Electronic commerce platform multi-source data security fusion method and system based on federated learning

Abstract

The application discloses an e-commerce platform multi-source data security fusion method and system based on federated learning, relates to the cross technical field of federated learning and data security, and comprises the following steps: collecting user behavior data, transaction data and user attribute data for preprocessing, and generating a multi-modal feature vector; based on the multi-modal feature vector, calculating the causal importance score of each multi-modal feature vector, generating a local score vector, and dynamically adjusting the privacy protection strength according to the local score vector to generate an encrypted feature vector; based on the fusion feature matrix, extracting each platform exclusive weight vector, calculating the relative contribution degree of the feature weight strength, generating a contribution proportion table, and performing zero-knowledge proof verification on the contribution proportion table to generate a data fusion audit report. The application realizes fine-grained dynamic regulation and control of the privacy protection strength, makes high-value features retain higher data utility, enhances the privacy protection of low-value features, and achieves the optimal balance between privacy protection and data precision.

Description

Technical Field

[0001] This invention relates to the field of federated learning and data security, and in particular to a method and system for the secure fusion of multi-source data on e-commerce platforms based on federated learning. Background Technology

[0002] In the field of multi-source data fusion in e-commerce platforms, federated learning technology is widely used because it enables cross-platform collaboration while ensuring data privacy. Existing solutions typically employ homomorphic encryption combined with differential privacy mechanisms: first, user behavior, transaction, and attribute data from each platform are standardized and preprocessed to generate feature vectors; then, Laplace noise is added based on a preset static privacy budget to achieve perturbation protection; finally, encrypted features are aggregated using a federated averaging algorithm. This type of method preserves data distribution characteristics through encryption operations, meeting basic privacy compliance requirements, while also providing technical support for cross-platform joint analysis by preserving data distribution characteristics through encrypted aggregation.

[0003] However, existing methods still have two limitations: First, static privacy budget allocation cannot dynamically adjust noise intensity based on feature-level causal importance, leading to the diminishing utility of high-value features due to excessive noise injection, or the risk of information leakage due to insufficient protection of low-sensitivity features; Second, the lack of a cross-platform contribution measurement mechanism based on causal inference makes it difficult to accurately quantify the relative contribution of feature weights from different platforms, affecting the fairness and interpretability of the fusion results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for secure fusion of multi-source data in e-commerce platforms based on federated learning to address the problems of imbalance between static privacy protection and feature utility, as well as inaccurate cross-platform contribution measurement.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides a method for secure fusion of multi-source data in e-commerce platforms based on federated learning. The method includes: collecting user behavior data, transaction data, and user attribute data for preprocessing to generate multimodal feature vectors; calculating the causal importance score of each multimodal feature vector based on the multimodal feature vectors to generate a local score vector, and dynamically adjusting the privacy protection strength according to the local score vector to generate an encrypted feature vector; calculating the global ATE value of each feature based on the encrypted feature vector, comparing it with a causal significance threshold, filtering the encrypted feature vectors to obtain a high-value causal feature subset; performing encrypted aggregation on the high-value causal feature subset to generate a global feature statistical vector, and performing cross-platform feature distribution alignment through federated learning to generate a fused feature matrix; extracting platform-specific weight vectors based on the fused feature matrix, calculating the relative contribution of feature weight strengths to generate a contribution ratio table, verifying the contribution ratio table with zero-knowledge proof, and generating a data fusion audit report.

[0008] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps for generating multimodal feature vectors are as follows:

[0009] The preprocessing includes data cleaning and standardization.

[0010] The preprocessed user behavior data, transaction data, and user attribute data are fused to generate multimodal feature vectors.

[0011] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps for calculating the causal importance score of each multimodal feature vector and generating a local score vector are as follows:

[0012] Modality separation is performed on the multimodal feature vectors to generate text feature sub-vectors, image feature sub-vectors, and numerical feature sub-vectors;

[0013] Based on text feature subvectors, the intervention effect value of text features on the target variable is calculated, and gradient-weighted class activation mapping is performed on image feature subvectors to extract visual causal significance. At the same time, counterfactual reasoning is performed on numerical feature subvectors to identify the average treatment effect.

[0014] By integrating intervention effect values, visual causal significance, and average treatment effects, a multimodal causal effect value set is generated.

[0015] Based on the multimodal causal effect value set, weight coefficients are assigned according to the feature modality type, the comprehensive causal importance score is calculated, and a local score vector is generated.

[0016] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps for dynamically adjusting the privacy protection strength based on the local score vector to generate the encrypted feature vector are as follows:

[0017] A score-privacy budget mapping rule is adopted to map local score vectors into feature-level privacy budget tables;

[0018] Based on the feature-level privacy budget table, a quantum true random number generator is invoked to generate Laplace noise distribution parameters and generate a privacy noise intensity vector;

[0019] Element-level noise weighting perturbation is performed on the corresponding elements of the privacy noise intensity vector and the multimodal feature vector to generate an initial noise injection feature vector;

[0020] Orthogonal random perturbations are added to the initial noise-injected feature vector, and KL divergence detection is performed to generate encrypted feature vectors.

[0021] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps of calculating the global ATE value of each feature based on the encrypted feature vector, comparing it with a causal significance threshold, filtering the encrypted feature vector, and obtaining a subset of high-value causal features are as follows.

[0022] Based on the encrypted feature vector, the global ATE value of each feature is calculated through a homomorphic encryption protocol to generate a local ATE vector;

[0023] The local ATE vector is decrypted and weighted averaged to generate a global ATE value vector. This vector is then compared with the causal significance threshold feature by feature to identify high-value features and generate a feature label vector.

[0024] The encrypted feature vector is filtered based on the feature label vector to generate a high-value causal feature subset.

[0025] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps for encrypted aggregation of high-value causal feature subsets to generate global feature statistical vectors are as follows:

[0026] Based on a subset of high-value causal features, the arithmetic mean and standard deviation of each feature are identified, and a local feature statistical vector is generated.

[0027] The local feature statistical vector is subjected to Paillier homomorphic encryption using the coordinating node to generate encrypted statistical ciphertext.

[0028] The encrypted statistical ciphertext is uploaded to the coordination node, where homomorphic accumulation and aggregation are performed to generate the encrypted global statistical ciphertext.

[0029] The coordinating node uses its private key to decrypt the encrypted global statistical ciphertext, obtains the original global statistical aggregate value, and performs standardization processing on the original global statistical aggregate value to generate a global feature statistical vector.

[0030] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the step of performing cross-platform feature distribution alignment through federated learning to generate a fused feature matrix includes the following steps:

[0031] Based on global feature statistical vectors, high-value causal feature subsets are standardized and aligned to generate aligned local feature sets.

[0032] The local feature sets are concatenated into an augmented matrix according to the feature dimensions, and then the dimensionality is reduced to a uniform dimension to generate a platform-level feature projection matrix.

[0033] The coordinating node receives the platform-level feature projection matrix and performs feature space convex combinatorial optimization based on the entropy weight method to generate a fused feature matrix.

[0034] As a preferred embodiment of the federated learning-based multi-source data secure fusion method for e-commerce platforms described in this invention, the steps of extracting platform-specific weight vectors based on the fusion feature matrix, calculating the relative contribution of feature weight strengths, and generating a contribution ratio table are as follows.

[0035] The fusion feature matrix is ​​analyzed, the exclusive weight vectors of each e-commerce platform are extracted, the Frobenius norm is calculated, and a feature weight intensity set is generated.

[0036] Based on the feature weight intensity set, relative contribution is calculated to generate a platform contribution ratio table.

[0037] As a preferred embodiment of the federated learning-based multi-source data security fusion method for e-commerce platforms described in this invention, the steps of verifying the contribution ratio table with zero-knowledge proof and generating a data fusion audit report are as follows:

[0038] The zero-knowledge proof certificate is obtained by verifying the contribution ratio table of the e-commerce platform with challenge seeds issued by the regulatory node using zero-knowledge proof.

[0039] The contribution ratio table of e-commerce platforms and zero-knowledge proof certificates are merged according to timestamps to generate a data fusion audit report.

[0040] Secondly, the present invention provides a multi-source data security fusion system for e-commerce platforms based on federated learning, including a data acquisition module, a feature encryption module, a feature filtering module, a cross-platform fusion module, and a joint audit module;

[0041] The data acquisition module is used to collect user behavior data, transaction data, and user attribute data for preprocessing and generating multimodal feature vectors.

[0042] The feature encryption module is used to calculate the causal importance score of each multimodal feature vector based on the multimodal feature vector, generate a local score vector, and dynamically adjust the privacy protection strength according to the local score vector to generate encrypted feature vectors.

[0043] The feature filtering module is used to calculate the global ATE value of each feature based on the encrypted feature vector, compare it with the causal significance threshold, filter the encrypted feature vector, and obtain a subset of high-value causal features.

[0044] The cross-platform fusion module is used to perform encrypted aggregation of high-value causal feature subsets, generate global feature statistical vectors, and perform cross-platform feature distribution alignment through federated learning to generate a fused feature matrix.

[0045] The joint audit module is used to extract the platform-specific weight vectors based on the fusion feature matrix, calculate the relative contribution of feature weight strength, generate a contribution ratio table, verify the contribution ratio table with zero-knowledge proof, and generate a data fusion audit report.

[0046] The beneficial effects of this invention are as follows: By dynamically adjusting the privacy protection strength to generate encrypted feature vectors based on local score vectors, fine-grained dynamic control of privacy protection strength is achieved, enabling high-value features to retain higher data utility and low-value features to enhance privacy protection, thus achieving an optimal balance between privacy protection and data accuracy; by calculating the global ATE value based on the encrypted feature vectors to filter high-value causal feature subsets, the optimal selection of causal-driven features within the encrypted domain is achieved, improving the business interpretability and fusion efficiency of high-value causal feature subsets. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for secure fusion of multi-source data in an e-commerce platform based on federated learning.

[0049] Figure 2 This is a schematic diagram of a multi-source data security fusion system for an e-commerce platform based on federated learning.

[0050] Figure 3 A flowchart for generating multimodal feature vectors.

[0051] Figure 4 A flowchart for generating a data fusion audit report. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for secure fusion of multi-source data in an e-commerce platform based on federated learning, including the following steps:

[0056] S1. Collect user behavior data, transaction data, and user attribute data, preprocess them, and generate multimodal feature vectors;

[0057] User behavior data includes click data, search data, and browsing time;

[0058] It should be noted that click data is generated by capturing the onClick event of user interface elements through front-end event listeners, recording the click coordinates and timestamp; search data is generated by intercepting the query string in the search box submission event, combining the session ID and search time; browsing duration is calculated by the difference between the hidden timestamp and the visible timestamp triggered by the page visibility API.

[0059] Transaction data includes transaction amount and product information;

[0060] It should be noted that the transaction amount is obtained in real time through the payment platform callback interface (such as Alipay / WeChat Pay), and the transaction record is generated by associating the user ID with the timestamp; the product information is generated by extracting the SKU code from the order database, calling the product center API to obtain details such as category / brand / price, and combining it with the purchase quantity to generate a standardized product feature vector.

[0061] User attribute data refers to basic demographic characteristics;

[0062] It should be noted that the basic demographic characteristics include age, gender, region, and membership level tags, which are obtained through user registration information, GPS location, and consumption behavior analysis, and are stored as feature vectors after being anonymized.

[0063] Preprocessing includes data cleaning and standardization.

[0064] It should be noted that data cleaning refers to completing missing timestamps and correcting abnormal dwell times in user behavior data, verifying the logical consistency between payment amounts and product prices in transaction data and filling in missing category codes, and verifying the matching of ID numbers and regions in user attribute data and standardizing date formats; standardization refers to performing Z-score standardization on numerical features (such as transaction amounts and browsing durations) after data cleaning, and performing One-Hot coding on categorical features (such as product categories);

[0065] The preprocessed user behavior data, transaction data, and user attribute data are fused using heterogeneous information to generate multimodal feature vectors;

[0066] Furthermore, based on preprocessed user behavior data, user click frequency is obtained by counting the number of clicks per unit time. Keywords are extracted from search data to identify search intent, and combined with browsing time to form a behavioral feature vector. Based on preprocessed user transaction data, transaction amount and product information are horizontally concatenated into a single vector to generate a transaction feature vector. Simultaneously, age segmentation, regional mapping, and membership level embedding are performed on preprocessed user attribute data to generate attribute feature vectors. The behavioral feature vector, transaction feature vector, and attribute feature vector are aligned and their dimensions are unified to generate dimension-aligned feature vectors. Finally, the correlation information of the dimension-aligned feature vectors is fused through a cross-modal attention mechanism to generate a multimodal feature vector.

[0067] It should be noted that the cross-modal attention mechanism is a computational method that uses a feature interaction matrix and a dynamic weight allocation algorithm to analyze the correlation between features of different modalities and dynamically allocate fusion weights, in order to improve the efficiency of multi-source information fusion and feature representation ability.

[0068] S2. Based on the multimodal feature vectors, calculate the causal importance score of each multimodal feature vector, generate a local score vector, and dynamically adjust the privacy protection strength according to the local score vector to generate an encrypted feature vector.

[0069] Modality separation is performed on the multimodal feature vectors to generate text feature sub-vectors, image feature sub-vectors, and numerical feature sub-vectors;

[0070] Based on the pre-embedded modal type labels in the multimodal feature vectors (e.g., text features are labeled "A", image features are labeled "B", and numerical features are labeled "C"), the multimodal feature vectors are separated into text feature vectors, image feature vectors, and numerical feature vectors through the type labels, generating preliminary modal sub-vector groups. The preliminary modal sub-vector groups are checked for missing values ​​to achieve data integrity verification, and sub-vectors with irregular dimensions are padded or truncated to the standard dimension length to generate standardized modal sub-vector groups. Based on the standardized modal sub-vector groups, the L2 norm of the text vectors is identified and normalized, the central region of the image feature vectors is selected and its size is adjusted, and the numerical feature vectors are linearly scaled to a uniform range according to the maximum and minimum values, finally generating completely separated text feature sub-vectors, image feature sub-vectors, and numerical feature sub-vectors.

[0071] It should be noted that modality type labels are identifiers set by associating metadata from the feature extraction stage with the corresponding dimension of the multimodal feature vector.

[0072] It should be noted that modal separation of multimodal feature vectors is performed to eliminate interference between different modal features and retain the unique characteristics of each modality, thereby improving the accuracy and interpretability of subsequent causal analysis.

[0073] Based on text feature subvectors, the intervention effect value of text features on the target variable is calculated, and gradient-weighted class activation mapping is performed on image feature subvectors to extract visual causal significance. At the same time, counterfactual reasoning is performed on numerical feature subvectors to identify the average treatment effect.

[0074] Furthermore, based on text feature sub-vectors, the expected differences in the target variable when specific text attributes change are identified, and the intervention effect value of text features on the target variable is obtained. Then, the intervention effect value is used as feature weights to perform gradient-weighted class activation mapping on the image feature sub-vectors. By identifying the gradient flow of the convolutional layer, key image regions (such as the brand logo area or user attention hotspot area in the main product image) are located, thereby extracting visual causal significance. Combining visual causal significance as a constraint (i.e., limiting the perturbation range of image features in counterfactual reasoning to only act on image regions with low visual causal significance), counterfactual reasoning is performed on the numerical feature sub-vectors to construct counterfactual samples and compare them with actual observation results, finally identifying the average processing effect of numerical features.

[0075] It should be noted that the target variable is a vector of quantitative results generated from key business metrics (such as user conversion rate and average order value) extracted from the e-commerce platform's business database after data cleaning and standardization. The e-commerce platform's business database is a structured dataset stored after the raw business logs collected by the platform's real-time transaction platform have been cleaned and transformed through the ETL pipeline, including user behavior data, transaction data, and attribute data.

[0076] By integrating intervention effect values, visual causal significance, and average treatment effects, a multimodal causal effect value set is generated.

[0077] Furthermore, the intervention effect value generated from text features, the visual causal significance of image feature extraction, and the average processing effect of numerical feature recognition are standardized to generate intermediate indicators with unified dimensions. The intermediate indicators are then aligned and fused in a multimodal manner through a cross-modal attention mechanism to construct a joint representation vector that includes the causal contributions of text, images, and numerical values. Based on the joint representation vector, a weighted integration calculation is performed to generate an initial multimodal causal effect value set. The causal consistency between the initial multimodal causal effect value set and the target variable is verified to generate the final multimodal causal effect value set.

[0078] Based on the multimodal causal effect value set, weight coefficients are assigned according to the feature modality type, a comprehensive causal importance score is calculated, and a local score vector is generated.

[0079] Furthermore, based on the multimodal causal effect value set, the information entropy of the causal contribution of text, image, and numerical modes is calculated respectively, and weight coefficients are assigned to each mode using the entropy weight method. Then, the weight coefficients of each mode are linearly combined with the causal effect values ​​of the corresponding modes in the multimodal causal effect value set to obtain a comprehensive causal importance score. The comprehensive causal importance score is expanded and mapped according to the text feature dimension, image feature dimension, and numerical feature dimension to generate a local score vector containing the specific causal contribution of each feature.

[0080] The expression for calculating the overall causal importance score is:

[0081]

[0082] Where S is the overall causal importance score; w i These are the weighting coefficients for each mode; e i It represents the total causal effect value for each modality; i is the modality index variable;

[0083] A score-privacy budget mapping rule is adopted to map local score vectors into feature-level privacy budget tables;

[0084] Furthermore, based on the local score vector, the causal importance score of each feature is linearly transformed to a fixed interval using the Min-Max normalization method to generate a normalized score vector. Subsequently, a score-privacy budget mapping rule is used to map each score in the normalized score vector to a privacy budget value, forming a privacy budget value vector. Next, each privacy budget value in the privacy budget value vector is matched with the corresponding feature name in the local score vector using key-value pair matching to assemble a structured feature-level privacy budget table. Finally, the feature-level privacy budget table is validated: it is checked whether all features (i.e., text, image, and numerical features) in the multimodal feature vector are assigned privacy budget values, ensuring that all privacy budget values ​​are non-negative and their sum does not exceed the global privacy budget value, and outputting a feature-level privacy budget table that meets the differential privacy requirements.

[0085] It should be noted that the score-privacy budget mapping rule is a strategy framework for dynamically allocating privacy budgets based on causal importance scores. It is defined by an exponential decay function, a global budget cap, a decay coefficient, and a modal weight lower limit. The global privacy budget value is the upper limit of the total privacy protection strength pre-set by data protection needs, privacy regulations, and application scenario risk assessments, and is directly determined according to the actual privacy protection strength requirements.

[0086] Based on the feature-level privacy budget table, a quantum true random number generator is invoked to generate Laplace noise distribution parameters and generate a privacy noise intensity vector;

[0087] Furthermore, based on the feature-level privacy budget table, the privacy budget values ​​of all features in the multimodal feature vector are extracted and assembled into a privacy budget value vector according to the feature order. Then, a quantum true random number generator (such as a hardware device based on quantum optical effects) is called to generate a high-entropy random seed. Next, the pseudo-random number generator is initialized using the high-entropy random seed. Finally, according to each privacy budget value in the privacy budget value vector (as the scale parameter of the Laplace distribution), the corresponding Laplace noise parameter is generated by the initialized pseudo-random number generator, and all the corresponding Laplace noise parameters are organized into a privacy noise intensity vector according to the feature order.

[0088] It should be noted that a pseudo-random number generator is a computing device that uses deterministic algorithms (such as cryptographic algorithms based on the AES-CTR mode) to expand a short, high-entropy random seed into a long random number sequence. The output sequence has statistical randomness but can be repeatedly generated, and is widely used for noise generation in security-sensitive scenarios. Feature order refers to the original dimensional order of multimodal feature vectors, i.e., text features first, image features second, and numerical features last.

[0089] Element-level noise weighting perturbation is performed on the corresponding elements of the privacy noise intensity vector and the multimodal feature vector to generate an initial noise injection feature vector;

[0090] Furthermore, based on each Laplace noise parameter in the privacy noise intensity vector, random noise samples conforming to the Laplace distribution are generated through an initialized pseudo-random number generator, forming a noise value vector with the same dimension as the multimodal feature vector. The noise value vector is aligned with the multimodal feature vector according to the feature dimension, and element-wise addition is performed on each feature value to achieve privacy budget-weighted noise perturbation, ultimately generating an initial noise injection feature vector that satisfies differential privacy protection.

[0091] Orthogonal random perturbations are added to the initial noise-injected feature vector, and KL divergence detection is performed to generate encrypted feature vectors;

[0092] Furthermore, based on the initial noise injection feature vector, a pseudo-random number generator is used to generate a random perturbation vector orthogonal to the initial noise covariance. Each element of the orthogonal random perturbation vector is then added to the corresponding element of the initial noise injection feature vector to achieve element-level noise superposition, generating an intermediate perturbation feature vector. The intermediate perturbation feature vector and the multimodal feature vector are normalized and converted into discrete probability distributions. Based on the discrete probability distributions of the intermediate perturbation feature vector and the multimodal feature vector, the KL divergence value is calculated using the KL divergence formula to obtain a privacy protection metric. The privacy protection metric is then compared with a preset privacy protection threshold. If the privacy protection metric is within the privacy protection threshold, the intermediate perturbation feature vector is determined to meet the privacy protection requirements (e.g., if the privacy protection metric is 0.12, it is within the privacy protection threshold, and the intermediate perturbation feature vector is determined to meet the privacy protection requirements), and an encrypted feature vector is output.

[0093] The expression for calculating the KL divergence value is:

[0094]

[0095] Where D is the KL divergence value, representing the degree of difference between the discrete probability distribution of the intermediate perturbation feature vector and the discrete probability distribution of the original multimodal feature vector; d is the total number of dimensions of the feature vector (e.g., the total number of text, image, and numerical features); k is the dimension index; P(k) is the normalized probability value of the original multimodal feature vector on the k-th dimension; ln is the natural logarithm operation, used to measure the relative entropy of the probability distribution; Q(k) is the normalized probability value of the intermediate perturbation feature vector on the k-th dimension.

[0096] It should be noted that the privacy protection threshold is the maximum acceptable KL divergence value upper limit set based on the privacy protection strength requirements, data sensitivity and availability needs of the specific application scenario, through domain expert experience or theoretical deduction. An exemplary value range is 0.05 to 0.2.

[0097] S3. Based on the encrypted feature vector, calculate the global ATE value of each feature and compare it with the causal significance threshold to filter the encrypted feature vector and obtain a subset of high-value causal features.

[0098] Based on the encrypted feature vector, the global ATE value of each feature is calculated through a homomorphic encryption protocol to generate a local ATE vector;

[0099] Furthermore, based on the encrypted feature vector, the encryption processing status (e.g., assigning 1 to accept processing and 0 to reject processing) and the encryption result variable (target value, such as user conversion rate) of each feature are first parsed using homomorphic decryption according to the modality type label, generating a set of encrypted data pairs for each feature. A homomorphic encryption accumulation operation is then performed on the encrypted data pair set for each feature: the encryption indicator value of the accept processing status (assigned "1") is homomorphically accumulated to obtain the number of encrypted samples; the encrypted feature values ​​of the accept processing status in the encryption result variable are homomorphically accumulated to obtain the sum of the encryption processing group results; simultaneously, the encryption indicator value of the reject processing status (assigned "0") is homomorphically accumulated to obtain the number of encrypted samples. Homomorphically sum the encrypted feature values ​​of the unprocessed states in the encrypted result variables to obtain the sum of the encrypted processing group results; based on the ratio of the sum of the encrypted processing group results of the processed states to the number of samples in the encrypted processing group, obtain the average result of the processed states; based on the sum of the encrypted processing group results of the unprocessed states to the number of samples in the encrypted processing group, obtain the average result of the unprocessed states; identify the difference between the average result of the processed states and the average result of the unprocessed states to obtain the encrypted ATE value of each feature; finally, decrypt the encrypted ATE value of each feature to obtain the plaintext average processing effect value, and combine the plaintext average processing effect values ​​of all features in feature order to generate a local ATE vector;

[0100] The expression for calculating the global ATE value of each feature is:

[0101]

[0102] Among them, G f It is the global ATE value of feature f; O 1,f It is the sum of the results of the processing states of feature f; O 0,f It is the sum of the results of the unprocessed states of feature f; N 1,f N is the number of samples in the processed state of feature f; 0,f It is the number of samples in the unprocessed state of feature f;

[0103] The local ATE vector is decrypted and weighted averaged to generate a global ATE value vector. This vector is then compared with the causal significance threshold feature by feature to identify high-value features and generate a feature label vector.

[0104] Furthermore, the local ATE vector is decrypted to generate a decrypted ATE vector containing the plaintext ATE value of each feature; a predefined feature weight coefficient vector is called, and each ATE value of the decrypted ATE vector is multiplied element-wise with the corresponding weight of the feature weight coefficient vector to obtain the global ATE value, which is then integrated to generate a global ATE value vector; each element in the global ATE value vector is compared feature-by-feature with a preset causal significance threshold: if the global ATE value is greater than the causal significance threshold, it is identified as a high-value feature; if the global ATE value is less than the causal significance threshold, it is identified as a non-high-value feature; the identified features are assigned binary values ​​(for example, setting a binary value of 1 at the position corresponding to a high-value feature and a binary value of 0 at the position corresponding to a non-high-value feature) to generate a feature label vector;

[0105] It should be noted that the feature weight coefficient vector is a set of weight values ​​pre-set based on feature reliability (such as data quality and measurement accuracy) and domain knowledge (such as business importance), containing the weight scalar of each specific feature in text, image, and numerical modalities; the causal significance threshold is a critical value jointly set based on statistical significance level, effect size standard, and domain prior knowledge, with an exemplary value range of 0.01 to 0.3;

[0106] A filtering operation is performed on the encrypted feature vector based on the feature label vector to generate a high-value causal feature subset;

[0107] Furthermore, based on the feature tag vector, each element is iteratively traversed, checking whether the binary value is 1, and recording the indices of all binary values ​​of 1 in a temporary index array to generate a high-value feature index list. Using the high-value feature index list as the access key, the corresponding encrypted feature values ​​are extracted from the encrypted feature vector by index and stored in a temporary value array to form a high-value feature value set. Finally, according to the index order in the high-value feature index list (maintaining the ascending order of the original feature order), the values ​​in the high-value feature value set are reorganized to generate a high-value causal feature subset.

[0108] S4. Encrypt and aggregate the high-value causal feature subset to generate a global feature statistical vector, and perform cross-platform feature distribution alignment through federated learning to generate a fused feature matrix;

[0109] Based on a subset of high-value causal features, the arithmetic mean and standard deviation of each feature are identified, and a local feature statistical vector is generated.

[0110] Furthermore, based on the high-value causal feature subset, each feature column is first traversed, and all encrypted feature values ​​of each feature column are numerically aggregated to obtain the arithmetic mean. The arithmetic mean is then stored as a feature mean vector in feature order. Based on the feature mean vector and the high-value causal feature subset, the squared deviation between the encrypted feature value and the corresponding feature mean is identified for each feature column. The sum of these deviations is then compared with the sample size, and the square root is taken to obtain the standard deviation. All standard deviations are stored as a feature standard deviation vector in feature order, and the feature mean vector and the feature standard deviation vector are merged in feature index order to generate a local feature statistical vector.

[0111] It should be noted that the feature column refers to the vertical data set divided according to the dimensions of the original multimodal feature vector within the high-value causal feature subset, including text feature columns, image feature columns, and numerical feature columns;

[0112] The local feature statistical vector is subjected to Paillier homomorphic encryption using the coordinating node to generate encrypted statistical ciphertext.

[0113] Furthermore, based on the coordinating node, the local feature statistical vector is first read, and a public key and a private key are generated using the Paillier key generation algorithm (the public key is used for encryption operations, and the private key is reserved for subsequent decryption). Then, the coordinating node uses the public key to perform Paillier encryption on the local feature statistical vector: for each value in the local feature statistical vector, the coordinating node selects a random number, and then uses the public key to obtain the ciphertext, generating an encrypted value vector. Finally, the encrypted value vector and the public key information are packaged and encapsulated to generate encrypted statistical ciphertext.

[0114] It should be noted that the coordinating node is a central server or control center responsible for coordinating the computation process, managing key distribution, and aggregating encryption results.

[0115] The encrypted statistical ciphertext is uploaded to the coordination node, where homomorphic accumulation and aggregation are performed to generate the encrypted global statistical ciphertext.

[0116] Furthermore, the local node transmits the encrypted statistical ciphertext to the coordinating node via a secure network channel. The coordinating node then collects all uploaded encrypted statistical ciphertexts and organizes them into an encrypted statistical ciphertext set according to their features. Based on this set, the coordinating node performs a homomorphic multiplication operation on all corresponding encrypted ciphertexts for each feature location, achieving homomorphic accumulation of the ciphertexts and generating a global encrypted statistical value. Finally, the global encrypted statistical values ​​of all features are combined to generate the global encrypted statistical ciphertext.

[0117] The coordinating node uses its private key to decrypt the encrypted global statistical ciphertext, obtains the original global statistical aggregate value, and performs standardization processing on the original global statistical aggregate value to generate a global feature statistical vector;

[0118] Furthermore, the coordinating node first uses its private key to decrypt each encrypted value in the encrypted global statistical ciphertext to obtain the original global statistical aggregate value vector. Based on the original global statistical aggregate value vector, the coordinating node identifies the overall mean and overall standard deviation, and performs Z-score standardization on each plaintext value in the original global statistical aggregate value vector according to the overall mean and overall standard deviation to generate a set of standardized values. Finally, the set of standardized values ​​is combined according to the feature order to generate a global feature statistical vector.

[0119] Based on global feature statistical vectors, high-value causal feature subsets are standardized and aligned to generate aligned local feature sets.

[0120] Furthermore, the global mean and global standard deviation corresponding to each feature in the global feature statistical vector are extracted to generate a standardized parameter set. Based on the standardized parameter set and the high-value causal feature subset, a standardized alignment calculation is performed on each feature value to generate a standardized feature value set. Finally, each feature value set in the standardized feature value set is reorganized according to the original feature index order to generate an aligned local feature set.

[0121] The local feature sets are concatenated into an augmented matrix according to the feature dimensions, and then the dimensionality is reduced to a uniform dimension to generate a platform-level feature projection matrix.

[0122] Furthermore, all feature vectors in the aligned local feature set are concatenated column-wise (i.e., each feature vector is a column of the matrix, and the encrypted feature is a row of the matrix) to form an initial augmented matrix. Based on the initial augmented matrix, the mean of each column is identified, and the mean of the corresponding column is removed from the initial augmented matrix to obtain the centered matrix. Each element in the centered matrix is ​​traversed, and the row and column indices are interchanged to generate a transpose matrix. The transpose matrix is ​​then multiplied by the centered matrix to obtain the covariance matrix. Based on the covariance matrix, the QR algorithm is used for iterative calculation: the covariance matrix is ​​decomposed using QR to obtain an orthogonal matrix and an upper triangular matrix, and the product of the orthogonal matrix and the upper triangular matrix is ​​calculated to generate a new matrix. This decomposition calculation is repeated until the matrix converges to an approximately diagonal form, and the diagonal elements are taken as eigenvalues. For each eigenvalue, the corresponding eigenvector is obtained through inverse iteration, and all eigenvalues ​​and eigenvectors are sorted according to the size of the eigenvalues ​​to generate an eigenvalue vector. Finally, the eigenvalue vector with the larger eigenvalue is selected to form the projection matrix, and the centered matrix is ​​multiplied by the projection matrix to generate the platform-level feature projection matrix.

[0123] It should be noted that the platform-level feature projection matrix can uniformly map high-dimensional features to a low-dimensional space through dimensionality reduction, eliminate differences in feature dimension and scale, retain the main causal information, improve computational efficiency, and enhance the comparability and generalization ability of cross-platform features.

[0124] The coordinating node receives the platform-level feature projection matrix and performs feature space convex combination optimization based on the entropy weight method to generate a fused feature matrix.

[0125] Furthermore, the coordinating node receives the platform-level feature projection matrix and traverses each feature column of the platform-level feature projection matrix. Based on the proportional relationship between each feature value and the sum of feature values ​​in the feature column, it identifies the probability distribution of each feature value. Based on the probability distribution, it uses the Shannon entropy formula to obtain the information entropy of each feature column, generates an entropy value vector, and obtains the weight coefficient of each feature using the entropy weight method based on the entropy value vector, generating a weight vector. Finally, based on the weight vector and the platform-level feature projection matrix, it performs a weighted summation operation on the feature values ​​of each encrypted feature (i.e., each row of the platform-level feature projection matrix) to obtain a fused feature value vector. The fused feature value vector is then reorganized into a single-column matrix according to the encrypted feature index order (where each row corresponds to the fused feature value of an encrypted feature) to generate a fused feature matrix.

[0126] S5. Based on the fusion feature matrix, extract the platform-specific weight vectors, calculate the relative contribution of feature weight strength, generate a contribution ratio table, verify the contribution ratio table with zero-knowledge proof, and generate a data fusion audit report.

[0127] The fusion feature matrix is ​​analyzed, the exclusive weight vectors of each e-commerce platform are extracted, the Frobenius norm is calculated, and a feature weight intensity set is generated.

[0128] Furthermore, based on the fused feature matrix and platform identifier, for each e-commerce platform, all rows belonging to each e-commerce platform are selected from the fused feature matrix to generate a unique feature sub-matrix; then, the arithmetic mean of all encrypted feature values ​​in the same feature column (i.e., the same dimension) in each unique feature sub-matrix is ​​identified to generate a unique weight vector; the Frobenius norm of each unique weight vector is calculated using the Frobenius norm formula, and all Frobenius norms are combined in the order of the platform identifier to generate a feature weight intensity set;

[0129] The expression for calculating the Frobenius norm is:

[0130]

[0131] Among them, V F is the Frobenius norm of the specific weight vector V; n is the length of the specific weight vector, representing the number of weight values ​​contained in the specific weight vector; v p It is the weight of the p-th feature in the dedicated weight vector V; p is the index of the weight value corresponding to the p-th feature in the dedicated weight vector.

[0132] It should be noted that the platform identifier is a unique label (such as platform name or ID) used to distinguish different e-commerce platforms. It is extracted directly from the metadata of the data source or assigned according to the data source.

[0133] Based on the feature weight intensity set, perform relative contribution calculation to generate a platform contribution ratio table;

[0134] Furthermore, the Frobenius norm of all platforms in the feature weight intensity set is traversed, and the Frobenius norms of all platforms are summed to obtain the total intensity value. Based on the feature weight intensity set and the total intensity value, the ratio of the Frobenius norm of each platform to the total intensity value is calculated to generate the relative contribution of each platform. Finally, based on the relative contribution of each platform and the platform identifier, the platform identifier and the corresponding relative contribution are combined into a key-value pair to generate a platform contribution ratio table.

[0135] The expression for calculating relative contribution is:

[0136]

[0137] Among them, E u It is the relative contribution of the u-th platform; F u R is the Frobenius norm of platform u; R is the total strength value, representing the sum of the Frobenius norms of all platforms; u is the platform index, used to identify different e-commerce platforms;

[0138] The zero-knowledge proof certificate is obtained by verifying the contribution ratio table of the e-commerce platform with challenge seeds issued by the regulatory node using zero-knowledge proof.

[0139] Furthermore, the challenge seed issued by the supervisory node is digitally signed using the supervisory node's private key to generate an issued challenge seed; based on the issued challenge seed and the e-commerce platform's contribution ratio table, a proof statement and computational evidence are constructed using a zero-knowledge proof protocol to generate a zero-knowledge proof; finally, the validity of the zero-knowledge proof is checked using the supervisory node's public key and the proof statement, and a zero-knowledge proof certificate is generated after successful verification.

[0140] It should be noted that the challenge seed is a random value generated by the supervisory node using a cryptographically secure random number generator. After being digitally signed, it is used as random input in zero-knowledge proofs to ensure the uniqueness and tamper-proof nature of the proofs, thereby enhancing the security and credibility of the verification process.

[0141] The contribution ratio table of e-commerce platforms and zero-knowledge proof certificates are merged according to timestamps to generate a data fusion audit report;

[0142] Furthermore, based on the e-commerce platform contribution ratio table, all platform identifiers and their corresponding relative contributions are extracted. Simultaneously, proof validity and credential metadata are extracted from the zero-knowledge proof credential. All extracted data are combined into an initial fusion data block. The current timestamp is obtained by calling a high-precision clock source (such as an atomic clock or GPS synchronization clock), and this timestamp is used as an additional field to associate and encapsulate with the initial fusion data block, generating a timestamp fusion data block. Finally, based on the timestamp fusion data block, a report header, data body, and verification information are added according to the standard audit format to generate a data fusion audit report.

[0143] This embodiment also provides a multi-source data security fusion system for e-commerce platforms based on federated learning, including: a data acquisition module, a feature encryption module, a feature filtering module, a cross-platform fusion module, and a joint audit module;

[0144] The data acquisition module is used to collect user behavior data, transaction data, and user attribute data for preprocessing and generating multimodal feature vectors.

[0145] The feature encryption module is used to calculate the causal importance score of each multimodal feature vector based on the multimodal feature vector, generate a local score vector, and dynamically adjust the privacy protection strength according to the local score vector to generate encrypted feature vectors.

[0146] The feature filtering module is used to calculate the global ATE value of each feature based on the encrypted feature vector, compare it with the causal significance threshold, filter the encrypted feature vector, and obtain a subset of high-value causal features.

[0147] The cross-platform fusion module is used to perform encrypted aggregation of high-value causal feature subsets, generate global feature statistical vectors, and perform cross-platform feature distribution alignment through federated learning to generate a fused feature matrix.

[0148] The joint audit module is used to extract the platform-specific weight vectors based on the fusion feature matrix, calculate the relative contribution of feature weight strength, generate a contribution ratio table, verify the contribution ratio table with zero-knowledge proof, and generate a data fusion audit report.

[0149] This embodiment also provides a computer device applicable to the multi-source data security fusion method for e-commerce platforms based on federated learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the multi-source data security fusion method for e-commerce platforms based on federated learning proposed in the above embodiment.

[0150] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0151] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for secure fusion of multi-source data in an e-commerce platform based on federated learning, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0152] In summary, this invention achieves fine-grained dynamic control of privacy protection strength by dynamically adjusting the privacy protection strength based on the local score vector to generate encrypted feature vectors. This allows high-value features to retain higher data utility while enhancing privacy protection for low-value features, thus achieving an optimal balance between privacy protection and data accuracy. Furthermore, by calculating the global ATE value based on the encrypted feature vectors to filter high-value causal feature subsets, this invention enables the optimal selection of causal-driven features within the encrypted domain, improving the business interpretability and fusion efficiency of high-value causal feature subsets.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A secure data fusion method for multi-source e-commerce platforms based on federated learning, characterized in that: include, Collect user behavior data, transaction data, and user attribute data, preprocess them, and generate multimodal feature vectors; Based on multimodal feature vectors, the causal importance score of each multimodal feature vector is calculated to generate a local score vector. The privacy protection strength is then dynamically adjusted based on the local score vector to generate an encrypted feature vector. Based on the encrypted feature vector, calculate the global ATE value of each feature and compare it with the causal significance threshold to filter the encrypted feature vector and obtain a subset of high-value causal features. A subset of high-value causal features is encrypted and aggregated to generate a global feature statistical vector, and cross-platform feature distribution alignment is performed through federated learning to generate a fused feature matrix. Based on the fusion feature matrix, extract the platform-specific weight vectors, calculate the relative contribution of feature weight strength, generate a contribution ratio table, verify the contribution ratio table with zero-knowledge proof, and generate a data fusion audit report.

2. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for generating multimodal feature vectors are as follows: The preprocessing includes data cleaning and standardization. The preprocessed user behavior data, transaction data, and user attribute data are fused to generate multimodal feature vectors.

3. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for calculating the causal importance score of each multimodal feature vector and generating a local score vector are as follows: Modality separation is performed on the multimodal feature vectors to generate text feature sub-vectors, image feature sub-vectors, and numerical feature sub-vectors; Based on text feature subvectors, the intervention effect value of text features on the target variable is calculated, and gradient-weighted class activation mapping is performed on image feature subvectors to extract visual causal significance. At the same time, counterfactual reasoning is performed on numerical feature subvectors to identify the average treatment effect. By integrating intervention effect values, visual causal significance, and average treatment effects, a multimodal causal effect value set is generated. Based on the multimodal causal effect value set, weight coefficients are assigned according to the feature modality type, the comprehensive causal importance score is calculated, and a local score vector is generated.

4. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for dynamically adjusting the privacy protection strength based on the local score vector and generating the encrypted feature vector are as follows: A score-privacy budget mapping rule is adopted to map local score vectors into feature-level privacy budget tables; Based on the feature-level privacy budget table, a quantum true random number generator is invoked to generate Laplace noise distribution parameters and generate a privacy noise intensity vector; Element-level noise weighting perturbation is performed on the corresponding elements of the privacy noise intensity vector and the multimodal feature vector to generate an initial noise injection feature vector; Orthogonal random perturbations are added to the initial noise-injected feature vector, and KL divergence detection is performed to generate encrypted feature vectors.

5. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps are as follows: Calculate the global ATE value of each feature based on the encrypted feature vector, compare it with the causal significance threshold, filter the encrypted feature vector, and obtain a subset of high-value causal features. Based on the encrypted feature vector, the global ATE value of each feature is calculated through a homomorphic encryption protocol to generate a local ATE vector; The local ATE vector is decrypted and weighted averaged to generate a global ATE value vector. This vector is then compared with the causal significance threshold feature by feature to identify high-value features and generate a feature label vector. The encrypted feature vector is filtered based on the feature label vector to generate a high-value causal feature subset.

6. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for encrypting and aggregating high-value causal feature subsets to generate global feature statistical vectors are as follows. Based on a subset of high-value causal features, the arithmetic mean and standard deviation of each feature are identified, and a local feature statistical vector is generated. The local feature statistical vector is subjected to Paillier homomorphic encryption using the coordinating node to generate encrypted statistical ciphertext. The encrypted statistical ciphertext is uploaded to the coordination node, where homomorphic accumulation and aggregation are performed to generate the encrypted global statistical ciphertext. The coordinating node uses its private key to decrypt the encrypted global statistical ciphertext, obtains the original global statistical aggregate value, and performs standardization processing on the original global statistical aggregate value to generate a global feature statistical vector.

7. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for performing cross-platform feature distribution alignment and generating a fused feature matrix through federated learning are as follows. Based on global feature statistical vectors, high-value causal feature subsets are standardized and aligned to generate aligned local feature sets. The local feature sets are concatenated into an augmented matrix according to the feature dimensions, and then the dimensionality is reduced to a uniform dimension to generate a platform-level feature projection matrix. The coordinating node receives the platform-level feature projection matrix and performs feature space convex combinatorial optimization based on the entropy weight method to generate a fused feature matrix.

8. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for extracting platform-specific weight vectors based on the fused feature matrix, calculating the relative contribution of feature weight strengths, and generating a contribution ratio table are as follows. The fusion feature matrix is ​​analyzed, the exclusive weight vectors of each e-commerce platform are extracted, the Frobenius norm is calculated, and a feature weight intensity set is generated. Based on the feature weight intensity set, relative contribution is calculated to generate a platform contribution ratio table.

9. The method for secure fusion of multi-source data in an e-commerce platform based on federated learning as described in claim 1, characterized in that: The steps for verifying the contribution ratio table using zero-knowledge proof and generating a data fusion audit report are as follows. The zero-knowledge proof certificate is obtained by verifying the contribution ratio table of the e-commerce platform with challenge seeds issued by the regulatory node using zero-knowledge proof. The contribution ratio table of e-commerce platforms and zero-knowledge proof certificates are merged according to timestamps to generate a data fusion audit report.

10. A multi-source data security fusion system for e-commerce platforms based on federated learning, based on the multi-source data security fusion method for e-commerce platforms based on federated learning as described in any one of claims 1 to 9, characterized in that: This includes a data acquisition module, a feature encryption module, a feature filtering module, a cross-platform integration module, and a joint audit module; The data acquisition module is used to collect user behavior data, transaction data, and user attribute data for preprocessing and generating multimodal feature vectors. The feature encryption module is used to calculate the causal importance score of each multimodal feature vector based on the multimodal feature vector, generate a local score vector, and dynamically adjust the privacy protection strength according to the local score vector to generate encrypted feature vectors. The feature filtering module is used to calculate the global ATE value of each feature based on the encrypted feature vector, compare it with the causal significance threshold, filter the encrypted feature vector, and obtain a subset of high-value causal features. The cross-platform fusion module is used to perform encrypted aggregation of high-value causal feature subsets, generate global feature statistical vectors, and perform cross-platform feature distribution alignment through federated learning to generate a fused feature matrix. The joint audit module is used to extract the platform-specific weight vectors based on the fusion feature matrix, calculate the relative contribution of feature weight strength, generate a contribution ratio table, verify the contribution ratio table with zero-knowledge proof, and generate a data fusion audit report.

Images