Method for secure acquisition of naive bayes model in 6g network

Abstract

The application is suitable for the field of information security technology, and provides a security acquisition method of a Naive Bayes model in a 6G network, comprising: acquiring an original data set in the 6G network; through a local user terminal node, at least encoding processing, splitting processing, mixing processing and random response processing are performed on the original data set to obtain a first data set; the first data set is used for sending to a super edge node for processing; through the super edge node, the first data set is sequentially subjected to aggregation processing, frequency statistics processing and encryption processing to obtain a second data set, and the second data set is used for sending to a center node for processing; through the center node, the second data set is sequentially subjected to aggregation processing and decryption processing to obtain a third data set; and the Naive Bayes model is trained by using the third data set. Through the above-mentioned manner, the usability and security of the Naive Bayes model can be effectively considered when the Naive Bayes model is deployed in the 6G network.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and in particular relates to a secure method for obtaining a Naive Bayes model in a 6G network. Background Technology

[0002] The sixth-generation mobile network (6G) is an extension of the fifth-generation mobile network (5G). Building upon the existing 5G network architecture, 6G deploys nodes (which can also be understood as computer devices) in different locations based on their computing power, forming a hierarchical architecture of local user terminal nodes, super edge nodes, and central nodes. Compared to the 5G network architecture, this hierarchical architecture not only achieves breakthroughs in network capacity and transmission speed, but also, in the field of Artificial Intelligence (AI) services, by combining large-scale distributed collaborative machine learning, can collect massive amounts of user data in real time and continuously optimize model capabilities based on this collected data. This real-time data collection for model generation / updating can provide users with high-quality online intelligent services.

[0003] Among the many models included in 6G networks, the Naive Bayes model ( The Naive Bayes Model (NBM) is a model based on the Naive Bayes algorithm. The Naive Bayes model assumes that features are independent of each other. Using a given training set, it learns the joint probability distribution from input to output, based on this assumption. Then, based on the learned model, it predicts the probability of event A given that event B has occurred. In practice, the Naive Bayes model can be applied to text classification, spam filtering, sentiment analysis, personalized webpage recommendations, disease diagnosis, and financial risk management.

[0004] However, in existing 6G networks, the privacy of user data is often overlooked when generating or updating Naive Bayes models using massive amounts of collected user data, posing a risk of leaking sensitive information. Although some engineers have considered introducing homomorphic encryption to protect user data privacy, in practice, homomorphic encryption requires processing massive amounts of user data, which is undoubtedly a significant overhead. Furthermore, traditional training methods for Naive Bayes models based on homomorphic encryption are not suitable for the layered architecture of 6G wireless networks. In other words, existing methods for obtaining Naive Bayes models in 6G networks cannot effectively balance usability and security. Summary of the Invention

[0005] This invention provides a secure method for obtaining Naive Bayes models in 6G networks, which can effectively balance the availability and security of Naive Bayes models when deploying them in 6G networks.

[0006] In a first aspect, the present invention provides a secure method for obtaining a Naive Bayes model in a 6G network, comprising:

[0007] Obtain the raw dataset from the 6G network through the local user terminal node;

[0008] The original dataset is subjected to at least encoding, splitting, promiscuous processing, and random response processing by the local user terminal node to obtain a first dataset; the first dataset is then sent to the super edge node for processing.

[0009] The first dataset is sequentially aggregated, frequency counted, and encrypted using the super edge node to obtain a second dataset, which is then sent to the central node for processing.

[0010] The second dataset is aggregated and decrypted sequentially through the central node to obtain the third dataset;

[0011] A Naive Bayes model is trained using the central node and the third dataset.

[0012] Secondly, the present invention provides a secure acquisition system for Naive Bayes models in a 6G network, comprising: a security control center 1, a local user terminal node 2, a super edge node 3, and a central node 4. The system is used to execute the methods described in the first aspect or any possible implementation thereof. Specifically, the device includes modules (or units) for executing the methods described in the first aspect or any possible implementation thereof.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: In this invention, by sequentially encoding, splitting, promiscuous, and randomly responding to the collected raw data at the local user terminal node to obtain the first dataset, the privacy and security of the first dataset after it flows out of the user domain can be ensured. Even if the first dataset is subsequently maliciously intercepted, sensitive user information included in the original dataset cannot be obtained based on the first dataset. Furthermore, using a local differential privacy method between the local user terminal node and the super edge node enables rapid collection and processing of massive amounts of data without leaking personal privacy, which is beneficial for saving computing resources. Moreover, using a homomorphic encryption model to re-encrypt the data between the super edge node and the central node, compared to collecting and encrypting the raw data at the local user terminal node (which typically incurs huge computing resource costs) before transmitting it to the super edge node, the secure acquisition method provided by this invention uses a local differential privacy method between the local user terminal node and the super edge node to perturb the raw data. This ensures user privacy and security while enabling rapid collection and processing of massive amounts of data, thereby allowing at least frequency statistics and encryption processing to be performed on the first dataset to obtain the second dataset. This approach extracts a small amount of feature data required for model training from super edge nodes and then encrypts this small amount of feature data, thereby greatly reducing the computational overhead and time of homomorphic encryption. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating the application scenario of the secure acquisition method of Naive Bayes model in 6G network provided by the present invention.

[0015] Figure 2 This is a flowchart illustrating the secure acquisition method of the Naive Bayes model in a 6G network provided by the present invention.

[0016] Figure 3 This is a schematic diagram of the process for obtaining the first dataset provided by the present invention;

[0017] Figure 4 This is an example diagram illustrating the adversary guessing probability under different differential privacy budget values ​​provided by this invention;

[0018] Figure 5 This is an example diagram illustrating data processing operations on the original dataset provided by the present invention;

[0019] Figure 6 This is a schematic diagram of the process for obtaining the second dataset provided by the present invention;

[0020] Figure 7 This is a schematic diagram of the process for obtaining a third dataset provided by the present invention;

[0021] Figure 8 This is a structural block diagram of the secure acquisition system for the Naive Bayes model in a 6G network provided by the present invention. Detailed Implementation

[0022] In the following description, specific details such as particular system structures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0023] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0026] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0028] In one possible implementation, the secure acquisition method of the Naive Bayes model in 6G networks provided by this invention can be applied to application scenarios for data processing in 6G networks.

[0029] The following combination Figure 1 Specific scenarios are described with examples.

[0030] Figure 1 This is a schematic diagram illustrating an application scenario of the secure acquisition method for Naive Bayes models in 6G networks provided by this invention. For example... Figure 1 As shown, the application scenarios for data processing in 6G networks include: a central node, multiple super edge nodes, and multiple local user terminal nodes. The central node, super edge nodes, and local user terminal nodes constitute the basic architecture of the 6G network. The central node, located at the network center, serves as the information transmission hub, receiving and processing data sent by the super edge nodes. Super edge nodes, situated between the central node and local user terminal nodes, can receive data from multiple local user terminal nodes and transmit the received data to the central node. Local user terminal nodes can directly interact with users to provide various services and collect user information during these interactions, which is then aggregated and sent to the corresponding super edge nodes. A super edge node can be associated with one or more local user terminal nodes, and a central node can be associated with at least two super edge nodes.

[0031] Local user terminal nodes can collect user data in real time and transmit it to the corresponding super edge nodes. After processing the received user data, the super edge nodes send the processed data to the central node. The central node then uses the received data to generate / update various models on the 6G network in real time. Among these models is the Naive Bayes model, a classic intelligent classification method that can learn from data in the 6G network to provide efficient classification services.

[0032] The current process for obtaining a Naive Bayes model involves: acquiring a large amount of raw data to generate a sample dataset, which includes multiple labeled samples, each corresponding to a set of features and a class label; then extracting the features of each sample in the sample dataset and converting them into a format that the Naive Bayes model can process, thus obtaining a training dataset; and finally training the Naive Bayes model based on the training dataset to obtain a trained Naive Bayes model. However, in the traditional plaintext training process, users' privacy information is exposed throughout, and sensitive user information can be easily stolen by attackers, posing a serious security problem. Although some technicians have recognized the importance of protecting the security of raw data, for example, by using homomorphic encryption technology to protect the privacy of raw data, in 6G networks, homomorphic encryption technology requires a lot of time to process massive amounts of raw data and consumes a lot of computing power during the computation process. In other words, using traditional methods for obtaining a Naive Bayes model in 6G networks cannot ensure the security of raw data while effectively maintaining the usability of the Naive Bayes model.

[0033] In view of this, the present invention proposes a secure method for obtaining Naive Bayes models in 6G networks, which can ensure the security of the original data while effectively taking into account the usability of the Naive Bayes model.

[0034] The following will refer to the appendix of this invention. Figure 2 -Appendix Figure 8 The technical solutions in this invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments.

[0035] Figure 2 This is a flowchart illustrating the secure acquisition method of the Naive Bayes model in a 6G network provided by the present invention. The method includes the following steps:

[0036] Step 100: Obtain the raw dataset from the 6G network through the local user terminal node.

[0037] Here, in a 6G network data classification scenario, local user terminal nodes collect raw data containing sensitive user information in real time. The types of raw data include at least images and text. It should be noted that "raw dataset" here refers to a collection of multiple raw data sets. This invention does not limit the specific method of collecting raw data.

[0038] After obtaining the raw data containing sensitive user information in step 100, the raw data needs to be preprocessed and perturbed in step 110 to protect the sensitive information contained in the raw data, so as to prevent the super edge node from obtaining the sensitive information.

[0039] Step 110: Through the local user terminal node, the original dataset is subjected to at least encoding, splitting, promiscuous processing and random response processing to obtain the first dataset; the first dataset is used to send to the super edge node for processing.

[0040] Each local user terminal node generates a first dataset after processing the original dataset. The term "first dataset" here is used to distinguish it from "second dataset" and "third dataset" below, and does not limit the scope of protection of this invention.

[0041] Here, at the local user terminal node, after obtaining the raw dataset without any processing, a series of data processing steps are required before data transmission to prevent the leakage of sensitive user information in the raw dataset. These data processing steps include at least encoding, splitting, promiscuous processing, and random response processing. Specifically, after preprocessing each piece of raw data in the raw dataset through encoding, splitting, and promiscuous processing, random response processing is applied to the preprocessed raw dataset. Random response processing is a type of data perturbation processing. After completing a series of data processing steps at the local user terminal node, the raw dataset generates a first dataset, which is then transmitted by the local user terminal node to the super edge node for further processing.

[0042] In one possible implementation, a local differential privacy method is used to perform random response processing. The local differential privacy method processes the preprocessed raw data based on a local differential privacy mechanism; it can also be understood as adding noise to the raw data based on a localized differential privacy algorithm. The local differential privacy method protects the privacy of the raw data by setting a differential privacy budget value, allowing data to be shared and analyzed within the required privacy budget. The smaller the differential privacy value, the higher the level of protection. In this invention, random response processing is completed on the client side, i.e., at the local user terminal node. This invention obtains a first dataset by adding noise to each data point in the preprocessed raw dataset using a local differential privacy method.

[0043] By sequentially encoding, splitting, promiscuous, and randomly responding to the collected raw data through local user terminal nodes to obtain the first dataset, the privacy and security of the first dataset after it flows out of the user domain can be ensured. Even if the first dataset is subsequently maliciously intercepted, sensitive user information included in the original dataset cannot be obtained based on the first dataset. Furthermore, all collected data only undergoes the processing in step 110 to meet privacy protection requirements, thereby ensuring that massive amounts of collected data can be processed in real time.

[0044] Steps 100-110 describe the relevant operations of the local user terminal node, including the acquisition of the original dataset and the generation of the first dataset. Step 120 will describe in detail the processing operations of the super edge node on the received first dataset.

[0045] Step 120: Through the super edge node, the first dataset is sequentially aggregated, frequency counted, and encrypted to obtain the second dataset, which is then sent to the central node for processing.

[0046] In this invention, a super edge node is associated with at least one local user terminal node. Here, the super edge node aggregates the first dataset sent by the local user terminal node and performs aggregation processing on the aggregated first dataset; the aggregation processing refers to merging multiple sets of first datasets into one dataset; then, frequency statistics processing is performed on the aggregated first dataset. Since the first dataset contains data after random response processing, frequency statistics processing is required after receiving the first dataset to obtain useful data information. In other words, the data in the first dataset is filled with noise and needs to undergo frequency statistics processing to remove the noise from each data point in the first dataset and restore the original content that each data point was meant to convey; that is, frequency statistics processing can extract the feature data required for model training; after restoring the data information, each data point is then encrypted to generate a second dataset.

[0047] In one possible implementation, a frequency counting function from a local differential privacy method is used to perform frequency counting on the aggregated first dataset. Then, based on a homomorphic encryption model, the frequency-counted first dataset is encrypted to obtain the second dataset. Specifically, the frequency counting function is used to count the frequency of each data point in the aggregated first dataset, and then an encryption key generated by the homomorphic encryption model is used to encrypt each data point in the frequency-counted first dataset to generate the second dataset.

[0048] Homomorphic encryption models operate on a first dataset based on homomorphic encryption algorithms. The steps of a homomorphic encryption algorithm can be summarized as follows: key generation, encrypting the data using the generated key, performing a specific type of computation (e.g., homomorphic addition) on the encrypted data, and decrypting the encrypted data. Homomorphic encryption allows the result of operations on encrypted data to remain the encrypted form of its plaintext; the final result is only obtained after decryption. This characteristic allows computational operations on data to be performed without exposing the plaintext content. Types of homomorphic encryption include partially homomorphic encryption and fully homomorphic encryption. The difference between partially and fully homomorphic encryption is that partially homomorphic encryption allows specific types of computations on encrypted data, such as addition or multiplication, but cannot support both operations simultaneously; fully homomorphic encryption allows any type of computation on encrypted data.

[0049] In one possible implementation, the super edge node receives only a first dataset. In this case, when processing the first dataset, aggregation is not performed; instead, frequency statistics and encryption are directly executed. For simplicity, the frequency statistics and encryption processes will not be described in detail here.

[0050] Compared to the method of collecting and encrypting raw data at local user terminal nodes (which typically incurs significant computational costs) before transmitting it to super edge nodes, the secure acquisition method provided by this invention uses a local differential privacy method between the local user terminal node and the super edge node to perturb the raw data. This allows for the rapid collection and processing of massive amounts of data while ensuring user privacy and security. The first dataset is then subjected to at least frequency statistics and encryption to obtain the second dataset. This approach extracts a small amount of feature data required for model training at the super edge node and then encrypts this small amount of feature data, thereby significantly reducing the computational overhead and time required for homomorphic encryption.

[0051] In step 120, the super edge node generates a second dataset, which is then transmitted to the central node. Step 130 details the processing of the second dataset at the central node.

[0052] Step 130: Through the central node, the second dataset is aggregated and decrypted sequentially to obtain the third dataset.

[0053] In this invention, the central node is associated with at least two super edge nodes. Therefore, the central node receives at least two sets of second datasets sent by the super edge nodes. The central node aggregates the second datasets sent by the super edge nodes, merging multiple sets of second datasets into one dataset. Then, a homomorphic encryption model is used to decrypt the aggregated second dataset (hereinafter referred to as the "aggregated second dataset") to generate a third dataset for training the Naive Bayes model. It should be noted that the homomorphic encryption model generates a corresponding decryption key while generating multiple encryption keys. The decryption key can only decrypt the aggregated data; it cannot decrypt unaggregated encrypted data, thus protecting the privacy of individual super edge nodes. The decryption key generated by the homomorphic encryption model can then be used to decrypt the aggregated global ciphertext training data.

[0054] By sending the second dataset to the central node, the central node performs aggregation and decryption processing on the second dataset, so that the central node can use the third dataset to update the Naive Bayes model, thereby providing efficient classification services for end users of the 6G network.

[0055] In step 130, the central node obtains the third dataset by processing the second dataset; step 140 will explain how to obtain the Naive Bayes model based on the generated third dataset in a 6G network.

[0056] Step 140: Train the Naive Bayes model using the third dataset through the central node.

[0057] The core of the Naive Bayes model is the Naive Bayes algorithm, whose core formula is expressed as: Where P(B|A) is the posterior probability, which is the probability of event B occurring given that event A has occurred; P(A,B) is the joint probability, which is the probability that events A and B occur simultaneously; P(A|B) is the conditional probability, which is the probability of event A occurring given that event B has occurred; P(A) is the antecedent probability, which is the probability of event A occurring; and P(B) is the antecedent probability, which is the probability of event B occurring.

[0058] Here, after obtaining the third dataset, a Naive Bayes model is trained using it. The training process for the Naive Bayes model can be briefly summarized as follows: Calculate the prior probabilities P(A), P(B), and conditional probability P(A|B) using the third dataset; multiply the prior and conditional probabilities to obtain the joint probability P(A,B); and then obtain the posterior probability based on Bayes' theorem. Here, the third dataset serves as the training data for the Naive Bayes model, and the posterior probability corresponds to the model parameters. Obtaining the posterior probability using the third dataset is equivalent to training the Naive Bayes model.

[0059] Training the Naive Bayes model on a third dataset can effectively protect the security of users' sensitive information, reduce the high computational overhead caused by the budget for protecting users' sensitive information, and provide real-time online classification services for end users in 6G networks.

[0060] In a 6G network, a local user terminal node can acquire multiple raw data in a short period of time. When collecting each raw data, it generates feature and category information related to that raw data. The feature and category information related to that raw data are represented by vectors. That is, the raw dataset collected by the local user terminal node contains at least one raw data, and each raw data includes at least a first feature vector and a first category vector.

[0061] As one possible embodiment, Figure 3 This is a schematic diagram of the process for obtaining the first dataset provided by the present invention. For example... Figure 3 As shown, step 110 above can be achieved through the following steps:

[0062] Step 111: Encode at least the first feature vector and the first category vector using the local user terminal node to obtain the first encoding matrix.

[0063] Here, the original dataset is obtained through the local user terminal node. Each piece of original data in the original dataset corresponds to a first feature vector. and a first-class vector The first feature vector and the first category vector in each piece of original data are encoded to obtain a first encoding matrix; wherein the first encoding matrix includes at least a first feature encoding matrix, a first category encoding matrix, and a first auxiliary encoding matrix, wherein the first auxiliary encoding matrix is ​​calculated based on the first feature encoding matrix and the first category encoding matrix.

[0064] in, This refers to the feature vector of the k-th original data in the original dataset collected by the i-th local user terminal node, representing the feature attribute of the k-th original data; d represents the total dimension of the features, i.e., the number of elements contained in the first feature vector; the elements in the first feature vector... Let j1 be the attribute value in the first feature vector, and It refers to the discrete domain of the j1-th attribute value in the first feature vector. This refers to the category vector of the k-th raw data in the raw dataset collected by the i-th local user terminal node, representing the category to which the k-th raw data belongs; t represents the total dimension of the categories, i.e., the number of elements contained in the first category vector; the elements in the first category vector... Let j2 be the attribute value of the first category vector, and It refers to the discrete domain of the j2-th attribute value in the first category vector.

[0065] One possible implementation uses one-hot encoding to encode the first feature vector and the first category vector separately. One-hot encoding is commonly used to process categorical data, representing categorical variables as numerical data. Specifically, it works by using an N-bit state register to encode N states, each with its own independent register bit, and only one bit being active at any given time. Based on one-hot encoding, the original data representation (e.g., text) can be converted into numerical form. Here, after one-hot encoding, the first feature encoding matrix and the first category encoding matrix are represented as matrices containing only 0 and 1 values. Specifically, one-hot encoding is used to encode the first feature vector, making the first feature vector... In in, This refers to the number of discrete elements of the j1-th attribute value, thus obtaining the first feature encoding matrix. in, Furthermore, one-hot encoding is used to encode the first category vector, so that the first category vector... In in This refers to the number of discrete elements of the j2-th attribute value, thus obtaining the first category coding matrix. in, And using the first feature encoding matrix and the first category coding matrix The first auxiliary coding matrix is ​​calculated. in

[0066] Specifically, the first feature encoding matrix First category coding matrix and the first auxiliary coding matrix They can be represented as follows:

[0067]

[0068]

[0069]

[0070] Among them, the first auxiliary coding matrix It uses the first category coding matrix Each row of elements in the matrix is ​​associated with the first feature encoding matrix. This is obtained by multiplying each element of the corresponding row in the matrix. The first auxiliary encoding matrix is ​​used. The process of obtaining the first row of elements will be explained in detail: using the first category coding matrix The first line respectively with In Multiply, then use the first row. respectively with In Multiply, up to the first row. respectively with In Multiplication; the result of multiple multiplications is used as the first auxiliary encoding matrix. The first row of elements. First auxiliary encoding matrix. The process of obtaining other row elements and the first auxiliary encoding matrix The process of obtaining the first row of elements is the same. For the sake of simplicity, it will not be repeated here.

[0071] This section uses one of the original data entries from the original dataset as an example to illustrate the specific encoding process. This encoding process applies to every single original data entry in the original dataset; for the sake of brevity, the encoding process for each individual original data entry will not be described in detail here.

[0072] After the encoding process of the original data is completed in step 111, the encoded original data will be split in step 112.

[0073] Step 112: The first encoding matrix is ​​split using the local user terminal node to obtain the first split matrix.

[0074] Here, the local user terminal node performs a splitting process on the first feature encoding matrix in the first encoding matrix to obtain a first feature splitting matrix. The first splitting matrix includes at least a first feature splitting matrix, a first category splitting matrix, and a first auxiliary splitting matrix. The first feature splitting matrix is ​​obtained by splitting the first feature encoding matrix, the first category splitting matrix is ​​obtained by splitting the first category encoding matrix, and the first auxiliary splitting matrix is ​​obtained by splitting the first auxiliary encoding matrix.

[0075] Here, after encoding the first feature vector and the first category vector using one-hot encoding in step 111, the first feature encoding matrix is... First category coding matrix and the first auxiliary coding matrix The first feature encoding matrix, containing only 0 and 1 values, is split into two parts. First category coding matrix and the first auxiliary coding matrix Split into only containing The matrix.

[0076] Now, using the first feature encoding matrix Taking an example, the specific splitting process will be explained. The splitting process also applies to the first-class coding matrix. and the first auxiliary coding matrix For the first feature encoding matrix Each row generates a single value containing only the same dimension. The vector, and for the first feature encoding matrix For every position in each row where the value is 1, the corresponding value in the generated vector should be set to 0. Then use the first feature encoding matrix Subtracting each row from the corresponding generated vector yields another vector. These two vectors constitute the first feature encoding matrix. The splitting result for the corresponding row. First feature encoding matrix. The splitting results of each row form the first feature splitting matrix.

[0077] Wherein, the first feature encoding matrix The expression for the k-th row L1 is: First, a vector with the same dimension as L1 is randomly generated. This vector contains only... and Compare with the first feature encoding matrix For all positions in the k-th row L1 where the value is 1, set the corresponding values ​​in the vector to a uniform value. This yields vector L2; subtracting L1 from L2 gives L3; L2 and L3 together form the first feature splitting matrix. The two rows in the matrix are: L2 is row 2k-1, and L3 is row 2k. The same process is applied to all other rows in the encoding matrix, resulting in a split matrix with twice the number of rows of the corresponding encoding matrix.

[0078] Here, vector L2 can be expressed as follows: The expression for L3 is: Furthermore, each element in L3 satisfies:

[0079] For the first feature encoding matrix respectively Each row, first category coding matrix Each row and the first auxiliary coding matrix Each row is split into its constituent parts to generate the first feature splitting matrix. The first category split matrix generated and the generated first auxiliary split matrix The expressions are as follows:

[0080]

[0081]

[0082]

[0083] After generating the first split matrix in step 112, the process of performing a scrambling process on the first split matrix will be explained in detail in step 113.

[0084] Step 113: Perform a scrambling process on the first split matrix through the local user terminal node to obtain the first scrambling matrix.

[0085] Randomization refers to rearranging the rows of data in a matrix in a random order. For example, swapping the data in the first row with the data in the fifth row, or moving the data from the original fourth row to the third row, and so on.

[0086] Here, after the first splitting matrix is ​​subjected to a promiscuous process, a first promiscuous matrix is ​​generated. The first promiscuous matrix includes at least a first feature promiscuous matrix, a first category promiscuous matrix, and a first auxiliary promiscuous matrix. The first feature promiscuous matrix... Based on the first feature split matrix The first-class hybrid matrix obtained by hybridization processing Based on the first category split matrix The first auxiliary hybrid matrix obtained by hybridization processing Based on the first auxiliary split matrix Obtained by mixing.

[0087] After the profanity processing is completed in step 113, the data preprocessing operation on the original dataset is completed. Then, in step 114, random response processing is performed on the preprocessed original dataset.

[0088] Step 114: Determine the random response probability through the local user terminal node, and perform random response processing on the first mixed matrix based on the random response probability to obtain the first dataset.

[0089] Here, the random response probability is used to perturb each element in the first promiscuous matrix to obtain the first dataset. Specifically, the first feature promiscuous matrix... Each element in the matrix undergoes random response processing to obtain the first random matrix. For the first category of mixed matrix Each element in the matrix undergoes random response processing to obtain a second random matrix. For the first auxiliary hybrid matrix Each element in the matrix undergoes random response processing to obtain a third random matrix. First random matrix Second random matrix and the third random matrix This forms the first dataset.

[0090] Alternatively, the process of applying a random response to the first mixed matrix can be expressed as follows:

[0091]

[0092] Where H refers to the perturbation mechanism, x i y refers to the original value in the first promiscuous matrix, y refers to the value after processing by the perturbation mechanism H in the local differential privacy method, f refers to the random response probability, and ε refers to the differential privacy budget value.

[0093] In one possible implementation, determining the random response probability includes: obtaining a differential privacy budget value; and generating the random response probability based on the differential privacy budget value and a random response probability generation function in a local differential privacy method.

[0094] Here, based on the selected local privacy differential privacy scheme, a differential privacy budget value ε that meets security requirements is determined. The smaller the differential privacy budget value ε, the higher the level of privacy protection; that is, the greater the noise added to the original dataset, and the lower the data usability. The determined differential privacy budget value ε is then input into the random response probability generation function Gen of the local differential privacy method. LDPIn (ε), the probability f of the random response is calculated.

[0095] To clearly illustrate the impact of the differential privacy budget value ε on data privacy protection, Figure 4 This is an example diagram illustrating the adversary guessing probability under different differential privacy budget values ​​provided by this invention. For example... Figure 4 As shown, three sets of original data after mixed processing were subjected to random response processing. This invention further split the three sets of data, while the control group did not split the three sets of data to compare the difference in privacy protection between splitting and not splitting. When the differential privacy budget value ε is 1, the adversary guess probability is 0.154 without splitting and 0.133 with splitting; when the differential privacy budget value ε is 2, the adversary guess probability is 0.245 without splitting and 0.165 with splitting; when the differential privacy budget value ε is 3, the adversary guess probability is 0.325 without splitting and 0.19 with splitting; when the differential privacy budget value ε is 4, the adversary guess probability is 0.368 without splitting and 0.202 with splitting; and when the differential privacy budget value ε is 5, the adversary guess probability is 0.388 without splitting and 0.207 with splitting. It is evident that the smaller the differential privacy budget value ε, the higher the degree of privacy protection for the original data. Furthermore, under the same privacy budget, compared to the scheme without data splitting, the data splitting scheme of this invention has a lower probability of the opponent guessing correctly, resulting in better privacy protection.

[0096] Figure 5 This is an example diagram illustrating data processing operations on the original dataset provided by the present invention. To more clearly understand the data preprocessing operations in steps 111-113, the following is an example... Figure 5 The four processes (1), (2), (3), and (4) will be further explained.

[0097] like Figure 5 (1) shows two original data points from the original dataset, and their corresponding data content is as follows:

[0098] m1 = [(1,2),(0)];

[0099] m2 = [(0,1),(1)].

[0100] (1,2) and (0,1) are the first feature vectors, and the size of the discrete domain of the corresponding data is 3; (0) and (1) are the first category vectors, and the size of the discrete domain of the corresponding data is 3.

[0101] After encoding the two original data points using step 111, Figure 5(2) shows the first encoding matrix generated after one-hot encoding, which is expressed as follows:

[0102] m1=[(0,1,0,0,0,1),(1,0,0)];

[0103] m2=[1,0,0,0,1,0),(0,1,0)].

[0104] Use step 112 to Figure 5 The first feature encoding matrix and the first category encoding matrix shown in (2) are split, and the resulting matrix is ​​as follows: Figure 5 As shown in (3): the first feature coding matrix m1 is split into m1' and m1", and the first category coding matrix is ​​split into m'2 and m2". The specific expression is:

[0105]

[0106]

[0107]

[0108]

[0109] Use step 113 to Figure 5 The first feature splitting matrix and the first category splitting matrix shown in (3) are mixed, and the resulting matrix is ​​as follows: Figure 5 As shown in (4):

[0110]

[0111]

[0112]

[0113]

[0114] The encoding, splitting, and promiscuous transmission preprocessing method used in this invention can reduce the accuracy of adversary guessing even with a high privacy budget, effectively protecting user privacy and reducing the high computational overhead caused by the privacy budget, making it suitable for use in real-world scenarios.

[0115] As one possible embodiment, Figure 6 This is a schematic diagram of the process for obtaining the second dataset provided by the present invention. For example... Figure 6 As shown, step 120 above can be achieved through the following steps:

[0116] Step 121: Receive the first dataset sent by the local user terminal node through the super edge node.

[0117] Here, the super edge node receives the first dataset sent by one or more local user terminal nodes. In one possible scenario, the super edge node may receive only one set of the first dataset.

[0118] After receiving the first dataset in step 121, the first dataset is aggregated in step 122. It should be noted that if the super edge node only receives one set of the first dataset, step 122 is skipped and step 123 is executed directly.

[0119] Step 122: Aggregate the first dataset through the super edge node to obtain the first aggregation matrix; the first aggregation matrix includes at least a first feature aggregation matrix, a second category aggregation matrix and a third auxiliary aggregation matrix.

[0120] Here, the super edge node aggregates multiple sets of first datasets received, which include multiple first random matrices, multiple second random matrices, and multiple third random matrices, thereby generating a first aggregate matrix. The first aggregate matrix includes at least a first feature aggregate matrix, a second category aggregate matrix, and a third auxiliary aggregate matrix.

[0121] In one possible implementation, the first dataset includes: multiple first random matrices, multiple second random matrices, and multiple third random matrices; based on this, step 122 includes: aggregating the multiple first random matrices by row concatenation through the super edge node to obtain the first feature aggregation matrix; aggregating the multiple second random matrices by row concatenation through the super edge node to obtain the second category aggregation matrix; and aggregating the multiple third random matrices by row concatenation through the super edge node to obtain the third auxiliary aggregation matrix.

[0122] Here, the multiple first random matrices generated in step 114 are... The aggregation process is performed by row concatenation to obtain the first feature aggregation matrix; and the multiple second random matrices generated in step 114 are processed. The aggregation process is performed by row concatenation to obtain the second category aggregation matrix; and the multiple third random matrices generated in step 114 are processed. The aggregation process is performed by concatenating rows to obtain the third auxiliary aggregation matrix.

[0123] After aggregating multiple first datasets in step 122, frequency statistics are performed on the aggregated first datasets in step 123.

[0124] Step 123: Using the super edge node and the frequency statistics function in the local differential privacy method, perform frequency statistics processing on at least each column of the first feature aggregation matrix, each column of the second category aggregation matrix, and each column of the third auxiliary aggregation matrix to obtain the first frequency set, the second frequency set, and the third frequency set.

[0125] Here, the frequency statistics function Freq(x) is used. i Statistical processing is performed on each column of the first feature aggregation matrix (x, n, f) to obtain the first frequency set; where x i This refers to the data in the first aggregation matrix, where n is the number of probability statistics entries, f is the probability of the random response, and the frequency statistics function Freq(x) is used. i The frequency count is calculated for each column of the second-category aggregation matrix using the formulas f(x,n,f) to obtain the second frequency set; and the frequency count function Freq(x,n,f) is used to calculate the frequency count. i The frequency statistics of each column of the third auxiliary aggregation matrix (n, f) are processed to obtain the third frequency set.

[0126] Alternatively, the processing procedure in step 123 can be expressed as follows:

[0127]

[0128]

[0129]

[0130] Among them, m=1,2,…,d, n=1,2,…,|U m |;l=1,2,...,t,s=1,2,...,|V l |; It represents the frequency statistics of indices m and n in the first frequency set of the i-th super edge node. It is the frequency statistics of the nodes with indices l and s in the second frequency set of the i-th super edge node. is the frequency statistics of the indices (m,n) and (l,s) in the third frequency set of the i-th super edge node, where f refers to the random response probability. and The data is from the first characteristic confounding matrix. and This is the data in the first category of the mixed matrix. and The data is from the first auxiliary promiscuous matrix, where H refers to the perturbation mechanism and START. i Refers to the starting number of the local user terminal node of the i-th super edge node, END iThe terminator refers to the termination number of the local user terminal node of the i-th super edge node, where e is the terminal number and K is the terminal number. e This refers to the number of data entries for the e-th terminal.

[0131] It should be understood that the first, second, and third frequency sets include multiple sets of frequency statistics. After the frequency statistics processing is completed in step 123, these values ​​are substituted into step 124 for merging.

[0132] Step 124: Determine the first merge matrix using the first frequency set, the second frequency set, and the third frequency set.

[0133] Here, the first merging matrix includes at least a first feature merging vector, a second category merging vector, and a third auxiliary merging vector, wherein the first feature merging vector is obtained by merging the first frequency set, the second category merging vector is obtained by merging the second frequency set, and the third auxiliary merging vector is obtained by merging the third frequency set.

[0134] Alternatively, the processing procedure in step 124 can be expressed as follows:

[0135]

[0136]

[0137]

[0138] Among them, NA (i) It is the first feature merge vector of the i-th super edge node, NB (i) It is the second-class merge vector of the i-th super edge node, NC (i) It is the third auxiliary merging vector of the i-th super edge node.

[0139] After the merging process is completed in step 124, the merged first dataset is encrypted in step 125 to generate the second dataset.

[0140] Step 125: Determine the encryption key through the super edge node, and encrypt the first merge matrix based on the encryption key to obtain the second dataset.

[0141] Here, a homomorphic encryption model is used to encrypt the first merging matrix. Specifically, different super edge nodes encrypt the first feature merging vector using different encryption keys they have obtained to obtain first feature encrypted data; they encrypt the second category merging vector using the same encryption key to obtain second category encrypted data; and they encrypt the third auxiliary merging vector using the same encryption key to obtain third auxiliary encrypted data. The first feature encrypted data, the second category encrypted data, and the third auxiliary encrypted data constitute the second dataset.

[0142] In one possible implementation, determining the encryption key includes: obtaining a homomorphic encryption security coefficient; and generating the encryption key based on the homomorphic encryption security coefficient and a key generation function in the homomorphic encryption model.

[0143] Here, based on the selected homomorphic encryption scheme, the homomorphic encryption security coefficient κ is determined; the determined homomorphic encryption security coefficient κ is then input into the key generation function Gen in the homomorphic encryption model. OU In (κ), a decryption key PK = (N, g, h) and an encryption key SK = (p, q, q', μ) are generated. Where N = p 2 q, g∈{1,2,…,n-1}, and h = g q 'modp 2 p, q, q' are large prime numbers, and μ=(L(g p-1 mod p 2 )) -1 mod p, After generating the encryption and decryption keys, the number of super edge nodes (NUM) is used as the basis for the process. edge Split the parameter N into NUM edge portions, and The encryption key for the i-th super edge node is calculated as follows: in s is A random number is selected from the given information. Each super edge node is based on the received encryption key. The first dataset generated in step 124 is encrypted.

[0144] Optionally, the process of encrypting the first dataset can be expressed as follows:

[0145]

[0146] Here, msg is the data to be encrypted in the first merge matrix. This refers to the encrypted value of the data msg, where Enc is the encryption function and r is the value in the encryption process. A value randomly selected from the data. It is the encryption key of the i-th super edge node.

[0147] Furthermore, the process of obtaining the first feature encrypted data, the second category encrypted data, and the third auxiliary encrypted data can be expressed as follows:

[0148]

[0149]

[0150]

[0151] in, This refers to the first feature used to encrypt data. This refers to the second category of encrypted data. This refers to third-party auxiliary encrypted data.

[0152] Based on steps 121-125, the second dataset can be obtained by operating on the first dataset generated after processing in steps 111-114.

[0153] By using a homomorphic encryption model to re-encrypt data between super edge nodes and central nodes, compared to the method of collecting and encrypting the original data at local user terminal nodes (which usually consumes huge computing resources) before transmitting it to super edge nodes, the method provided by this invention, which performs at least frequency statistics processing and encryption processing on the first dataset to obtain the second dataset, can extract a small amount of feature data required for model training and then encrypt the small amount of feature data, thereby greatly reducing the computational overhead and computation time of homomorphic encryption.

[0154] As one possible embodiment, Figure 7 This is a schematic diagram of the process for obtaining a third dataset provided by the present invention. For example... Figure 7 As shown, step 130 above can be achieved through the following steps:

[0155] Step 131: Receive the second dataset sent by the super edge node through the central node.

[0156] Here, the central node is associated with at least two super edge nodes, and thus receives the second dataset sent by multiple super edge nodes. After receiving the second dataset in step 131, the second dataset is aggregated in step 132.

[0157] Step 132: Through the central node, aggregate the second dataset to obtain global encrypted training data; the global encrypted training data includes at least a fourth feature aggregation matrix, a fifth category aggregation matrix, and a sixth auxiliary aggregation matrix.

[0158] Here, the central node cannot decrypt each second dataset individually; it must first aggregate the received multiple sets of second datasets before it can be correctly decrypted. This homomorphic scheme of multiple encryption public keys and a single decryption private key effectively ensures the privacy and security of data at the super edge node.

[0159] In one possible implementation of the encryption model, the second dataset includes: multiple first-feature encrypted data, multiple second-category encrypted data, and multiple third-auxiliary encrypted data; step 132 includes: aggregating the multiple first-feature encrypted data through the central node to obtain the fourth-feature aggregation matrix; aggregating the multiple second-category encrypted data through the central node to obtain the fifth-category aggregation matrix; and aggregating the multiple third-auxiliary encrypted data through the central node to obtain the sixth-auxiliary aggregation matrix.

[0160] Optionally, the processing flow of step 132 can be expressed as follows:

[0161]

[0162]

[0163]

[0164] In step 132, after completing the aggregation process of multiple sets of the second dataset, in step 133, the aggregated second dataset is decrypted. Step 133: Determine the decryption key through the central node, and based on the decryption key, decrypt at least the fourth feature aggregation matrix, the fifth category aggregation matrix, and the sixth auxiliary aggregation matrix to obtain the third dataset.

[0165] Here, in step 125, the key generation function in the homomorphic encryption model generates multiple encryption keys and one decryption key. The multiple encryption keys are distributed to each super edge node, and the decryption key is sent to the central node to decrypt the aggregated matrix to obtain the third dataset for training the Naive Bayes model.

[0166] In one possible implementation, step 133 includes: generating the decryption key based on the key generation function in the homomorphic encryption model; decrypting the fourth feature aggregation matrix based on the decryption key to obtain fourth feature training data; decrypting the fifth category aggregation matrix based on the decryption key to obtain fifth category training data; decrypting the sixth auxiliary aggregation matrix based on the decryption key to obtain sixth auxiliary training data; and using at least the dataset consisting of the fourth feature training data, the fifth category training data, and the sixth auxiliary training data as the third dataset.

[0167] Therefore, based on steps 131-134, the processing procedure in step 140 can be expressed as follows:

[0168]

[0169]

[0170]

[0171] Pr(a m,n =0|b l,s =1)=1-Pr(a m,n =1|b l,s =1);

[0172] Pr(a m,n =0|b l,s =0)=1-Pr(a m,n =1|b l,s =0).

[0173] To address the issue that methods for obtaining Naive Bayes models in 6G networks cannot effectively balance usability and security, this invention proposes collecting data at local user terminal nodes and using a local differential privacy method to rapidly collect raw data while ensuring user privacy. The processed first dataset is then sent to a super edge node. At the super edge node, a homomorphic encryption model is used to encrypt the aggregated data before sending it to the central node for decryption. The decrypted data is then used to deploy the Naive Bayes model. Compared to homomorphic encryption of massive amounts of data at the user terminal, the encoding, splitting, and promiscuous transmission preprocessing method used in this invention reduces the accuracy of adversary guesses even with a high privacy budget, effectively protecting user privacy and reducing the high computational overhead associated with privacy budgets, making it suitable for real-world applications. Furthermore, the secure acquisition method of the Naive Bayes model in the 6G network provided by this invention can quickly collect massive amounts of data while ensuring user privacy and security, and use the user data collected in the network to update the network-side model in real time, thereby enabling the deployment of the Naive Bayes model in the 6G wireless network and allowing the network-side model to provide users with highly accurate classification and recognition services.

[0174] Furthermore, the present invention also provides a data classification method, characterized in that it includes: acquiring data to be classified in a 6G network; inputting the data to be classified into a model obtained by a secure acquisition method based on the Naive Bayes model in a 6G network provided by the present invention, and obtaining a classification result.

[0175] The model obtained based on the secure acquisition method of the Naive Bayes model in the 6G network provided by this invention can be used for text classification, spam filtering, sentiment analysis, personalized webpage recommendation, disease diagnosis, and financial risk management, etc.

[0176] Corresponding to the secure acquisition method of the Naive Bayes model in the 6G network described in the above embodiments, Figure 8 This is a block diagram of the secure acquisition system for the Naive Bayes model in a 6G network provided by this invention. For ease of explanation, only the parts relevant to this invention are shown. (Refer to...) Figure 8 The system includes: a security control center 1, local user terminal nodes 2, a super edge node 3, and a central node 4. Among them:

[0177] Security Control Center 1 is used to complete system initialization, set appropriate privacy budget and security parameters, provide random response probability parameters for local user terminal node 2, provide data encryption keys for super edge node 3, and provide data decryption keys for central node 4.

[0178] Security Control Center 1 includes:

[0179] The security factor selection module is used to select a security factor based on different privacy budgets and homomorphic encryption security requirements; the random response probability generation module is used to generate the corresponding random response probability based on the privacy budget value selected by the security factor selection module.

[0180] The random response probability distribution module is used to distribute random response probabilities to user terminal 2;

[0181] The key generation module is used to generate a homomorphic encryption system and homomorphic encryption security parameters based on the selected security coefficient; generate an encryption key and split the encryption key according to the number of super edge nodes 3; and generate a decryption key for the central node 4 by combining the encryption key of the homomorphic encryption system and the security parameters.

[0182] The key distribution module is used to distribute encryption keys to super edge nodes 3 and send decryption keys to central nodes 4.

[0183] Local user terminal node 2 includes:

[0184] The raw data acquisition module is used to acquire raw datasets from the 6G network through local user terminal nodes;

[0185] The local training data acquisition module is used to perform at least encoding, splitting, promiscuous processing and random response processing on the original dataset through the local user terminal node to obtain a first dataset; the first dataset is used to send to the super edge node for processing.

[0186] Optionally, the local training data acquisition module is configured to: encode at least the first feature vector and the first category vector using the local user terminal node to obtain a first encoding matrix; split the first encoding matrix using the local user terminal node to obtain a first split matrix; perform promiscuous processing on the first split matrix using the local user terminal node to obtain a first promiscuous matrix; determine the random response probability using the local user terminal node, and perform random response processing on the first promiscuous matrix based on the random response probability to obtain a first dataset.

[0187] Optionally, determining the random response probability includes: obtaining a differential privacy budget value; and generating the random response probability based on the differential privacy budget value and the random response probability generation function in the local differential privacy method.

[0188] Super edge node 3 is used to perform aggregation processing, frequency statistics processing and encryption processing on the first dataset in sequence to obtain a second dataset, which is then sent to the central node for processing.

[0189] Optionally, the super edge node corresponds to at least two local user terminal nodes; the super edge node 3 is used to: receive the first dataset sent by the local user terminal nodes through the super edge node; perform aggregation processing on the first dataset through the super edge node to obtain a first aggregation matrix; the first aggregation matrix includes at least a first feature aggregation matrix, a second category aggregation matrix, and a third auxiliary aggregation matrix; through the super edge node, using the frequency statistics function in the local differential privacy method, perform frequency statistics processing on at least each column of data in the first feature aggregation matrix, each column of data in the second category aggregation matrix, and each column of data in the third auxiliary aggregation matrix to obtain a first frequency set, a second frequency set, and a third frequency set; determine a first merge matrix through the first frequency set, the second frequency set, and the third frequency set; determine an encryption key through the super edge node, and encrypt the first merge matrix based on the encryption key to obtain a second dataset.

[0190] Optionally, determining the encryption key includes: obtaining a homomorphic encryption security coefficient; and generating the encryption key based on the homomorphic encryption security coefficient and a key generation function in the homomorphic encryption model.

[0191] Optionally, the first dataset includes: multiple first random matrices, multiple second random matrices, and multiple third random matrices; the step of aggregating the first dataset through the super edge node to obtain a first aggregated matrix includes: aggregating the multiple first random matrices through the super edge node in a row-wise concatenation manner to obtain a first feature aggregated matrix; aggregating the multiple second random matrices through the super edge node in a row-wise concatenation manner to obtain a second category aggregated matrix; and aggregating the multiple third random matrices through the super edge node in a row-wise concatenation manner to obtain a third auxiliary aggregated matrix.

[0192] Central node 4 is used to perform aggregation and decryption processing on the second dataset in sequence to obtain the third dataset; and is also used to train a Naive Bayes model using the third dataset.

[0193] Optionally, the central node corresponds to at least two super edge nodes; the central node 4 is used to: receive the second dataset sent by the super edge nodes through the central node; perform aggregation processing on the second dataset through the central node to obtain global encrypted training data; the global encrypted training data includes at least a fourth feature aggregation matrix, a fifth category aggregation matrix, and a sixth auxiliary aggregation matrix; determine a decryption key through the central node, and based on the decryption key, perform decryption processing on at least the fourth feature aggregation matrix, the fifth category aggregation matrix, and the sixth auxiliary aggregation matrix to obtain a third dataset.

[0194] Optionally, the second dataset includes: multiple first-feature encrypted data, multiple second-category encrypted data, and multiple third-auxiliary encrypted data; the step of aggregating the second dataset through the central node to obtain global encrypted training data includes: aggregating the multiple first-feature encrypted data through the central node to obtain the fourth-feature aggregation matrix; aggregating the multiple second-category encrypted data through the central node to obtain the fifth-category aggregation matrix; and aggregating the multiple third-auxiliary encrypted data through the central node to obtain the sixth-auxiliary aggregation matrix.

[0195] Optionally, the step of determining the decryption key through the central node, and based on the decryption key, decrypting at least the fourth feature aggregation matrix, the fifth category aggregation matrix, and the sixth auxiliary aggregation matrix to obtain a third dataset includes: generating the decryption key based on the key generation function in the homomorphic encryption model; decrypting the fourth feature aggregation matrix based on the decryption key to obtain fourth feature training data; decrypting the fifth category aggregation matrix based on the decryption key to obtain fifth category training data; decrypting the sixth auxiliary aggregation matrix based on the decryption key to obtain sixth auxiliary training data; and using at least the dataset consisting of the fourth feature training data, the fifth category training data, and the sixth auxiliary training data as the third dataset.

[0196] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A secure method for obtaining a Naive Bayes model in a 6G network, characterized in that, include: Obtain the raw dataset from the 6G network through the local user terminal node; The original dataset is subjected to at least encoding, splitting, promiscuous processing, and random response processing by the local user terminal node to obtain a first dataset; the first dataset is then sent to the super edge node for processing. The first dataset is sequentially aggregated, frequency counted, and encrypted using the super edge node to obtain a second dataset, which is then sent to the central node for processing. The second dataset is aggregated and decrypted sequentially through the central node to obtain the third dataset; A Naive Bayes model is trained using the central node and the third dataset. The original dataset contains at least one original data point, and each original data point includes at least a first feature vector and a first category vector. The first dataset is obtained by performing at least encoding, splitting, promiscuous processing, and random response processing on the original dataset through the local user terminal node, including: The local user terminal node encodes at least the first feature vector and the first category vector to obtain a first encoding matrix. The first encoding matrix includes at least a first feature encoding matrix, a first category encoding matrix, and a first auxiliary encoding matrix, wherein the first auxiliary encoding matrix is ​​calculated based on the first feature encoding matrix and the first category encoding matrix. The first encoding matrix is ​​split using the local user terminal node to obtain a first split matrix; the first split matrix includes at least a first feature split matrix, a first category split matrix, and a first auxiliary split matrix. The first splitting matrix is ​​promiscuous through the local user terminal node to obtain a first promiscuous matrix; the first promiscuous matrix includes at least a first feature promiscuous matrix, a first category promiscuous matrix, and a first auxiliary promiscuous matrix. The random response probability is determined through the local user terminal node, and random response processing is performed on the first hybrid matrix based on the random response probability to obtain the first dataset.

2. The secure acquisition method as described in claim 1, characterized in that, Determining the probability of a random response includes: Obtain the differential privacy budget value; The random response probability is generated based on the differential privacy budget value and the random response probability generation function in the local differential privacy method.

3. The secure acquisition method as described in claim 1, characterized in that, The super edge node corresponds to at least one of the local user terminal nodes; the second dataset is obtained by sequentially performing aggregation processing, frequency statistics processing, and encryption processing on the first dataset through the super edge node, including: The super edge node receives the first dataset sent by the local user terminal node; The first dataset is aggregated using the super edge node to obtain a first aggregation matrix; the first aggregation matrix includes at least a first feature aggregation matrix, a second category aggregation matrix, and a third auxiliary aggregation matrix. Using the super edge node, the frequency statistics function in the local differential privacy method is used to perform frequency statistics processing on at least each column of the first feature aggregation matrix, each column of the second category aggregation matrix, and each column of the third auxiliary aggregation matrix, to obtain the first frequency set, the second frequency set, and the third frequency set respectively. The first merging matrix is ​​determined using the first frequency set, the second frequency set, and the third frequency set; The encryption key is determined through the super edge node, and the first merge matrix is ​​encrypted based on the encryption key to obtain the second dataset.

4. The secure acquisition method as described in claim 3, characterized in that, The determination of the encryption key includes: Obtain the homomorphic encryption security factor; The encryption key is generated based on the homomorphic encryption security coefficient and the key generation function in the homomorphic encryption model.

5. The secure acquisition method as described in claim 3, characterized in that, The first dataset includes: multiple first random matrices, multiple second random matrices, and multiple third random matrices; the aggregation process of the first dataset through the super edge node to obtain the first aggregated matrix includes: The first feature aggregation matrix is ​​obtained by aggregating the multiple first random matrices through the super edge node in a row concatenation manner. By using the super edge node, the multiple second random matrices are aggregated in a row-by-row manner to obtain the second category aggregated matrix; The super edge node is used to aggregate the multiple third random matrices by row concatenation to obtain the third auxiliary aggregate matrix.

6. The secure acquisition method as described in claim 1, characterized in that, The central node corresponds to at least two super edge nodes; The third dataset is obtained by sequentially aggregating and decrypting the second dataset through the central node, including: The central node receives the second dataset sent by the super edge node; The second dataset is aggregated through the central node to obtain global encrypted training data; the global encrypted training data includes at least a fourth feature aggregation matrix, a fifth category aggregation matrix, and a sixth auxiliary aggregation matrix. The decryption key is determined through the central node. Based on the decryption key, at least the fourth feature aggregation matrix, the fifth category aggregation matrix, and the sixth auxiliary aggregation matrix are decrypted to obtain the third dataset.

7. The secure acquisition method as described in claim 6, characterized in that, The second dataset includes: multiple first-feature encrypted data, multiple second-category encrypted data, and multiple third-auxiliary encrypted data; the aggregation process performed on the second dataset through the central node to obtain global ciphertext training data includes: The central node is used to aggregate the multiple first-feature encrypted data to obtain the fourth-feature aggregation matrix. The central node is used to aggregate the multiple second-category encrypted data to obtain the fifth-category aggregation matrix. The central node is used to aggregate the multiple third-aid encrypted data to obtain the sixth-aid aggregation matrix.

8. The secure acquisition method as described in claim 6, characterized in that, The process involves determining a decryption key through the central node, and based on the decryption key, decrypting at least the fourth feature aggregation matrix, the fifth category aggregation matrix, and the sixth auxiliary aggregation matrix to obtain a third dataset, including: The decryption key is generated based on the key generation function in the homomorphic encryption model; Based on the decryption key, the fourth feature aggregation matrix is ​​decrypted to obtain the fourth feature training data; Based on the decryption key, the fifth category aggregation matrix is ​​decrypted to obtain the fifth category training data; Based on the decryption key, the sixth auxiliary aggregation matrix is ​​decrypted to obtain the sixth auxiliary training data; The dataset consisting of at least the fourth feature training data, the fifth category training data, and the sixth auxiliary training data shall be used as the third dataset.

9. A data classification method, characterized in that, include: Acquire data to be classified from a 6G network; The data to be classified is input into the Naive Bayes model obtained by the method described in any one of claims 1 to 8 to obtain the classification result.

Images