Privacy information dissemination control strategy generation method and device

Abstract

The invention provides a privacy information dissemination control strategy generation method and device. The method comprises the following steps: acquiring information forwarding data of a user and social network structure data; performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to perform feature extraction on the social network structure data to obtain user social features; selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into an information propagation model for propagation prediction to obtain propagation probability matrixes corresponding to the different seed sets; and determining a privacy propagation comprehensive income according to the propagation probability matrix, and generating a privacy information propagation control strategy based on the seed set with the maximum privacy propagation comprehensive income. Through the information sharing method and device, the problems that in the prior art, when information is shared in a social network, the privacy strategy making efficiency is low, the privacy leakage risk is difficult to overcome, and the privacy requirements of different users are met are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of information dissemination, and in particular to a method and device for generating a privacy information dissemination control strategy. Background Art

[0002] Social networks have become an indispensable part of modern society and have greatly changed people's information exchange patterns. Users share various types of information, including but not limited to text, pictures, videos, etc., through platforms such as Weibo, Twitter, and Facebook to interact and communicate with others. However, as users enjoy the benefits and fun brought by social platforms, the risk of privacy leakage has also become increasingly prominent. Serious privacy leakage problems may lead to reputation damage, relationship breakdown, and other consequences, which may harm the physical and mental health of the parties involved.

[0003] In social networks, after the information owner selects a seed node (usually from the first-level neighbor nodes) and publishes the information, the information propagation process is not controlled by the information owner. After multiple circulations, the information may be known by users with higher threat levels, causing privacy risks. At the same time, if the information is seen by more users, it will bring higher influence to the information owner. However, users usually lack sufficient background knowledge and risk awareness to fully understand the potential risks of publishing a piece of personal information, so it is difficult to formulate a wise privacy strategy.

[0004] In the prior art, solutions to privacy issues when sharing information in social networks can be mainly divided into two categories: the first category is to evaluate the risk of information released by users; the second category is to automatically generate privacy policies based on features. For the first category of solutions, although they have made contributions to the quantification of privacy risks and help users objectively measure the risks of privacy information, these solutions still require users to have sufficient domain knowledge and invest a lot of time and energy to manage privacy information. Users may find it difficult to effectively use these quantitative results due to lack of professional knowledge or insufficient time, and the privacy needs of each user are different, resulting in unsatisfactory efficiency and effectiveness in privacy policy formulation. For the second category of solutions, although the automatic generation of new policies through historical privacy policies can reduce the user's workload, this solution ignores the fact that users may have formulated unreasonable privacy policies in the past, and may continue to repeat the same mistakes in the future, resulting in the continued existence of privacy leakage risks. Summary of the invention

[0005] The present invention provides a method and device for generating a privacy information dissemination control strategy, which is used to solve the problems in the prior art of low efficiency in formulating privacy strategies and difficulty in overcoming the risk of privacy leakage and meeting user privacy needs when sharing information in social networks.

[0006] The present invention provides a method for generating a privacy information dissemination control strategy, the method comprising the following steps: Obtain user information forwarding data and social network structure data; Performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to perform feature extraction on the social network structure data to obtain user social features; Selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into an information propagation model for propagation prediction, to obtain propagation probability matrices corresponding to different seed sets; The comprehensive benefit of privacy propagation is determined according to the propagation probability matrix, and a privacy information propagation control strategy is generated based on a seed set with the maximum comprehensive benefit of privacy propagation.

[0007] In some embodiments, determining the comprehensive benefit of privacy propagation according to the propagation probability matrix includes: Calculate the influence benefit of each seed node in the seed set based on the propagation probability matrix to obtain the influence benefit corresponding to the seed set; Calculate the privacy risk loss of each seed node in the seed set based on the propagation probability matrix to obtain the privacy risk loss corresponding to the seed set; Preset a benefit weight and a loss weight, weight the influence benefit by the benefit weight to obtain a benefit weighted value, and weight the privacy risk loss by the loss weight to obtain a loss weighted value; The difference between the weighted value of the gain and the weighted value of the loss is taken as the comprehensive benefit of privacy dissemination.

[0008] In some embodiments, the pre-training process of the social network structure data model includes: Obtain data samples corresponding to the user's social network structure; Sampling the initialized potential social variables of each user node in the data sample from a standard normal distribution to obtain the potential social variables corresponding to the data sample; Decoding the latent social variables through a graph-generating neural network in the social network structure data model to obtain a reconstructed social network structure; Encoding the reconstructed social network structure through an autoencoder in the social network structure data model to obtain distribution parameters corresponding to the potential social variables; A joint loss value of the social network structure data model is constructed according to the reconstructed social network structure and the distribution parameters, and the joint loss value is used to perform back propagation in the social network structure data model to update the parameters of the social network structure data model.

[0009] In some embodiments, constructing a joint loss value of the social network structure data model according to the reconstructed social network structure and the distribution parameter includes: Constructing a reconstruction loss value of the social network structure according to the difference value between the reconstructed social network structure and the data sample; Constructing a KL divergence loss value according to the difference between the distribution value of the standard normal distribution and the distribution parameter; The sum of the reconstruction loss value and the KL divergence loss value is used as the joint loss value of the social network structure data model.

[0010] In some embodiments, the training process of the information propagation model includes: Obtain information cascade data sets and corresponding user social network structure data; Calling a pre-trained social network structure data model to extract corresponding training social variables from the user social network structure data; Cut each piece of information cascade data in the information cascade data set according to a preset number of seed nodes to obtain an expanded data set, wherein each cascade sample in the expanded data set carries label data; Calling the information propagation model to encode each cascade sample in the expanded data set to obtain a time impact variable; The training social variable and the time influence variable are concatenated to obtain input features, and the input features are predicted by the information propagation model to obtain a predicted propagation probability corresponding to each cascade sample; A cross entropy loss value is constructed based on the predicted propagation probability and the label data, and back propagation is performed in the information propagation model through the cross entropy loss value to update the parameters of the information propagation model.

[0011] In some embodiments, the step of cutting each piece of information cascade data in the information cascade data set according to a preset number of seed nodes to obtain an expanded data set includes: Preset multiple different seed node numbers k; For each information cascade data with a length of n in the information cascade data set, it is cut into a seed sequence with a length of k and label data with a length of nk according to the number of seed nodes k, where k is an integer greater than 1 and less than n, and n is an integer greater than 2.

[0012] The seed sequence and the label data are combined into corresponding cascade samples, and different cascade samples obtained by cutting each piece of information cascade data are combined into an expanded data set.

[0013] In some embodiments, the propagation probability matrix is ​​predicted by a privacy policy generation model based on a joint attention fusion network, the privacy policy generation model includes a social feature extraction model and a joint attention fusion network model, and the training process of the privacy policy generation model includes: Obtain user forwarding information dataset and corresponding user social network structure samples; Performing data expansion on the information propagation cascade data in the user forwarding information data set to obtain an expanded information propagation cascade data set; Position encoding is performed on the expanded information propagation cascade data set through the social feature extraction model to obtain cascade time series features, and feature extraction is performed on the user social network structure sample through the social feature extraction model to obtain user predicted social features; The cascade time series feature is concatenated with the user prediction social feature to obtain an information propagation feature vector, and the information propagation feature vector is input into the joint attention fusion network model for prediction to obtain a prediction probability matrix; generating a corresponding prediction privacy policy based on the propagation probability matrix; A final loss value of the privacy policy generation model is constructed according to the predicted privacy policy and the actual privacy policy, and the final loss value is used for backpropagation in the privacy policy generation model to update the parameters of the social feature extraction model and the joint attention fusion network model.

[0014] The present invention also provides a device for generating a privacy information dissemination control strategy, the device comprising the following modules: An acquisition module is used to acquire user information forwarding data and social network structure data; An extraction module, used to aggregate and analyze the information forwarding data to generate information cascade data, and call a pre-trained social network structure data model to extract features from the social network structure data to obtain user social features; A prediction module, used to select different seed sets from the information cascade data, and input the seed sets and the user social features into the information propagation model for propagation prediction, so as to obtain propagation probability matrices corresponding to different seed sets; A generation module is used to determine the comprehensive benefit of privacy propagation according to the propagation probability matrix, and generate a privacy information propagation control strategy based on a seed set with the maximum comprehensive benefit of privacy propagation.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any of the above methods for generating a privacy information dissemination control strategy is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for generating a privacy information dissemination control strategy as described in any one of the above is implemented.

[0017] The present invention also provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, the method for generating a privacy information dissemination control strategy as described above is implemented.

[0018] The method and device for generating a privacy information dissemination control strategy provided by the present invention obtains the user's information forwarding data and social network structure data, combines the extracted user social features and constructed information cascade data, calculates the comprehensive benefits of privacy dissemination, and formulates a privacy information dissemination control strategy while considering the maximum comprehensive benefits of privacy dissemination. It not only realizes the automated formulation of privacy policies and improves the efficiency of privacy policy formulation. It can also formulate a dissemination control strategy with the highest privacy dissemination benefits in combination with the user's social features, avoids possible risks of privacy leakage, and better meets the needs of different users. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of the method for generating a privacy information propagation control strategy provided by the present invention.

[0021] Figure 2 It is a structural schematic diagram of the privacy information dissemination control strategy generation device provided by the present invention.

[0022] Figure 3 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Combine the following Figure 1-Figure 2The invention describes a method and device for generating a privacy information dissemination control strategy. Figure 1 is a flow chart of the method for generating a privacy information dissemination control strategy provided by the present invention, such as Figure 1 As shown, the method includes the following steps 101 to 104.

[0025] Step 101: Obtain user information forwarding data and social network structure data.

[0026] First, in a social network platform, the social network platform can be a social software or a social website, such as a Weibo or a Tieba. When a user disseminates information, such as forwarding information, publishing information, or sharing information, the user's information forwarding data and social network structure data are obtained at this time. Among them, the information forwarding data specifically includes data items such as forwarding time, ID of the forwarding object (other users), and forwarding events (content). Social network structure data can be constructed based on the user's social relationship graph in the social network platform. For example, the user's social relationship graph can be determined through information such as the user's circle of friends, other user object lists (friend lists), etc., and represented using a graph data structure to characterize the social relationship between the current user and other user objects.

[0027] Step 102: aggregate and analyze the information forwarding data to generate information cascade data, and call a pre-trained social network structure data model to extract features from the social network structure data to obtain user social features.

[0028] Next, the information forwarding data is aggregated and analyzed to generate information cascade data, which is a sequence of forwarding events arranged in chronological order, where each forwarding event is described by the ID and forwarding time of the forwarding user object. Aggregation analysis can be indicator aggregation, which extracts the corresponding indicators such as the ID, forwarding time, and forwarding event of the forwarding user object from the information forwarding data. These indicators can be used as corresponding seed nodes, which will be explained below.

[0029] As for the social network structure data, the pre-trained social network structure data model is called to extract features from the social network structure data to obtain user social features. The social network structure data model here is actually a variational auto encoder model (VAE), which can capture the potential characteristics of user social homogeneity contained in the social network, such as user intimacy, interest similarity and other information, and extract accurate user social feature representations for the user's social network structure data. In order to extract user social features more accurately, the social network structure data model needs to be pre-trained in the embodiment of the present invention. The pre-trained social network structure data model is obtained by collecting a large number of data samples corresponding to the user's social network structure. The pre-training process of the social network structure data model is introduced below.

[0030] First, we obtain the data sample corresponding to the user social network structure. In the user social network structure, there are a set of nodes and edges composed of multiple users. The nodes represent users, and the edges are used to characterize the social relationships between user nodes, such as intimacy, interest similarity, etc. These social relationships can characterize the social characteristics of users. For a current user node, it has different social relationships with each other different user node. According to this rule, we can determine that the social relationship distribution of the current user satisfies the standard normal distribution. Therefore, here, we can sample the initialization potential social variables of each user node in the data sample from the standard normal distribution to obtain the potential social variables corresponding to the data sample.

[0031] Initialize latent social variables for each user node in the sampled data Then, the initialization potential social variables of each user node are They are spliced ​​together as the potential social variable corresponding to the current user, that is, the potential social variable Z corresponding to the data sample.

[0032] Next, the potential social variable Z is decoded through the graph generation neural network f_dec in the social network structure data model to obtain the reconstructed social network structure G. Here, the social variable Z is reconstructed through the graph generation neural network f_dec, that is, it is restored to the corresponding reconstructed social network structure G.

[0033] Then, the reconstructed social network structure G is encoded by the autoencoder in the social network structure data model to obtain the distribution parameters corresponding to the latent social variable Z. The distribution parameters can be specifically initialized as the latent social variable The corresponding mean or variance.

[0034] Next, we construct the joint loss value of the social network structure data model based on the reconstructed social network structure and distribution parameters. The joint loss value consists of two parts: one is the reconstruction loss of the social network structure, and the other is the KL (Kullback-Leibler) divergence loss of the distribution parameters, which are described in detail below.

[0035] For the reconstruction loss of the data sample, the reconstruction loss value of the social network structure is constructed according to the difference between the reconstructed social network structure G and the data sample. Here, the difference between the social network structure reconstructed by the model and the original user social network structure in the data sample is compared, and the reconstruction loss value is calculated based on this difference to train the social network structure reconstruction ability of the social network structure data model.

[0036] As for the KL divergence loss of the distribution parameters, the KL divergence loss value is constructed based on the difference between the distribution value of the standard normal distribution and the distribution parameter. Here, the standard normal distribution refers to the standard normal distribution that corresponds to the user social relationship in the original user social network structure of the data sample. The distribution value in this standard normal distribution is compared with the distribution parameter obtained by model encoding, and the corresponding KL divergence is calculated to measure the difference between the two distributions, thereby constructing the corresponding KL divergence loss, which is used to train the social network structure data model's ability to extract social relationship features.

[0037] Finally, the two parts of the loss are combined to train the social network structure data model, that is, the sum of the reconstruction loss value and the KL divergence loss value is used as the joint loss value of the social network structure data model, and the joint loss value is used to backpropagate in the social network structure data model to update the parameters of the social network structure data model.

[0038] During training, the corresponding iterative training rounds are preset. In each round, the joint loss value is used to perform back propagation in the social network structure data model. During back propagation, the gradient of the model is calculated according to the joint loss, and the gradient is optimized by using the gradient optimization algorithm to update the parameters of the social network structure data model. When the joint loss begins to converge in a certain iterative round or reaches the preset iterative training round, the training is stopped, and the pre-training process of the social network structure data model ends.

[0039] After training, the social network structure data model has excellent feature extraction capabilities and can accurately extract the user's social characteristics based on the input social network structure data.

[0040] Step 103: Select different seed sets from the information cascade data, and input the seed sets and user social features into the information propagation model for propagation prediction to obtain propagation probability matrices corresponding to different seed sets.

[0041] Through step 102, after extracting user social features using the pre-trained social network structure data model, the propagation effect is predicted in combination with the user social features, so different seed sets are selected from the information cascade data, and the seed sets and user social features are input into the information propagation model for propagation prediction, and the propagation probability matrix corresponding to different seed sets is obtained.

[0042] First, different seed sets are selected from the information cascade data. The selection of seed sets here adopts an iterative process. By setting the information cascade data into different seed sets, different seed sets and user social features are input into the information propagation model one by one for propagation prediction, and the propagation probability matrix corresponding to different seed sets is obtained. In the information cascade data, the information forwarding time is used as the seed node, then multiple information cascade data obtain the corresponding seed set, and different seed sets can be generated according to the selected seed node. Here, the ID of the forwarding user object or the forwarding event can also be used as the seed node, which is not limited here.

[0043] For example, the information cascade data is divided into two parts, one of which is a seed set S corresponding to the seed nodes, and the other is a candidate set C. In each iteration, a seed node v is selected from the candidate set and added to the seed set S to obtain a new seed set S+v. In this way, a different new seed set can be obtained in each iteration.

[0044] Each time a new seed set is selected, the seed set and the user's social features are input into the information propagation model for propagation prediction to obtain the propagation probability matrix corresponding to the seed set. Therefore, each time a seed set is selected through iteration, the propagation prediction can be used to obtain the propagation probability matrix corresponding to the seed set, thereby obtaining the propagation probability matrix corresponding to different seed sets.

[0045] Here, the information propagation model combines the user's social characteristics to predict different information cascade data and obtain the propagation probability of the user forwarding event. The specific network structure of the information propagation model can be a multi-layer perceptron (MLP), which can be trained by collecting a large amount of information cascade data sets and corresponding user social network structure data. The training process of the information propagation model is introduced below.

[0046] First, obtain the information cascade data set and the corresponding user social network structure data, then call the pre-trained social network structure data model, for example, the pre-trained social network structure data model used in step 102, and extract the corresponding training social variables from the user social network structure data, that is, extract the potential social variables corresponding to the user social network structure data. This training social variable is used to help the information propagation model predict the propagation nature.

[0047] Next, the information cascade data set is preprocessed. Considering the limited number of information cascade data sets, data expansion is required, and corresponding data labels need to be constructed. Therefore, in the embodiment of the present invention, each piece of information cascade data in the information cascade data set is cut according to the preset number of seed nodes to obtain an expanded data set.

[0048] Specifically, firstly, a plurality of different seed node numbers k are preset as the standard for data cutting. Then, for each information cascade data with a length of n in the information cascade data set, it is cut into a seed sequence with a length of k and a label data Label with a length of nk according to the seed node number k. Among them, k is an integer greater than 1 and less than n, and n is an integer greater than 2.

[0049] For example, the preset number of seed nodes k is 3, and the length n of each piece of information cascade data in the information cascade data set is 5. Then, according to the number of nodes 3, the information cascade data of length 5 is split into a seed sequence of length 3 and a label data of length 2.

[0050] Then the seed sequence and the label data are combined into corresponding cascade samples. In this way, each time the number of seed nodes k is set, a cascade sample is obtained, and one piece of information cascade data can be expanded to obtain n-1 cascade samples. Finally, the different cascade samples obtained by cutting each piece of information cascade data are combined into an expanded data set, and each cascade sample in the expanded data set carries label data.

[0051] After the data set is expanded, the next step is the training process. Here, the information propagation model is called to encode each cascade sample in the expanded data set to obtain the time-influence variable , and then the time-dependent variables Concatenate with training social variables to get input features , and the input features Input to the information propagation model, through the information propagation model to input features Make a prediction and get the predicted propagation probability corresponding to each cascade sample.

[0052] Specifically, the input feature The input is sent to a multi-layer perceptron, which consists of multiple fully connected layers. Each fully connected layer processes the input features through a linear transformation function and a nonlinear activation function ReLU, and the last fully connected layer outputs the corresponding predicted propagation probability. , represents the propagation probability matrix corresponding to each cascade sample.

[0053] Finally, according to the predicted propagation probability Cross entropy loss value with label data Label Here, the actual propagation probability and predicted propagation probability of the label data Label of each cascade sample are calculated. , calculate the cross entropy loss value, expressed as the following formula (1): (1)

[0054] In the above formula (1), N represents the total number of cascade samples in the expanded data set, i represents the i-th cascade sample, represents the predicted propagation probability of the i-th cascade sample, Indicates the true propagation probability corresponding to the label data Label of the i-th cascade sample.

[0055] During training, the cross entropy loss value Back propagation is performed in the information propagation model to update the parameters of the information propagation model. Specifically, the corresponding iterative training rounds are first preset, and in each round, the cross entropy loss value is used to Back propagation is performed in the information propagation model. During back propagation, the cross entropy loss value is used. Compute the gradient of the information propagation model , and the gradient optimization algorithm is used to optimize the gradient, which is expressed as the following formula (2): (2)

[0056] In the above formula (2), t represents the tth training round, represents the learning rate, represents the gradient optimization parameter, Represents the cross entropy loss value.

[0057] In each iteration, the cross entropy loss is calculated by the output predicted propagation probability, and then the gradient is calculated by the cross entropy loss, and the gradient is optimized to update the parameters of the social network structure data model. When the cross entropy loss begins to converge in a certain iteration or reaches the preset iteration training round, the training stops.

[0058] The trained information propagation model can combine the user's social characteristics and input into the information cascade data to accurately predict the propagation probability. Therefore, according to the selected different seed sets, the user's social characteristics can be combined to accurately predict the propagation probability matrix corresponding to different seed sets, so as to formulate the corresponding propagation strategy later.

[0059] Step 104: Determine the comprehensive benefit of privacy propagation according to the propagation probability matrix, and generate a privacy information propagation control strategy based on the seed set with the largest comprehensive benefit of privacy propagation.

[0060] In step 103, after the propagation probability matrix corresponding to the information cascade data is predicted by the information propagation model, the comprehensive benefit of privacy propagation is determined based on the propagation probability matrix. Here, the comprehensive benefit of privacy propagation is the benefit result obtained by comprehensively considering the social influence benefit and privacy risk loss. First, based on the propagation probability matrix, the influence benefit of each seed node in the seed set is calculated to obtain the influence benefit corresponding to the seed set. , this influence benefit is the propagation benefit obtained by the current user after forwarding the forwarding event corresponding to each seed node in the seed set to other user objects. Finally, the influence benefit calculated by the seed set is expressed as the following formula (3): (3)

[0061] In the above formula (3), n represents the total number of seed nodes in the seed set s, V represents the set of all users in the user's social network structure data, represents the i-th user in V. represents the kth seed node in the seed set s, It is calculated based on the propagation probability matrix and represents the influence gain obtained by propagating the kth seed node in the seed set s to the i-th user in V. It represents the influence gain obtained after propagating the kth seed node in the seed set s.

[0062] On the other hand, the privacy risk loss of each seed node in the seed set is calculated based on the propagation probability matrix to obtain the privacy risk loss corresponding to the seed set. . Because there may be some blacklisted users in the user's social relationship, if information is forwarded to these blacklisted users, there will inevitably be corresponding privacy leakage risks. Therefore, when forwarding information, it is also necessary to consider the risk loss that may be caused by privacy leakage. Therefore, the privacy risk loss here is the privacy risk loss value that may exist after the current user forwards the forwarding event corresponding to each seed node in the seed set to the blacklisted user node. Finally, the privacy risk loss calculated by the seed set is expressed as the following formula (4): (4)

[0063] In the above formula (3), n represents the total number of all users in the user's social network structure data. Represents the social network structure data of the user, the i-th user among all users. represents the kth seed node in the seed set s, represents the privacy risk loss value that may be generated by propagating the forwarding event corresponding to the seed node to the i-th user, B represents the set of blacklisted users, represents the bth blacklist user in the blacklist user set, It is calculated based on the propagation probability matrix, which specifically indicates the probability that when a forwarding event is propagated to the i-th user, it may be propagated to the b-th blacklist user.

[0064] When disseminating information, it is necessary to comprehensively consider the loss value caused by influence and the risk loss caused by privacy leakage. Because some users are more interested in gaining influence and are not very concerned about the possible privacy risks, while some users are too concerned about the security of private information and are not very concerned about gaining influence. Therefore, it is necessary to set corresponding weights according to the user's privacy wishes to weigh the two. Here, the preset benefit weight (1- ) and loss weights, which are based on non-income weights To calculate. Rather than the income weight It is determined according to the scale of blacklist users. Here, the number of users in all user sets in the user social network structure data is calculated, and the number of blacklist users in the blacklist user set is calculated to obtain the corresponding ratio, and then the ratio and non-benefit weight are calculated. The product of , and thus the loss weight is obtained.

[0065] Next, the influence benefit is weighted by the benefit weight to obtain the benefit weighted value, and the privacy risk loss is weighted by the loss weight to obtain the loss weighted value. The difference between the benefit weighted value and the loss weighted value is taken as the comprehensive benefit of privacy dissemination, denoted as , expressed as the following formula (5): (5)

[0066] In the above formula (5), G represents the number of users in the user social network structure data of all users, B represents the number of blacklisted users in the blacklisted user set, represents the non-revenue weight, represents the income weight, represents the influence benefit, which is calculated by the above formula (3). represents the privacy risk loss, which is calculated by the above formula (4).

[0067] Through the above steps, the corresponding privacy propagation comprehensive benefit can be calculated for each different seed set selected iteratively, and finally a privacy information propagation control strategy is generated based on the seed set with the largest privacy propagation comprehensive benefit. When the privacy propagation comprehensive benefit is the largest, the user's privacy needs can be best met. At this time, according to the seed nodes in the seed set with the largest privacy propagation comprehensive benefit, the corresponding factor information propagation strategy is generated, that is, the relevant information is propagated according to all the seed nodes in the seed set.

[0068] For example, if the seed node is the forwarding time, the corresponding forwarding event will be forwarded to the forwarding user object ID according to the corresponding forwarding time. If the seed node is the ID of the forwarding user object, the corresponding forwarding event will be forwarded according to the forwarding user object with the corresponding ID within the forwarding time. If the seed node is the forwarding event, the corresponding forwarding event will be forwarded to the forwarding user object ID within the forwarding time.

[0069] In this way, based on the user's information forwarding data and social network structure data, a privacy information dissemination control strategy that maximizes the comprehensive benefits of user privacy dissemination can be formulated, and then users can be reminded or recommended on the corresponding social network platform. When users are disseminating information, they can perform corresponding information dissemination operations based on reminders or recommendations.

[0070] The embodiments of the present invention can not only automatically predict the privacy dissemination benefits and reduce the burden of privacy management on users, but also perform quantitative analysis and formulate a dissemination control strategy with the highest privacy dissemination benefits based on user social characteristics and user privacy wishes, thereby improving the efficiency of privacy strategy formulation, avoiding possible privacy leakage risks, better meeting users' privacy information needs, and enhancing their social value in social networks.

[0071] In other social network scenarios, such as the Twitter platform, since users may need to spread information multiple times, and different information spreads have a certain order of precedence, different information spread orders may generate different privacy risks. Considering that the user social network structure data model and the information spread model are trained separately using different data, there may be limitations in separately training the models to predict privacy risks. That is, the user social features extracted by the social network structure data model may not be suitable for the information spread model to predict multiple information cascade data with time series features, resulting in poor results of the generated privacy policy, and the separate model training process is cumbersome.

[0072] Based on the needs of the above social network scenarios such as the Twitter platform, an embodiment of the present invention also provides a privacy policy generation model based on a joint attention fusion network to generate a propagation probability matrix, that is, the propagation probability matrix obtained in the above step 103 can also be predicted by a privacy policy generation model based on a joint attention fusion network. The privacy policy generation model specifically includes a social feature extraction model and a joint attention fusion network model. The social feature extraction model is specifically used to extract user social features, and the joint attention fusion network model is specifically used to predict the propagation probability matrix. The training process of the privacy policy generation model is introduced below.

[0073] First, we obtain a user forwarding information dataset and a user social network structure sample. Here, we can collect various user forwarding information in real time on the Twitter platform as a user forwarding information dataset, and while collecting user forwarding information, we can extract the user's social network structure as a user social network structure sample.

[0074] After the data collection is completed, the information propagation cascade data in the user forwarding information data set is expanded to obtain the expanded information propagation cascade data set. Here, the data expansion process can refer to the data set expansion process during the training of the above-mentioned information propagation model, which will not be repeated here. Finally, the expanded information propagation cascade data can be obtained.

[0075] The social feature extraction model is used to positionally encode the expanded information propagation cascade dataset to obtain cascade time series features, and the social feature extraction model is used to extract features of the user's social network structure to obtain the user's predicted social features.

[0076] Specifically, the information propagation cascade data set still needs to select different seed sets first. The selection method is similar to the above step 103 and will not be repeated here. Each seed set includes multiple expanded cascade samples. Then, for each seed set, the social feature extraction model of the privacy policy generation model is used to position encode each expanded cascade sample in the seed set. The position feature of each expanded cascade sample can characterize the cascade time series features that may exist when multiple pieces of information are propagated. For the user social network structure sample, the social feature extraction model is used to extract features of the user social network structure sample to obtain the user's predicted social features.

[0077] Next is the feature concatenation process, where the cascaded time series features and the user predicted social features are concatenated to obtain the information propagation feature vector, and the information propagation feature vector is input into the joint attention fusion network model for prediction to obtain the prediction probability matrix. The joint attention fusion network model can perform information propagation prediction for the information propagation feature vector and obtain the corresponding prediction probability matrix.

[0078] Specifically, the joint attention fusion network model performs attention calculation on the input information propagation feature vector. In the network model, the attention layer can be used to perform cross-attention calculations on the cascaded temporal features and the user prediction social features in the information propagation feature vector, and then the calculated results are fused through a fusion layer (for example, a multi-layer perceptron), and finally the fusion layer outputs the corresponding prediction probability matrix.

[0079] Finally, the corresponding predicted privacy strategy is generated based on the propagation probability matrix. That is, the influence benefit and privacy risk loss of the corresponding seed set are calculated according to the predicted probability matrix, and the corresponding comprehensive privacy propagation benefit is further calculated. Then, the seed set with the largest comprehensive privacy propagation benefit is selected to formulate a predicted privacy strategy.

[0080] When constructing the loss of the privacy policy generation model, the optimal actual privacy policy can be pre-formulated as the true label based on the user forwarding information dataset and the user social network structure sample. Then, the final loss value of the privacy policy generation model is constructed based on the predicted privacy policy and the actual privacy policy. By comparison, the difference between the predicted privacy policy generated by the privacy policy generation model and the actual privacy policy is calculated to construct the corresponding final loss value. For example, the difference between the comprehensive benefits of privacy dissemination corresponding to the two strategies can be compared to construct the final loss value.

[0081] During training, the final loss value is used to back-propagate in the privacy policy generation model to update the parameters of the social feature extraction model and the joint attention fusion network model. The total number of training iterations can still be preset here. In each training iteration, the corresponding final loss value is calculated based on the generated predicted privacy policy, and then the final loss value is used to back-propagate in the privacy policy generation model. During the back-propagation process, the gradient of the model is calculated, and the gradient optimization algorithm is called to optimize the gradient, so as to simultaneously update the parameters of the social feature extraction model and the joint attention fusion network model. When a certain iteration round ends, the final loss value begins to converge or reaches the preset total number of training iterations, and the training stops.

[0082] The trained privacy policy generation model can be deployed and applied on the Twitter platform. When users need to disseminate information multiple times, it can collect multiple user forwarding information in real time, and generate a privacy policy with the maximum comprehensive privacy dissemination benefit based on the user's social network structure, and then remind or recommend it on the Twitter platform. Users can adjust the information forwarding process according to the privacy policy.

[0083] In the embodiment of the present invention, even when a user spreads information multiple times on certain social platforms, such as the Twitter platform, a privacy policy with the highest privacy dissemination benefit can be automatically generated through a privacy policy generation model. Thus, a dissemination control strategy with the highest privacy dissemination benefit is formulated in combination with the user's social characteristics and the user's privacy intention, thereby improving the efficiency of privacy policy formulation, avoiding possible privacy leakage risks, meeting the privacy information needs of different users, and enhancing the social value in social networks. Moreover, only a single privacy policy generation model is used for prediction, which not only reduces the difficulty of model training, but also effectively improves the generation effect of privacy policies.

[0084] The following is a description of a device for generating a private information propagation control strategy provided by the present invention. The device for generating a private information propagation control strategy described below and the method for generating a private information propagation control strategy described above can be referred to in correspondence with each other.

[0085] like Figure 2 As shown, the device for generating a privacy information dissemination control strategy of the present invention includes the following modules: an acquisition module 201, an extraction module 202, a prediction module 203, and a generation module 204. Specifically, the acquisition module 201 is used to obtain the user's information forwarding data and social network structure data; the extraction module 202 is used to perform aggregation analysis on the information forwarding data, generate information cascade data, and call the pre-trained social network structure data model to extract features from the social network structure data to obtain user social features; the prediction module 203 is used to select different seed sets from the information cascade data, and input the seed sets and the user social features into the information dissemination model for dissemination prediction to obtain the dissemination probability matrix corresponding to different seed sets; the generation module 204 is used to determine the comprehensive benefit of privacy dissemination according to the dissemination probability matrix, and generate a privacy information dissemination control strategy based on the seed set with the maximum comprehensive benefit of privacy dissemination.

[0086] It should be noted that the beneficial effects of the privacy information propagation control strategy generation device here and the privacy information propagation control strategy generation method mentioned above can correspond to each other, so the beneficial effects of the privacy information propagation control strategy generation device are not repeated here.

[0087] Figure 3 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 3As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 320 and the memory 330 communicate with each other through the communication bus 340. The processor 310 may call the logic instructions in the memory 330 to execute the method for generating a privacy information propagation control strategy, which includes: obtaining information forwarding data and social network structure data of the user; performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to extract features of the social network structure data to obtain user social features; selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into the information propagation model for propagation prediction to obtain propagation probability matrices corresponding to different seed sets; determining the comprehensive benefit of privacy propagation according to the propagation probability matrix, and generating a privacy information propagation control strategy based on the seed set with the largest comprehensive benefit of privacy propagation.

[0088] In addition, the logic instructions in the above-mentioned memory 330 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0089] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the privacy information dissemination control strategy generation method provided by the above methods, which method includes: obtaining user information forwarding data and social network structure data; performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to perform feature extraction on the social network structure data to obtain user social features; selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into the information dissemination model for dissemination prediction to obtain the dissemination probability matrices corresponding to different seed sets; determining the comprehensive benefit of privacy dissemination according to the dissemination probability matrix, and generating a privacy information dissemination control strategy based on the seed set with the largest comprehensive benefit of privacy dissemination.

[0090] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the privacy information dissemination control strategy generation method provided by the above-mentioned methods, the method comprising: obtaining user information forwarding data and social network structure data; performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to perform feature extraction on the social network structure data to obtain user social features; selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into the information dissemination model for dissemination prediction to obtain the dissemination probability matrices corresponding to different seed sets; determining the comprehensive benefit of privacy dissemination according to the dissemination probability matrix, and generating a privacy information dissemination control strategy based on the seed set with the largest comprehensive benefit of privacy dissemination.

[0091] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0092] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a privacy information dissemination control strategy, characterized in that: The method comprises: Obtain user information forwarding data and social network structure data; Performing aggregation analysis on the information forwarding data to generate information cascade data, and calling a pre-trained social network structure data model to perform feature extraction on the social network structure data to obtain user social features; Selecting different seed sets from the information cascade data, and inputting the seed sets and the user social features into an information propagation model for propagation prediction, to obtain propagation probability matrices corresponding to different seed sets; The comprehensive benefit of privacy propagation is determined according to the propagation probability matrix, and a privacy information propagation control strategy is generated based on a seed set with the maximum comprehensive benefit of privacy propagation.

2. The method for generating a privacy information dissemination control strategy according to claim 1, characterized in that: The determining of the comprehensive benefit of privacy propagation according to the propagation probability matrix includes: Calculate the influence benefit of each seed node in the seed set based on the propagation probability matrix to obtain the influence benefit corresponding to the seed set; Calculate the privacy risk loss of each seed node in the seed set based on the propagation probability matrix to obtain the privacy risk loss corresponding to the seed set; Preset a benefit weight and a loss weight, weight the influence benefit by the benefit weight to obtain a benefit weighted value, and weight the privacy risk loss by the loss weight to obtain a loss weighted value; The difference between the weighted value of the gain and the weighted value of the loss is taken as the comprehensive benefit of privacy dissemination.

3. The method for generating a privacy information dissemination control strategy according to claim 1, characterized in that: The pre-training process of the social network structure data model includes: Obtain data samples corresponding to the user's social network structure; Sampling the initialized potential social variables of each user node in the data sample from a standard normal distribution to obtain the potential social variables corresponding to the data sample; Decoding the latent social variables through a graph-generating neural network in the social network structure data model to obtain a reconstructed social network structure; Encoding the reconstructed social network structure through an autoencoder in the social network structure data model to obtain distribution parameters corresponding to the potential social variables; A joint loss value of the social network structure data model is constructed according to the reconstructed social network structure and the distribution parameters, and the joint loss value is used to perform back propagation in the social network structure data model to update the parameters of the social network structure data model.

4. The method for generating a privacy information dissemination control strategy according to claim 3, characterized in that: The step of constructing a joint loss value of the social network structure data model according to the reconstructed social network structure and the distribution parameter includes: Constructing a reconstruction loss value of the social network structure according to the difference value between the reconstructed social network structure and the data sample; Constructing a KL divergence loss value according to the difference between the distribution value of the standard normal distribution and the distribution parameter; The sum of the reconstruction loss value and the KL divergence loss value is used as the joint loss value of the social network structure data model.

5. The method for generating a privacy information dissemination control strategy according to claim 1, characterized in that: The training process of the information propagation model includes: Obtain information cascade data sets and corresponding user social network structure data; Calling a pre-trained social network structure data model to extract corresponding training social variables from the user social network structure data; Cut each piece of information cascade data in the information cascade data set according to a preset number of seed nodes to obtain an expanded data set, wherein each cascade sample in the expanded data set carries label data; Calling the information propagation model to encode each cascade sample in the expanded data set to obtain a time impact variable; The training social variable and the time influence variable are concatenated to obtain input features, and the input features are predicted by the information propagation model to obtain a predicted propagation probability corresponding to each cascade sample; A cross entropy loss value is constructed based on the predicted propagation probability and the label data, and back propagation is performed in the information propagation model through the cross entropy loss value to update the parameters of the information propagation model.

6. The method for generating a privacy information dissemination control strategy according to claim 5, characterized in that: The step of cutting each piece of information cascade data in the information cascade data set according to the preset number of seed nodes to obtain an expanded data set includes: Preset multiple different seed node numbers k; For each piece of information cascade data with a length of n in the information cascade data set, cut it into a seed sequence with a length of k and label data with a length of nk according to the number of seed nodes k, wherein k is an integer greater than 1 and less than n, and n is an integer greater than 2; The seed sequence and the label data are combined into corresponding cascade samples, and different cascade samples obtained by cutting each piece of information cascade data are combined into an expanded data set.

7. The method for generating a privacy information dissemination control strategy according to claim 1, characterized in that: The propagation probability matrix is ​​predicted by a privacy policy generation model based on a joint attention fusion network. The privacy policy generation model includes a social feature extraction model and a joint attention fusion network model. The training process of the privacy policy generation model includes: Obtain user forwarding information dataset and corresponding user social network structure samples; Performing data expansion on the information propagation cascade data in the user forwarding information data set to obtain an expanded information propagation cascade data set; Position encoding is performed on the expanded information propagation cascade data set through the social feature extraction model to obtain cascade time series features, and feature extraction is performed on the user social network structure sample through the social feature extraction model to obtain user predicted social features; The cascade time series feature is concatenated with the user prediction social feature to obtain an information propagation feature vector, and the information propagation feature vector is input into the joint attention fusion network model for prediction to obtain a prediction probability matrix; generating a corresponding prediction privacy policy based on the propagation probability matrix; A final loss value of the privacy policy generation model is constructed according to the predicted privacy policy and the actual privacy policy, and the final loss value is used for backpropagation in the privacy policy generation model to update the parameters of the social feature extraction model and the joint attention fusion network model.

8. A device for generating a privacy information dissemination control strategy, characterized in that: The device comprises: An acquisition module is used to acquire user information forwarding data and social network structure data; An extraction module, used to aggregate and analyze the information forwarding data to generate information cascade data, and call a pre-trained social network structure data model to extract features from the social network structure data to obtain user social features; A prediction module, used to select different seed sets from the information cascade data, and input the seed sets and the user social features into the information propagation model for propagation prediction, so as to obtain propagation probability matrices corresponding to different seed sets; A generation module is used to determine the comprehensive benefit of privacy propagation according to the propagation probability matrix, and generate a privacy information propagation control strategy based on a seed set with the maximum comprehensive benefit of privacy propagation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for generating a privacy information dissemination control strategy as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for generating a privacy information dissemination control strategy as described in any one of claims 1 to 7 is implemented.

Images