Privacy protection scheme for power marketing data feature sharing
Through the distributed deep learning architecture and the differential privacy protection solution of adaptive noise addition, the privacy leakage and illegal access problems in power marketing data sharing are solved, the security sharing and privacy protection of power marketing data are realized, and the operational efficiency and service quality of power enterprises are improved.
Patent Information
- Application Number
- CN202510347881.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-24
- Publication Date
- 2025-07-08
AI Technical Summary
There is a risk of sensitive information leakage and illegal access during the sharing process of power marketing data, making it difficult to achieve effective security and privacy protection.
Adopting a distributed deep learning architecture, combined with the differential privacy protection scheme of adaptive noise addition, we use the two-way protection of central differential privacy and localized differential privacy to achieve secure data transmission and model optimization.
Effectively protect the privacy of power marketing data, ensure the security and availability of data in the sharing process, and achieve the improvement of power enterprise operation efficiency and service quality.
Smart Images
Figure CN120277714A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data security and privacy protection, and specifically to a privacy protection scheme for sharing characteristics of power marketing data. Background Art
[0002] Power marketing data refers to various data generated during the power marketing process, covering information in multiple aspects such as power users, power transactions, and power services. The application of power marketing data covers multiple important fields, including load forecasting, electricity bill management, customer service, marketing, equipment maintenance, etc. Therefore, the management and analysis of power marketing data are crucial for the operation and management of power enterprises. By effectively managing and analyzing power marketing data, the operation efficiency and service quality of power enterprises can be improved, and the market competitiveness can be enhanced, which is of great significance for the operation, management, and service optimization of power enterprises.
[0003] However, since power marketing data contains a lot of sensitive or private data, the sharing of these data brings many conveniences and innovation opportunities while also facing a series of security problems. On the one hand, the sharing of power marketing data is prone to the leakage of sensitive information. Power marketing data contains a large amount of sensitive information, such as user personal information, electricity consumption habits, business secrets, etc. If the data cannot be effectively controlled during the processes of collection, transmission, storage, processing, and use, it may lead to the leakage of these sensitive information, causing great damage to personal privacy and commercial interests. On the other hand, in the face of evolving cyber threats, power marketing data sharing faces the risk of illegal access. Hackers or unauthorized users may attempt to enter the system to obtain, tamper with, or destroy data. To address such problems, effective preventive measures are needed, such as establishing a secure access control and authentication mechanism, restricting system access rights, and implementing network isolation.
[0004] In order to effectively protect security and privacy while maximizing the value of power marketing data during the sharing and use processes, considering the characteristics of large volume, fast growth, strong real-time nature, and high value density of power data, the simple data sharing can be upgraded to a more secure feature sharing mode by constructing a trusted data space. For this reason, this patent proposes a privacy protection method for sharing characteristics of power marketing data, which realizes the local storage and processing of data based on a distributed deep learning architecture; at the same time, in order to achieve the rapid convergence of the model, a differential privacy protection scheme based on adaptive noise addition is proposed during the feature transfer process. Summary of the Invention
[0005] Aiming at the deficiencies of the prior art, the present invention provides a privacy protection method for sharing characteristics of power marketing data, which solves the problems raised in the above background art.
[0006] To achieve the above object, the present invention is realized through the following technical solutions: The method includes the following specific steps:
[0007] S1: Complete the initialization settings, construct a central server in the trusted data space as a centralized model, and deploy the power marketing data owners as distributed nodes to the terminal models;
[0008] S2: Through the data acquisition module, obtain the power marketing data to be shared. Then, through the data preprocessing module, numericalize, normalize, perform One Hot encoding, and regularize the data features to form standardized data and store it in the local server;
[0009] S3: Each client learning node performs model learning based on the local data, obtains the relevant tensors of the local model, where the tensors include neural network weight information, and then iteratively optimizes the model through the loss function to obtain the optimal local model parameters;
[0010] S4: The distributed nodes add adaptive noise to the obtained local model parameters, and then transfer the local model parameters to the server node;
[0011] S5: After receiving the model parameters of each distributed learning node, the server node obtains the global model parameters through weighted aggregation, adds noise based on central differential privacy, and distributes the noisy global model parameters to each client;
[0012] S6: After receiving the global model parameters returned by the server, the client nodes reconstruct the model and call the local data for further training to obtain a new round of local model parameters;
[0013] S7: Repeat and iterate the execution of S4 to S6 until the task ends or the number of iteration rounds is reached.
[0014] A further improvement of the technical solution of the present invention lies in: S1 further includes:
[0015] S1.1: Set the global noise σ A , sensitivity C C parameters in the initialization stage. Through the initially set privacy budget ε, calculate the noise of the i-th client as σ i and obtain the adaptive noise σ i ' of the i-th client. The calculation formula is as follows:
[0016]
[0017] Based on σ i ', calculate the privacy budget ε i consumed by the client. Through the privacy budget ε iAdjust the size to meet the noise requirements in different environments. By setting the minimum value of adaptive noise σ low to control the lower limit of the added noise. At this time, the size of the added noise σ i is expressed as:
[0018] σ″ i = max(σ i ', σ low ) (2)
[0019] S1.2: Build a central server as a centralized model in the trusted data space, and deploy the power marketing data owner as a distributed node to the terminal model.
[0020] A further improvement of the technical solution of the present invention lies in: S2 further includes:
[0021] S2.1: Obtain complete and accurate marketing data through the data acquisition function of the internal system, including customer information, market data, and sales data;
[0022] S2.2: Preprocess the power marketing data to be shared through the data preprocessing module, including deleting blank values and outliers. The preprocessed data is stored in a preset storage structure and data verification is performed;
[0023] S2.3: Extract features from the data to be shared, remove irrelevant and redundant features, and retain useful information;
[0024] S2.4: Numerize, normalize, One Hot encode, and regularize the results of feature extraction to form standardized data and store it in the local server.
[0025] A further improvement of the technical solution of the present invention lies in: S3 further includes:
[0026] S3.1: According to the parameter conditions preset during initialization, the data owner starts the local model to perform local training based on the standardized data generated in S2, and obtains relevant tensors of the local model, and this tensor includes neural network weight information.
[0027] A further improvement of the technical solution of the present invention lies in: S4 further includes:
[0028] S4.1: The local client uses the method of adding adaptive noise to add perturbations to the transmitted gradients, so as to achieve the purpose that the larger the gradient change, the larger the added noise, and the smaller the gradient change, the smaller the added noise. The formula for the local client to add noise is:
[0029]
[0030] C L= median(||w1||2,||w2||2,...,||w i ||2) (5)
[0031] Wherein, in local differential privacy federated learning, is the model parameter of client i in the t-th round of communication, is its L2 norm, C is the clipping threshold, and after clipping, the model parameter is adjusted to The clipping coefficient C in formula (5) L is the median of the L2 norm of the gradient set, is a Gaussian distribution with a mean of 0 and a variance of is the model parameter after clipping and adding noise;
[0032] S4.2: The local model transmits the noisy model parameter to the server node.
[0033] A further improvement of the technical solution of the present invention is that: said S5 further includes:
[0034] S5.1: On the central server, the formula for weighted aggregation of model parameters and adding central differential privacy is as follows:
[0035]
[0036] Wherein, σ C represents the size of the noise added by central differential privacy, σ l represents the sum of the noises already added by the clients. By setting the global noise and the adaptive noise of the clients, an adaptive noise σ C is also added to the central server during the release process of the global model parameters to achieve two-way privacy protection;
[0037] S5.2 Sends the aggregated model parameters with added noise back to each client model.
[0038] A further improvement of the technical solution of the present invention is that: said S6 further includes:
[0039] S6.1: The distributed node receives the global model parameters returned by the server;
[0040] S6.2: The client reorganizes the local model, based on the obtained global model parameters, calls the local data for training, and obtains a new round of local model parameters.
[0041] Beneficial effects
[0042] 1. The present invention proposes a feature sharing privacy protection method for power marketing data. Through the power marketing data feature sharing privacy protection scheme, feature sharing is used to replace simple data sharing, and the security and privacy of the parameter transfer process are protected by adaptive noise during the feature sharing process, effectively realizing data availability without visibility during the power marketing data sharing process.
[0043] 2. The present invention proposes a feature sharing privacy protection method for power marketing data. Under this architecture, the owner of the power marketing data serves as the terminal model of the distributed system. While ensuring local data storage, the marketing data is analyzed using a deep learning model. Based on the security model of differential privacy, a privacy protection scheme based on adaptive noise is proposed. After the distributed learning model obtains the local model parameters, noise is added to the feature interaction data by adding adaptive noise, and the noisy data is sent to the server node. The model parameters are optimized by weighted aggregation of multiple client data, and after adding central differential privacy noise, it is transmitted to each distributed node. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Attached Figure 1 is a flowchart of the present invention;
[0045] Attached Figure 2 is the model architecture diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0046] Various exemplary embodiments, features, and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.
[0047] The special term "exemplary" herein means "serving as an example, embodiment, or illustration". Any embodiment described herein as "exemplary" does not have to be construed as superior to or better than other embodiments.
[0048] In addition, in order to better illustrate the present application, numerous specific details are given in the following specific embodiments. Those skilled in the art should understand that the present application can be implemented without some specific details. In some instances, methods, means, and elements well known to those skilled in the art are not described in detail in order to highlight the gist of the present application.
[0049] To achieve the above-mentioned invention object, the present invention proposes a privacy protection scheme for feature sharing of power marketing data. The following further describes the present application with reference to Figure 1 —2.
[0050] The present invention provides a feature sharing privacy protection method for power marketing data, characterized in that the method comprises the following specific steps:
[0051] S1: Complete the initialization settings. Construct a central server as a centralized model in the trusted data space, and deploy terminal models with power marketing data owners as distributed nodes.
[0052] S2: Through the data acquisition module, obtain the power marketing data to be shared. Then, through the data preprocessing module, numerically transform, normalize, perform One Hot encoding, and regularize the data features to form standardized data and store it in the local server.
[0053] S3: Each client learning node performs model learning based on local data to obtain relevant tensors of a local model, which include information such as neural network weights. Then, through the cross-entropy loss function, the model is iteratively optimized to further obtain the optimal local model parameters, and its calculation is shown in formula (8):
[0054]
[0055] where L represents the loss function, y i and p i are the corresponding events and their probabilities.
[0056] S4: The distributed nodes add adaptive noise to the obtained local model parameters, and then transfer the local model parameters to the server node.
[0057] S5: After receiving the model parameters of each distributed learning node, the server node obtains the global model parameters through weighted aggregation. Then, based on central differential privacy, noise addition is completed, and the noisy global model parameters are distributed to each client.
[0058] S6: After receiving the global model parameters returned by the server, the client node reorganizes the model and further trains it by invoking local data to obtain a new round of local model parameters.
[0059] S7: Repeat steps S4 to S6 iteratively until the task ends or the iteration rounds are reached.
[0060] S1 further includes:
[0061] S1.1: In the initialization stage, set the global noise σ A based on Gaussian noise, calculate the sensitivity C C and other parameters based on the initial L2 norm. Through the initially set privacy budget ε (evenly valued between 0.01 and 10), calculate the noise for the i-th client as σ i。Further, according to Adam, obtain the adaptive noise σ of the i-th client i ', and its calculation formula is as follows:
[0062]
[0063] Through σ i ', the privacy budget ε consumed by this client can be calculated i . By adjusting the size of the privacy budget ε i , the noise requirements in different environments can be achieved. Since there is a risk of insufficient noise addition when the gradient change is small, this solution controls the lower limit of the added noise by setting the minimum value of the adaptive noise σ low , improving the global privacy protection ability. At this time, the size of the added noise σ i ” is expressed as:
[0064] σ″ i = max(σ i ', σ low ) (2)
[0065] S1.2: Build a central server as a centralized model in the trusted data space, and deploy the power marketing data owner as a distributed node for the terminal model.
[0066] S2 also includes:
[0067] S2.1: Obtain complete and accurate marketing data through the data collection function of the internal system, including customer information, market data, sales data, etc. Among them, customer information mainly includes the basic attributes of customers, electricity consumption characteristics, and interaction behavior data; market data mainly includes market demand, competitive analysis, and policy and price data, etc.; sales data covers aspects such as contracts and transactions, sales activities, and channel management.
[0068] S2.2: Preprocess the power marketing data that needs to be shared through the data preprocessing module, including deleting blank values and outliers (data that does not conform to the business logic of power marketing data), etc. The preprocessed data is stored in a pre-set storage structure, and data verification is performed to ensure the accuracy and integrity of the data.
[0069] S2.3: Extract features from the data that needs to be shared, remove irrelevant and redundant features, and retain the information that is most useful for model training.
[0070] S2.4: Numeralize, normalize, One Hot encode, and regularize the results of feature extraction to form standardized data and store it in the local server.
[0071] S3 also includes:
[0072] S3: According to the parameter conditions preset during initialization, the data owner starts local training of the local model using the standardized data generated in S2, and obtains relevant tensors of a local model, which include information such as neural network weights. Specifically, this solution adopts a basic CNN structure, mainly including an input layer, a convolutional layer 1 (16 3×3 filters, ReLU activation), a max pooling layer (2×2), a convolutional layer 2 (32 3×3 filters, ReLU), a max pooling layer (2×2), a flattening layer, a fully connected layer (128 neurons, ReLU), and an output layer (2 neurons, Softmax).
[0073] S4 also includes:
[0074] S4.1: The local client adds perturbations to the transmitted gradients in an adaptive noise addition manner to achieve the purpose that the larger the gradient change, the larger the added noise, and the smaller the gradient change, the smaller the added noise. The formula for the local client to add noise is as follows:
[0075]
[0076]
[0077] C L =median(||w1||2,||w2||2,...,||w i ||2) (5)
[0078] where, in local differential privacy federated learning is the model parameter of client i in the t-th round of communication, is its L2 norm, C is the clipping threshold, and after clipping, the model parameter is adjusted to In the above formula, the clipping coefficient C L is the median of the L2 norm of the gradient set, is a Gaussian distribution with a mean of 0 and a variance of is the model parameter after clipping and adding noise.
[0079] S4.2: The local model transmits the noise-added model parameters to the server node.
[0080] S5 also includes:
[0081] S5.1: On the central server, the formula for weighted aggregation of model parameters and adding central differential privacy is as follows:
[0082]
[0083] where, σ C represents the magnitude of the noise added by central differential privacy, σl It represents the sum of the noises added by the client. By setting the global noise and the adaptive noise of the client, an adaptive noise σ is also added to the central server during the publication process of the global model parameters C to achieve two-way privacy protection.
[0084] S5.2: Send the aggregated model parameters added with noise back to each client model.
[0085] S6 also includes:
[0086] S6.1: The distributed node receives the global model parameters returned by the server.
[0087] S6.2: The client reorganizes the local model, based on the obtained global model parameters, calls the local data for training, and obtains a new round of local model parameters.
[0088] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit. Among them, the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, it can run the invention content of a feature sharing privacy protection method for power marketing data provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.
[0089] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of a computer program and its corresponding general hardware platform. Based on such an understanding, the technical solutions in the embodiments of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a computer program, that is, a software product. The computer program software product can be stored in the storage medium, including several instructions to enable a device (which can be a personal computer, a server, a single-chip microcomputer, an MCU, or a network device, etc.) including a data processing unit to execute the methods described in each embodiment or some parts of the embodiments of the present invention.
[0090] The present invention provides a feature sharing privacy protection method for power marketing data. There are many methods and ways to specifically implement this technical solution. The above is only the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by the prior art.
Claims
1. A feature sharing privacy protection method for power marketing data, characterized in that, The method includes the following specific steps: S1: Complete the initialization settings. Build a central server in the trusted data space as a centralized model, and deploy terminal models with power marketing data owners as distributed nodes; S2: Through the data acquisition module, obtain the power marketing data to be shared. Then, through the data preprocessing module, numerically transform, normalize, perform One Hot encoding, and regularize the data features to form standardized data and store it in the local server; S3: Each client learning node performs model learning based on local data to obtain relevant tensors of the local model, where the tensors include neural network weight information. Then, through the loss function, the model is iteratively optimized to obtain the optimal local model parameters; S4: The distributed nodes add adaptive noise to the obtained local model parameters, and then transfer the local model parameters to the server node; S5: After receiving the model parameters of each distributed learning node, the server node obtains global model parameters through weighted aggregation, adds noise based on central differential privacy, and distributes the noisy global model parameters to each client; S6: After receiving the global model parameters returned by the server, the client node reorganizes the model and further trains it by calling local data to obtain a new round of local model parameters; S7: Repeatedly and iteratively execute S4 to S6 until the task ends or the number of iteration rounds is reached.
2. The feature sharing privacy protection method for power marketing data according to claim 1, characterized in that S1 further includes: S1.1: Set the global noise σ in the initialization stage A and the sensitivity C C parameters. The noise for the i-th client is obtained as σ i through calculation using the initially set privacy budget ε, and the adaptive noise σ i ' for the i-th client is obtained. Its calculation formula is as follows: Based on σ i Calculate the privacy budget ε consumed by the client i , by adjusting the size of the privacy budget ε i To meet the noise requirements in different environments, by setting the adaptive noise minimum value σ low To control the lower limit of the added noise, at this time the added noise size σ i Is expressed as: σ i ” = max(σ i ', σ low ) (2) S1.2: Build a central server in the trusted data space as a centralized model, and deploy terminal models with power marketing data owners as distributed nodes.
3. The feature sharing privacy protection method for power marketing data according to claim 1, characterized in that, S2 further includes: S2.1: Obtain complete and accurate marketing data, including customer information, market data, and sales data, through the data acquisition function of the internal system; S2.2: Preprocess the power marketing data to be shared through the data preprocessing module, including deleting blank values and outliers. The preprocessed data is stored in a preset storage structure and data verification is performed; S2.3: Extract features from the data to be shared, remove irrelevant and redundant features, and retain useful information; S2.4: Numerically transform, normalize, perform One Hot encoding, and regularize the results of feature extraction to form standardized data and store it in the local server.
4. A feature sharing privacy protection method for power marketing data according to claim 1, wherein S3 further includes: S3.1: According to the parameter conditions preset during initialization, the data owner starts the local model to perform local training based on the standardized data generated in S2, and obtains relevant tensors of the local model. The relevant tensors include neural network weight information.
5. A feature sharing privacy protection method for power marketing data according to claim 1, characterized in that, S4 further includes: S4.1: The local client uses the method of adding adaptive noise to add perturbations to the transmitted gradients, so as to achieve the purpose that the larger the gradient change, the larger the added noise, and the smaller the gradient change, the smaller the added noise. The formula for the local client to add noise is: Among them, in local differential privacy federated learning, is the model parameter of client i in the t-th round of communication, is its L2 norm, C is the clipping threshold, and after clipping, the model parameter is adjusted to In the above formula, the clipping coefficient C L is the median of the L2 norm of the gradient set, is a Gaussian distribution with a mean of 0 and a variance of is the model parameter after clipping and adding noise; S4.2: The local model transfers the noisy model parameters to the server node.
6. The feature sharing privacy protection method for power marketing data according to claim 1, wherein S5 further includes: S5.1: On the central server, the formula for weighted aggregation of model parameters and adding central differential privacy is as follows: Among them, σ C represents the magnitude of the noise added by central differential privacy, and σ l represents the sum of the noises that have been added by the client. By setting the global noise and the client's adaptive noise, an adaptive noise σ C is also added to the central server during the release process of the global model parameters, so as to achieve two-way privacy protection; S5.2 Send the aggregated parameters of the model with added noise back to each client model.
7. The feature sharing privacy protection method for power marketing data according to claim 1, characterized in that The said S6 further includes: S6.1: The distributed node receives the global model parameters returned by the server; S6.2: The client reorganizes the local model, based on the obtained global model parameters, calls the local data for training, and obtains a new round of local model parameters.