Intelligent fusion analysis method for cross-platform user portraits with enhanced privacy calculation

Abstract

The invention discloses a privacy calculation enhanced cross-platform user portrait intelligent fusion analysis method, which comprises the following steps that: each platform collects behavior characteristic data of a user, constructs a respective encrypted characteristic matrix based on the behavior characteristic data, and stores the encrypted characteristic matrix in a database; receiving a fusion analysis request, matching the initial encryption feature matrix from the database, and obtaining associated encryption feature matrixes in other databases based on feature alignment; tensor product operation is carried out on the initial encryption feature matrix and the associated encryption feature matrix, and a fusion feature matrix is obtained; decrypting and deducing the fused feature matrix, and obtaining a user portrait; complex interaction of multi-platform features can be captured through tensor product operation, an accurate user portrait containing high-order association is generated in combination with decryption and deduction on the premise of ensuring data privacy, the bottleneck of the traditional technology in feature interaction depth and privacy protection strength is broken through, and the user experience is improved. And a safe and efficient technical path is provided for cross-industry and cross-field data fusion.

Description

Technical Field

[0001] The present invention relates to the field of information processing technologies, and in particular to a privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method. Background Art

[0002] With the rapid development of mobile Internet and Internet of Things technologies, users have generated massive amounts of behavioral data across multiple terminals and scenarios, including social platforms, e-commerce platforms, information platforms, and financial platforms. This data includes multi-dimensional features such as browsing history, transaction information, social interactions, and device fingerprints. Building accurate user portraits has become a core requirement for Internet companies to achieve refined operations, personalized services, and risk management. By integrating fragmented user behavioral data and extracting user demographic attributes, interest preferences, spending power, and other labels, companies can significantly improve marketing conversion rates, optimize product design, and reduce operating costs. For example, e-commerce platforms can accurately recommend products that meet users' potential needs by combining their social behavior and consumption records; financial institutions can more efficiently assess credit risks by integrating users' consumption habits and credit records. However, user data is stored in a dispersed manner across different platforms, forming data silos. User portraits on a single platform are difficult to fully portray the user's full picture. Cross-platform data integration has become a key path to breaking through the bottleneck of portrait accuracy.

[0003] Current cross-platform data fusion faces multiple challenges: first, the direct circulation of raw data can easily lead to privacy leaks. For example, if user interest tags on social platforms are directly integrated with consumption records on e-commerce platforms without protection, the user's sensitive behavior trajectory may be exposed; second, the user identification systems of different platforms are independent and lack public identification, making it difficult to directly associate the behavioral data of the same user on different platforms. At the same time, the feature dimensions, data formats and semantic definitions of each platform vary significantly, making it difficult to directly align the feature space; finally, the deep correlation of users' cross-platform behaviors is mostly achieved by superimposing matrices through splicing, simple weighting and other methods. It can only handle linear relationships and cannot capture complex nonlinear correlations, resulting in low efficiency in high-order feature fusion. Summary of the Invention

[0004] In view of this, the present invention proposes a privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method, which can efficiently realize cross-platform data fusion and obtain accurate user portraits while meeting strict privacy protection requirements.

[0005] The technical solution of the present invention is achieved as follows:

[0006] A privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method includes the following steps:

[0007] Step S1: Each platform collects user behavior feature data, builds its own encrypted feature matrix based on the behavior feature data, and stores it in the database;

[0008] Step S2: receiving a fusion analysis request, matching an initial encrypted feature matrix from a database, and obtaining associated encrypted feature matrices in other databases based on feature alignment;

[0009] Step S3: performing a tensor product operation on the initial encrypted feature matrix and the associated encrypted feature matrix to obtain a fused feature matrix;

[0010] Step S4: decrypt and deduce the fused feature matrix to obtain the user portrait.

[0011] Preferably, the specific steps of step S1 include:

[0012] Step S11: Each platform collects users' clicks, browsing, transactions, and social interactions on the platform as raw data;

[0013] Step S12: After removing duplicates, filling missing values, and filtering outliers from the original data, the original data is divided into numerical features and categorical features;

[0014] Step S13: Standardize and vectorize the numerical features and categorical features respectively, encrypt the processed features, combine them into an encrypted feature matrix, and store them in the database of each platform.

[0015] Preferably, the specific steps of step S13 are:

[0016] Step S131: normalize the numerical features and use a deep learning model to map the categorical features into low-dimensional dense vectors;

[0017] Step S132: encrypt the normalized numerical features and each element in the low-dimensional dense vector using a homomorphic encryption method;

[0018] Step S133: Combine the encrypted numerical features and low-dimensional dense vectors into an encrypted feature matrix and store it in the database of each platform.

[0019] Preferably, the specific steps of step S2 include:

[0020] Step S21: When any platform receives a fusion analysis request, it parses the user information contained in the fusion analysis request and matches the initial encrypted feature matrix from the current platform database based on the user information.

[0021] Step S22: determine whether the user information contains the user public identifier. If so, use an explicit alignment method to find the associated encryption feature matrix; if not, use an implicit alignment method to find the associated encryption feature matrix.

[0022] Preferably, the specific steps of using the explicit alignment method to find the associated encrypted feature matrix are: querying the device fingerprint or hashed ID from the user information as the user's public identifier, querying based on the user's public identifier in other platforms, and outputting the encrypted feature matrix of the corresponding user in the database as the associated encrypted feature matrix.

[0023] Preferably, when the user information or other platforms do not contain the user's public identification, the specific steps of using the implicit alignment method to find the associated encrypted feature matrix are: mapping the user's behavioral feature data to the public semantic space, calculating the cosine similarity between the user's behavioral feature data corresponding to the initial encrypted feature matrix and the behavioral feature data of users on other platforms, obtaining the encrypted feature matrix with the highest similarity in other platforms based on the cosine similarity, and outputting it as the associated encrypted feature matrix.

[0024] Preferably, the specific steps of step S3 include:

[0025] Step S31: homomorphically multiply the elements in the initial encryption feature matrix and the associated encryption feature matrix;

[0026] Step S32: Arrange all element-by-element product results into a block matrix to form an encrypted tensor;

[0027] Step S33: Flatten the multidimensional encrypted tensor into a two-dimensional matrix form, and obtain the fusion feature matrix while keeping the order of elements unchanged.

[0028] Preferably, the step S31 performs relinearization or analog-to-digital switching after each homomorphic multiplication.

[0029] Preferably, the specific steps of step S4 include:

[0030] Step S41: Based on the preset decryption key shards, the fused feature matrix is ​​collaboratively decrypted through a secure multi-party computing protocol to obtain a plaintext feature matrix;

[0031] Step S42: reconstruct the plaintext feature matrix into a user-feature two-dimensional table, and perform inverse transformation and feature restoration on the numerical features and categorical features in the user-feature two-dimensional table respectively;

[0032] Step S43: Generate several user key tags based on the results of the reverse transformation and feature restoration, and generate a user clustering heat map based on the user key tags for display.

[0033] Preferably, before displaying the user clustering heat map, step S43 verifies whether the user key tags contain sensitive information and performs hierarchical desensitization processing on the sensitive information.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention provides a privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method. After each platform collects the user's behavioral feature data, it constructs its own encrypted feature matrix and stores it. The encrypted feature matrix in the database of each platform can query the associated encrypted feature matrix in other platforms. When receiving a fusion analysis request, the corresponding associated encrypted feature matrix can be found, and then the tensor product operation is performed on the initial encrypted feature matrix and the associated encrypted feature matrix to capture the complex interaction of multi-platform features. Finally, by decrypting and deducing the fused feature matrix, an accurate user portrait containing high-order correlations can be generated under the premise of ensuring data privacy. This breaks through the bottleneck of traditional technology in feature interaction depth and privacy protection strength, and provides a safe and efficient technical path for cross-industry and cross-field data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 This is a flowchart of a privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method of the present invention;

[0038] Figure 2 This is a flowchart of step S1 of the privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method of the present invention;

[0039] Figure 3 This is a flowchart of step S13 of the privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method of the present invention;

[0040] Figure 4 This is a flowchart of step S2 of the privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method of the present invention;

[0041] Figure 5 This is a flowchart of step S3 of the privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method of the present invention;

[0042] Figure 6This is a flowchart of step S4 of the privacy computing enhanced cross-platform user portrait intelligent fusion analysis method of the present invention. DETAILED DESCRIPTION

[0043] In order to better understand the technical content of the present invention, a specific embodiment is provided below, and the present invention is further described in conjunction with the accompanying drawings.

[0044] See also Figures 1 to 6 The present invention provides a privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method, which includes the following steps:

[0045] Step S1: Each platform collects user behavior feature data, builds its own encrypted feature matrix based on the behavior feature data, and stores it in the database;

[0046] Step S2: receiving a fusion analysis request, matching an initial encrypted feature matrix from a database, and obtaining associated encrypted feature matrices in other databases based on feature alignment;

[0047] Step S3: performing a tensor product operation on the initial encrypted feature matrix and the associated encrypted feature matrix to obtain a fused feature matrix;

[0048] Step S4: decrypt and deduce the fused feature matrix to obtain the user portrait.

[0049] With the rapid development of mobile Internet, a variety of online platforms are emerging in an endless stream. For individual users, the number of platforms they use is also increasing. For ordinary users, the platforms they use daily have evolved from early single social or shopping tools to a multi-platform combination covering instant messaging, content consumption, and online transactions. In the fields of financial management, health management, vocational education, etc., the platform usage scenarios of deep users have further expanded to vertical fields such as bank apps, medical consultation platforms, and professional knowledge payment platforms, forming cross-domain, multi-dimensional digital behavior trajectories. Different platforms always maintain relative independence. In order to improve user experience and operational efficiency, different platforms usually build user portraits based on their own data. By collecting user browsing, clicks, transactions, social interactions and other behavioral data, they use machine learning algorithms to generate a label system that includes demographic attributes, interest preferences, and consumption tendencies, thereby achieving personalized recommendations.

[0050] However, based on business independence and data security considerations, each platform has generally built a closed data management system, resulting in the formation of data islands for user behavior data on different platforms. For example, user interest tags recorded by social platforms, content consumption preferences captured by short video platforms, and consumption capacity data accumulated by e-commerce platforms are all stored in separate platforms. There is a lack of efficient and secure association mechanisms between them. Although the user portrait of a single platform can improve service accuracy in local scenarios, it is difficult to fully portray the user's real needs and potential characteristics due to the lack of correlation analysis of cross-platform behavior data. For example, on social platforms, users show a continuous focus on new energy vehicles, but frequently browse auto parts on e-commerce platforms. The user portrait of a single platform may label the user as a general interest user or an auto parts consumer. However, if the user portraits of the two platforms can be combined, the user can be accurately positioned as a potential buyer of new energy vehicles. The present invention can achieve the above-mentioned fusion of cross-platform user portraits. Before constructing the user portrait, based on privacy protection For the purpose of this, each platform collects the behavioral feature data of the user, builds its own encrypted feature matrix based on the behavioral feature data, and stores the encrypted feature matrix in its own database. That is, each platform builds its own database to store the user's encrypted feature matrix. When a fusion analysis request is received, the initial encrypted feature matrix and the associated encrypted feature matrix can be obtained from the database. The initial encrypted feature matrix is ​​obtained by the current platform based on user information matching, and the associated encrypted feature matrix is ​​obtained from other platforms by feature alignment. After obtaining the encrypted feature matrix and the associated encrypted feature matrix, the encrypted feature matrix and the associated encrypted feature matrix are tensor-producted to capture the complex interaction of multi-platform features and obtain a fusion feature matrix. Finally, after decrypting and deducing the fusion feature matrix, the user portrait can be accurately obtained, which can significantly improve the privacy protection in the process of cross-platform user data fusion, and provide a safe and efficient technical path for cross-industry and cross-field data fusion.

[0051] Preferably, the specific steps of step S1 include:

[0052] Step S11: Each platform collects users' clicks, browsing, transactions, and social interactions on the platform as raw data;

[0053] Step S12: After removing duplicates, filling missing values, and filtering outliers from the original data, the original data is divided into numerical features and categorical features;

[0054] Step S13: Standardize and vectorize the numerical features and categorical features respectively, encrypt the processed features, combine them into an encrypted feature matrix, and store them in the database of each platform.

[0055] Each platform collects users' clicks, browsing, transactions, social interactions and other behaviors on the platform as raw data in real time and in batches through point statistics, API interfaces or log capture, and then cleans the raw data. The cleaning process includes deduplication, filling missing values ​​and filtering outliers such as ultra-high frequency invalid clicks to improve data quality. After the data cleaning is completed, the raw data is classified into numerical features and categorical features. Numerical features include age, consumption amount and purchase frequency, etc. Categorical features include gender, region and preferences, etc. The dimensions of numerical features of different platforms are different, so the numerical features need to be standardized or normalized to ensure the numerical stability of features of different dimensions in subsequent tensor product operations. Categorical features need to be converted into vector expressions. Categorical features can be converted into numerical vectors through deep learning models to unify the feature representation space. After the numerical features and categorical features are decrypted, they can be combined into an encrypted feature matrix and stored in the database for use.

[0056] Preferably, the specific steps of step S13 are:

[0057] Step S131: normalize the numerical features and use a deep learning model to map the categorical features into low-dimensional dense vectors;

[0058] Step S132: encrypt the normalized numerical features and each element in the low-dimensional dense vector using a homomorphic encryption method;

[0059] Step S133: Combine the encrypted numerical features and low-dimensional dense vectors into an encrypted feature matrix and store it in the database of each platform.

[0060] Numerical features are directly standardized, while categorical features need to be vectorized. The present invention uses deep learning models such as Word2Vec or BERT to map categorical features into low-dimensional dense vectors. For example, when the categorical feature is "digital", it is converted into a low-dimensional dense vector of [0.3, 0.7, -0.2] through a deep learning model. Different vector combinations represent different categorical features. After the numerical features and categorical features are processed, the numerical features and the low-dimensional dense vectors are encrypted respectively using a homomorphic encryption method. The homomorphic encryption method generates a homomorphic encrypted public-private key pair, the public key is used for encryption, and the private key is used for decryption. The numerical features are directly encrypted by the public key, and each element in the low-dimensional dense vector is also directly encrypted by the public key. The encrypted dense vector is used as the row or column of the matrix, and after combining with the numerical features, an encrypted feature matrix can be obtained by combining the encrypted feature matrix, which is then sent to the database for storage. At the same time, the private key is uploaded to a preset third-party decryption agency for storage for subsequent decryption processes.

[0061] Preferably, the specific steps of step S2 include:

[0062] Step S21: When any platform receives a fusion analysis request, it parses the user information contained in the fusion analysis request and matches the initial encrypted feature matrix from the current platform database based on the user information.

[0063] Step S22: determine whether the user information contains the user public identifier. If so, use an explicit alignment method to find the associated encryption feature matrix; if not, use an implicit alignment method to find the associated encryption feature matrix.

[0064] In the daily operation and maintenance of each platform, if it is necessary to focus on analyzing the portraits of some users or groups, a fusion analysis request can be initiated actively, and the fusion analysis request can be sent to any platform. The operation and maintenance personnel can choose a more relevant platform to send the fusion analysis request and package the user information in the fusion analysis request. The user information includes the user's IP, user ID, basic information, device information, etc. When the fusion analysis request is received by the platform, the user information can be extracted and the own database can be queried based on the user information. The encrypted feature matrix of the user is matched from the database and the encrypted feature matrix is ​​output as the initial encrypted feature matrix. The initial encrypted feature matrix is ​​part of the construction of the fusion feature matrix, and then it is necessary to By integrating the encryption feature matrices of other platforms, operation and maintenance personnel can specify the corresponding platform, or the current platform can automatically search for other platforms. When searching for the corresponding encryption feature matrix, user information needs to be aligned, including two alignment methods. The first is the explicit alignment method, and the second is the implicit alignment method. The explicit alignment method requires obtaining the user's public identifier from the user information. The user's public identifier can be used to quickly determine the corresponding content in other platforms, thereby determining the associated encryption feature matrix. If the user's public identifier is not included, the implicit alignment method needs to be used. The search principle of the privacy alignment method is to obtain it based on the correlation of user behavior. Based on the explicit alignment method and the implicit alignment method, the associated encryption feature matrix in other platforms can be quickly obtained.

[0065] Preferably, the specific steps of using the explicit alignment method to find the associated encrypted feature matrix are: querying the device fingerprint or hashed ID from the user information as the user's public identifier, querying based on the user's public identifier in other platforms, and outputting the encrypted feature matrix of the corresponding user in the database as the associated encrypted feature matrix.

[0066] The explicit alignment method requires first obtaining the public identification of the user to be fused. First, query the user information to see whether it contains the device fingerprint or hashed ID. If so, it is recorded as the user's public identification. Then, the corresponding user can be found in other platforms based on the user's public identification. If the corresponding user exists, the encrypted feature matrix of the user stored in the database can be directly output as the associated encrypted feature matrix.

[0067] Preferably, when the user information or other platforms do not contain the user's public identification, the specific steps of using the implicit alignment method to find the associated encrypted feature matrix are: mapping the user's behavioral feature data to the public semantic space, calculating the cosine similarity between the user's behavioral feature data corresponding to the initial encrypted feature matrix and the behavioral feature data of users on other platforms, obtaining the encrypted feature matrix with the highest similarity in other platforms based on the cosine similarity, and outputting it as the associated encrypted feature matrix.

[0068] However, user information may not necessarily contain device fingerprints or hashed IDs. Similarly, users on other platforms may not necessarily have device fingerprints or hashed IDs. When user information or other platforms do not contain user public identifiers, an implicit alignment method is required to find the associated encrypted feature matrix. First, the behavioral feature data of the corresponding user collected by the current platform is mapped to the public semantic space. At the same time, the behavioral feature data collected by other platforms is also mapped to the same public semantic space. Then, the similarity between the behavioral feature data of each other platform and the behavioral feature data of the current user is calculated. The calculation method uses the cosine similarity calculation method. In addition to the cosine similarity calculation method, the Pearson correlation coefficient calculation method or the Jaccard coefficient calculation method can also be used. By calculating the similarity, the correlation between the encrypted feature matrix of each platform and the initial encrypted feature matrix can be determined. Then, based on the similarity threshold, the encrypted feature matrix with the highest correlation can be determined and output as the associated encrypted feature matrix. The corresponding platform can be specified based on the operation and maintenance personnel's selection. If the operation and maintenance personnel do not specify, the encrypted feature matrices with the highest similarity can be selected for output, reducing the number of associated encrypted feature matrices and reducing the amount of computation.

[0069] Preferably, the specific steps of step S3 include:

[0070] Step S31: homomorphically multiply the elements in the initial encrypted feature matrix and the associated encrypted feature matrix, and perform relinearization or modular switching after each homomorphic multiplication;

[0071] Step S32: Arrange all element-by-element product results into a block matrix to form an encrypted tensor;

[0072] Step S33: Flatten the multidimensional encrypted tensor into a two-dimensional matrix form, and obtain the fusion feature matrix while keeping the order of elements unchanged.

[0073] After obtaining the initial encryption feature matrix and the associated encryption feature matrix, a tensor product operation needs to be performed on the initial encryption feature matrix and the associated encryption feature matrix. The tensor product operation uses the initial encryption feature matrix as a benchmark and homomorphically multiplies multiple associated encryption feature matrices with it. When each associated encryption feature matrix is ​​homomorphically multiplied with the initial encryption feature matrix, several block matrices are obtained. Each block matrix is ​​an encryption tensor. After performing homomorphic multiplication with different associated encryption feature matrices in sequence, several multidimensional encryption tensors can be obtained. The multidimensional encryption tensors are then converted into a two-dimensional matrix form and arranged according to the order of the elements of the initial encryption feature matrix to obtain a fused feature matrix. The tensor product operation retains the high-order interaction relationship between the original matrix features, thereby facilitating the capture of complex interactions of multi-platform features. Homomorphic multiplication may cause noise growth to exceed a threshold. Therefore, after each homomorphic multiplication, it is necessary to reduce the noise through relinearization or modulus switching to ensure that the ciphertext noise is controlled within the decryption threshold, thereby ensuring the correctness of subsequent operations.

[0074] Preferably, the specific steps of step S4 include:

[0075] Step S41: Based on the preset decryption key shards, the fused feature matrix is ​​collaboratively decrypted through a secure multi-party computing protocol to obtain a plaintext feature matrix;

[0076] Step S42: reconstruct the plaintext feature matrix into a user-feature two-dimensional table, and perform inverse transformation and feature restoration on the numerical features and categorical features in the user-feature two-dimensional table respectively;

[0077] Step S43: Generate several user key tags based on the results of the reverse transformation and feature restoration, and generate a user clustering heat map based on the user key tags for display.

[0078] The decryption and deduction process will not be carried out on any platform. The present invention introduces a trusted third party authorized by the platform or a decryption node agreed upon by multiple platforms for unified decryption to minimize data exposure. The trusted third party or decryption node shards according to the preset decryption key and collaboratively decrypts the fused feature matrix through a secure multi-party computing protocol to obtain a plaintext feature matrix. The plaintext feature matrix can be divided and reconstructed into a user-feature two-dimensional table according to the original feature dimension. The user-feature two-dimensional table contains numerical features and categorical features. The numerical features are reverse transformed to restore the original dimension of the features, while the categorical features are restored. For example, they are decoded into text features through a preset mapping table, or feature semantics are identified through cluster analysis. User key tags are generated based on the reverse transformed numerical features and the restored categorical features. In addition, the user's composite attributes or preference patterns can be extracted from the cross features generated by the tensor product, and user key tags are also generated. All user key tags are projected into a low-dimensional space after dimensionality reduction through t-SNE, and a user clustering heat map or tag cloud is generated to intuitively display the user feature distribution and depict an accurate user portrait.

[0079] Preferably, before displaying the user clustering heat map, step S43 verifies whether the user key tags contain sensitive information and performs hierarchical desensitization processing on the sensitive information.

[0080] The generated user key tags may contain some sensitive information, such as medical records, religious beliefs, political views, etc. This information is private information for users and is not convenient to display intuitively. Therefore, these sensitive information can be desensitized in a graded manner. Sensitive information that is completely prohibited can be directly deleted, sensitive information that is partially allowed but needs to be weakened can be blurred, and noise can be injected into some sensitive information, so as to control the risk of privacy leakage within the allowable range.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method, characterized by: The following steps are involved: Step S1: Each platform collects user behavior feature data, builds its own encrypted feature matrix based on the behavior feature data, and stores it in the database; Step S2: receiving a fusion analysis request, matching an initial encrypted feature matrix from a database, and obtaining associated encrypted feature matrices in other databases based on feature alignment; Step S3: performing a tensor product operation on the initial encrypted feature matrix and the associated encrypted feature matrix to obtain a fused feature matrix; Step S4: decrypt and deduce the fused feature matrix to obtain the user portrait.

2. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 1 is characterized in that: The specific steps of step S1 include: Step S11: Each platform collects users' clicks, browsing, transactions, and social interactions on the platform as raw data; Step S12: After removing duplicates, filling missing values, and filtering outliers from the original data, the original data is divided into numerical features and categorical features; Step S13: Standardize and vectorize the numerical features and categorical features respectively, encrypt the processed features, combine them into an encrypted feature matrix, and store them in the database of each platform.

3. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 2 is characterized in that: The specific steps of step S13 are: Step S131: normalize the numerical features and use a deep learning model to map the categorical features into low-dimensional dense vectors; Step S132: encrypt the normalized numerical features and each element in the low-dimensional dense vector using a homomorphic encryption method; Step S133: Combine the encrypted numerical features and low-dimensional dense vectors into an encrypted feature matrix and store it in the database of each platform.

4. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 1 is characterized in that: The specific steps of step S2 include: Step S21: When any platform receives a fusion analysis request, it parses the user information contained in the fusion analysis request and matches the initial encrypted feature matrix from the current platform database based on the user information. Step S22: determine whether the user information contains the user public identifier. If so, use an explicit alignment method to find the associated encryption feature matrix; if not, use an implicit alignment method to find the associated encryption feature matrix.

5. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 4 is characterized in that: The specific steps of using the explicit alignment method to find the associated encrypted feature matrix are: querying the device fingerprint or hashed ID from the user information as the user's public identifier, querying based on the user's public identifier in other platforms, and outputting the encrypted feature matrix of the corresponding user in the database as the associated encrypted feature matrix.

6. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 5 is characterized in that: When the user information or other platforms do not contain the user's public identifier, the specific steps of using the implicit alignment method to find the associated encrypted feature matrix are: mapping the user's behavioral feature data to the public semantic space, calculating the cosine similarity between the user's behavioral feature data corresponding to the initial encrypted feature matrix and the behavioral feature data of users on other platforms, obtaining the encrypted feature matrix with the highest similarity in other platforms based on the cosine similarity, and outputting it as the associated encrypted feature matrix.

7. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 1 is characterized in that: The specific steps of step S3 include: Step S31: homomorphically multiply the elements in the initial encryption feature matrix and the associated encryption feature matrix; Step S32: Arrange all element-by-element product results into a block matrix to form an encrypted tensor; Step S33: Flatten the multidimensional encrypted tensor into a two-dimensional matrix form, and obtain the fusion feature matrix while keeping the order of elements unchanged.

8. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 7 is characterized in that: The step S31 performs relinearization or analog-to-digital switching after each homomorphic multiplication.

9. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 1 is characterized in that: The specific steps of step S4 include: Step S41: Based on the preset decryption key shards, the fused feature matrix is ​​collaboratively decrypted through a secure multi-party computing protocol to obtain a plaintext feature matrix; Step S42: reconstruct the plaintext feature matrix into a user-feature two-dimensional table, and perform inverse transformation and feature restoration on the numerical features and categorical features in the user-feature two-dimensional table respectively; Step S43: Generate several user key tags based on the results of the reverse transformation and feature restoration, and generate a user clustering heat map based on the user key tags for display.

10. The privacy computing-enhanced cross-platform user portrait intelligent fusion analysis method according to claim 9, characterized in that: Before displaying the user clustering heat map, step S43 verifies whether the user key tags contain sensitive information and performs hierarchical desensitization processing on the sensitive information.

Images