual y; secondly, the security of the encrypted information, which cannot restore the sensitive information contained in the original data, or in other words, the original information itself does not contain sensitivity, but the differential attack can be inferred through the original information to infer the sensitive information Sex, which is undesirable, for this reason, encrypted information must keep sensitive information confidential.

Step 2. Model network construction

In the deep learning model, see Figure 1 and Figure 2, the non-sensitive input data is input into the variational autoencoder for conversion to obtain mean and variance data, and the obtained mean and variance are used to construct a simulated Gaussian distribution, from the Gaussian distribution random sampling in order to obtain the characteristic feature, denoted as z, the characteristic feature is divided into two parts, respectively recorded as zx part and za part, use the said characteristic feature to fit the transmission information, and use the za part data to fit the sensitive input data , by generating an adversarial network for training to improve the independence of the zx part and the za part.

Specifically, a variational autoencoder and a generative confrontation network are used to desensitize the data feature representation. First, an autoencoder model is built. Formulas (3)-(5) describe the transformation of data through a variational autoencoder. process. The variational autoencoder consists of a mean encoder and a variance encoder. Both encoders define a 2-layer neural network, use the relu function as the interlayer activation function, use the sigmoid function as the tail layer activation function, and use the predicted mean and variance to construct a simulated distribution, and the final characteristic features are randomly sampled from the distribution.

Divide the obtained representation into two parts zx and za, use [zx+za] to fit the original data, and use za to fit the sensitive input data a, formulas (6)-(8) describe this adversarial learning process. Finally, the independence of zx and za is improved. The final output zx is the encrypted feature with sensitive input removed.

u＝sigmoid(W _u2 ×relu(W _u1 ×x+b _u1 )+b _u2 ) (1)

v＝sigmoid(W _v2 ×relu(W _v1 ×x+b _v1 )+b _v2 ) (2)

zx,za=split(z) (4)

为拟合后的传输信息，

为拟合后的敏感输入数据，W_*i表示处理*的第i个网络层权重，其中b_*i表示处理*的第i个网络层的偏移量，*为u、v、x、a。以上操作是为了去敏表征，将变分自动编码器和对抗网络一起作为表征模型，因此表征模型整体包括如下几个子网络：输出数据均值的神经网络，输出数据方差的神经网络，输出无敏感输入信息还原的神经网络和输出敏感输入信息还原的神经网络。In the formula, u and v respectively represent the mean and variance of the data,

is the fitted transmission information,

is the sensitive input data after fitting, W _*i represents the weight of the i-th network layer that processes *, where b _*i represents the offset of the i-th network layer that processes *, and * is u, v, x, a . The above operation is to desensitize the representation, and the variational autoencoder and the confrontation network are used together as the representation model, so the representation model as a whole includes the following sub-networks: the neural network that outputs the mean value of the data, the neural network that outputs the variance of the data, and the output without sensitive input Neural Network for Information Restoration and Neural Network for Output Sensitive Input Information Restoration.

Step 3. Model optimization

The zx part of the final output is used as the encrypted feature that has removed the sensitive input to input the regression network, and then the predicted value of power demand is obtained, and the efficiency of the predicted value of power demand is checked to determine whether it meets the preset requirements. If the preset requirements are not met , the deep learning model needs to be further optimized until the preset requirements are met.

Specifically, model optimization includes the following:

In the variational autoencoder, each column of features is sampled from a Gaussian distribution generated by its own unique predicted mean and variance. By constraining the relationship between the generated Gaussian distribution and the standard Gaussian distribution, the representation features are more independent.

Use the characterization features to fit the transmission information corresponding to a neural network that outputs non-sensitive input information restoration, and constrains the carrying of information in the representation through a simple mean square error loss function; the output has no sensitive input information restoration neural network The loss function of the network is calculated as follows:

Use the za part of the data to fit the sensitive input data corresponding to a neural network that outputs sensitive input information restoration, and constrains the information carried in the za part of the representation through a simple mean square error loss function; the neural network that outputs sensitive input information restoration The network loss function is calculated as follows:

The loss function that makes the zx part and the za part independent of each other is calculated as follows:

where p(x|z) is the decoding distribution, q(z|x) is the encoding distribution, p(z) is the characteristic prior, p(a|za) is the sensitive information decoding distribution, za _k is the kth A prediction-sensitive input representation dimension, the relationship between z, zx, za is z=[zx, za];

Through the above three loss functions, the overall loss function of encryption representation is obtained as:

Loss＝Loss _VAE + Loss _a +λ×Loss _z

In the formula, λ is a parameter that controls the degree of elimination of input information, and the value range of λ is [0,1]. The larger the λ, the deeper the optimization for desensitization is. If the desensitization is improved, the useful information contained in the characterization will be reduced. Therefore, it is also very important to choose an appropriate λ as the optimization balance parameter.

回归网络f对应的损失函数如下：Through the final representation zx, a regression network f is used to output the individual evaluation value, namely

The loss function corresponding to the regression network f is as follows:

是回归阈值。in,

is the regression threshold.

Efficiency inspection is carried out on the obtained forecasted value of electricity demand to judge whether it meets the preset requirements. The efficiency inspection includes the inspection of the availability of encrypted information and the inspection of data containing sensitive information indicators.

Wherein, the investigation of the availability of the encrypted information is carried out using the following formula,

是企业通过深度学习模型给出的加密信息预测的个体评定值；y_t是企业通过实际信息预测的个体评定值。In the formula, _At is the efficiency of the encrypted information to restore the individual t, and the larger the value, the lower the information loss rate of the encrypted information and the more efficient the encryption transmission.

is the individual evaluation value predicted by the encrypted information given by the enterprise through the deep learning model; y _t is the individual evaluation value predicted by the enterprise through the actual information.

The data contains sensitive information index inspection using the following formula,

是压缩数据消除敏感属性能力的指标，其值越大，说明压缩信息中的敏感信息越少，

是企业依据深度学习模型给出加密信息表征做出的评估结果(即

)；var(x)是x的方差统计值；T是个体总数；k是敏感输入类型；k_t表示个体t的敏感属性k数值分布；

是x的平均值；m是规定的敏感信息数量。In the formula,

is an indicator of the ability to compress data to eliminate sensitive attributes. The larger the value, the less sensitive information in the compressed information.

It is the evaluation result made by the enterprise based on the deep learning model to give the encrypted information representation (that is,

); var(x) is the variance statistical value of x; T is the total number of individuals; k is the sensitive input type; k _t represents the numerical distribution of the sensitive attribute k of individual t;

is the average value of x; m is the specified amount of sensitive information.

If the availability of encrypted information meets the requirements but the data contains sensitive information indicators do not meet the requirements, it is necessary to increase the value of λ for learning optimization again; Learning optimization, through continuous adjustment and optimization, until the requirements for the availability of encrypted information and the requirements for data containing sensitive information indicators are met at the same time.

In the embodiment of the present invention, the main work of step 1 is to mathematically model the problem; the main work of step 2 is to build each network framework of the model, the output of the representation is realized by the variational autoencoder, and the information separation of the representation is achieved by strengthening The independence between different parts is achieved, and the representation information is preserved by the decoding network; the main work of step three is to optimize and iterate different neural networks with their own optimization goals. In a preferred embodiment of the present invention, referring to Figure 3, the network module is initialized after the mathematical modeling, and then the representation model is trained, that is, the variational autoencoder and the confrontation network are trained, and the desired value is achieved by adjusting the setting λ Mask the information ability, and then further train the regression model, by adjusting the setting λ and training the representation model, so that the regression model meets the expected accuracy requirements, and finally output the representation model and the regression model, and the model input data is sequentially passed through the representation model and the regression model To get the predicted value of electricity demand.

An embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and operable on the processor, and the above-mentioned safe transmission of electric power information is realized when the processor executes the computer program method. The idea of this device embodiment is the same as the working process of the detection method in the above embodiment, and the entire content of the above detection method embodiment is incorporated into this device embodiment by citing the full text, and will not be described again.

An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above method for safely transmitting power information is implemented. The idea of this embodiment of the storage medium is the same as the working process of the detection method in the above embodiment, and the entire content of the above embodiment of the detection method is incorporated into the embodiment of the storage medium by citing the full text, and will not be described again.

The power information secure transmission method provided by the present invention utilizes the means of representation in deep learning to desensitize and informatize the transmitted data, realizes secure transmission of information, and prevents differential attacks. The present invention characterizes data based on a variational autoencoder and a generative confrontation network, encrypts information with a deep learning model, improves the protection of sensitive information, and prevents more serious information violations.

The power information security transmission method provided by the present invention adopts variational autoencoder and generation confrontation network to reduce the correlation between sensitive information and transmission information, and separates the information of transmission data, so that the sensitive information in the data does not enter the transmission link, and requires Sensitive information cannot be inferred from non-sensitive information in transit.

The above is only a preferred embodiment of the present invention, and does not limit the scope of its patents. Any equivalent structure or equivalent process transformation made by using the description of the present invention and the contents of the accompanying drawings is directly or indirectly used in other related technical fields. All are included in the scope of patent protection of the present invention in the same way.

Claims (10)

1. A safe transmission method for electric power information with data characterization desensitization is characterized by comprising the following steps:

establishing a deep learning model for the transmission information in a machine learning manner, and providing corresponding model input data (x, a, y) and model output data

The definition of (1) is to divide the transmission information into non-sensitive input data and sensitive input data, x is the non-sensitive input data, a is the sensitive input data, x and a are distinguished by the information sensitive category predefined by the user privacy, y is the task label to be predicted,

evaluating values of individuals given by the enterprises through the deep learning model;

in a deep learning model, inputting the non-sensitive input data into a variational automatic encoder for conversion to obtain mean value and variance data, constructing simulated Gaussian distribution by using the obtained mean value and variance, randomly sampling from the Gaussian distribution to obtain a characteristic feature, recording the characteristic feature as z, dividing the characteristic feature into two parts, recording the two parts as a zx part and a za part respectively, fitting the transmission information by using the characteristic feature, fitting the sensitive input data by using the data of the za part, training by generating an antagonistic network to improve the independence of the zx part and the za part, inputting the zx part output after training as an encrypted feature with sensitive input removed into a regression network to obtain the encrypted feature with sensitive input removed, and obtaining the mean value and the variance data

To pair

Carrying out efficiency investigation to judge whether the preset requirements are met, and if the preset requirements are not met, further optimizing the deep learning modelAnd transmitting the optimized zx part as encryption information until the preset requirement is met.

2. The method for safely transmitting power information according to claim 1, wherein the efficiency investigation comprises encrypted information availability investigation and data content sensitive information indicator investigation, wherein the encrypted information availability investigation is performed by using the following formula,

in the formula, A _t Is the efficiency of the cryptographic information pair to recover the individual t, y _t Is an individual evaluation value predicted by enterprises through actual information,

the individual evaluation value is predicted by the encryption information given by the enterprise through a deep learning model;

the data-sensitive information index investigation is performed by using the following formula,

in the formula,

is an index of the ability of the compressed data to eliminate sensitive attributes;

the enterprise gives out the evaluation result made by the encrypted information representation according to the deep learning model,

var (x) is the variance statistic for x; t is the total number of regions; k is a sensitive input type; k is a radical of _t Sensitivity attribute representing individual tk value distribution;

is the average value of x and m is the specified amount of sensitive information.

3. The method for safely transmitting power information according to claim 1, wherein the variable automatic encoder comprises a mean encoder and a variance encoder, the mean encoder and the variance encoder both define a 2-layer neural network, both use a relu function as an interlayer activation function, and both use a sigmoid function as a tail layer activation function,

the conversion formula corresponding to the mean encoder is as follows:

u＝sigmoid(W _u2 ×relu(W _u1 ×x+b _u1 )+b _u2 )

the conversion formula corresponding to the variance encoder is as follows:

v＝sigmoid(W _v2 ×relu(W _v1 ×x+b _v1 )+b _v2 )

in the formula, u and v represent the mean and variance of data, W _*i I-th network layer weight representing processing, wherein b _*i The offset of the ith mesh layer representing the process x is u or v.

4. The power information secure transmission method according to claim 2,

and fitting the transmission information by using the characterization characteristics, wherein a corresponding fitting formula is as follows:

and (3) fitting the sensitive input data by using the Za partial data, wherein the corresponding fitting formula is as follows:

in the formula,

for the transmission information after the fitting to be performed,

for the fitted sensitive input data, W _*i I-th network layer weight representing processing, wherein b _*i The offset of the ith network layer representing process is x or a.

5. The method for safely transmitting power information according to claim 4, wherein the characterization feature is used to fit the transmission information to a neural network without sensitive input information restoration output, and the loss function of the neural network without sensitive input information restoration output is calculated as follows:

fitting sensitive input data to a neural network corresponding to the restoration of the output sensitive input information by using the za part data, wherein the loss function of the neural network restored by the output sensitive input information is calculated as follows:

the loss function that makes the zx and za portions independent of each other is calculated as follows:

where p (x | z) is the decoding distribution, q (z | x) is the encoding distribution, p (z) is the characterising prior, p (a | za) is the sensitive information decoding distribution, za _k Is the kth prediction-sensitive input characterization dimensionThe relationship between degree, z, zx, za is z = [ zx, za =]；

By the three loss functions, the overall loss function when the encrypted representation is obtained is as follows:

Loss＝Loss _VAE +Loss _a +λ×Loss _z

in the formula, λ is a parameter for controlling the degree of elimination of the sensitive information.

6. The method for safely transmitting the electric power information according to claim 5, wherein if the availability of the encrypted information meets the requirement and the availability of the data containing the sensitive information does not meet the requirement, the lambda value needs to be increased for learning optimization again, and if the availability of the data containing the sensitive information meets the requirement and the availability of the encrypted information does not meet the requirement, the lambda value needs to be decreased for learning optimization again, wherein the lambda value range is [0,1].

7. The method according to claim 1, wherein the regression network is denoted as f, and the corresponding loss function is as follows:

wherein,

is the regression threshold.

8. The method for safely transmitting the power information as claimed in claim 1, wherein the transmission information is communication data between a power supply company and a user, the non-sensitive input data comprises user power information, power utilization time and a user number, and the sensitive input data comprises a user name, a user identity card number, a user gender, a user address and a user contact way.

9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for secure transmission of power information according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium on which a computer program is stored, the computer program, when being executed by a processor, implementing the power information secure transmission method according to any one of claims 1 to 8.

Images