A digital product and virtual equity user privacy protection recommendation method based on federated learning

Abstract

This invention discloses a privacy-preserving recommendation method for digital products and virtual rights based on federated learning. Within the framework of federated learning, each data holder locally constructs privacy-preserving computation anchors and trains a model to generate an encrypted joint feature matrix. This avoids the privacy leakage risks associated with direct data sharing and integrates data features extracted from unique user groups across platforms, uncovering correlations and complementarities between data to achieve effective feature fusion. Secondly, by combining latent semantic mining of user behavior trajectories with cross-domain interest graph transfer learning to construct a dynamic weight allocation recommendation model, it can deeply analyze users' potential intentions, discover potential interest associations, and flexibly adjust weights based on real-time behavior and scenarios, improving the dynamic adaptability and accuracy of recommendations. In the recommendation result generation stage, a composite encryption strategy encrypts intermediate parameters, increasing the difficulty of decryption and preventing the reverse engineering of original data, comprehensively protecting the privacy and security of user data and recommendation results, and reducing the risk of privacy leakage.

Description

Technical Field

[0001] This invention relates to the field of privacy protection technology for digital products and virtual rights users, and in particular to a recommendation method for protecting the privacy of digital products and virtual rights users based on federated learning. Background Technology

[0002] With the rapid development of digital technology, digital products (such as software applications and online services) and virtual rights (such as virtual currency and membership privileges) are playing an increasingly important role in people's lives. As a key bridge connecting users with digital products and virtual rights, recommendation systems can provide personalized recommendations based on users' interests and needs, thereby improving user experience and platform revenue.

[0003] However, user privacy protection issues are becoming increasingly prominent in the recommendation process. Existing federated learning-based recommendation methods still have some shortcomings when dealing with recommendations for digital products and virtual rights. For example, how to effectively integrate the data features of various platforms to build a more accurate recommendation model during cross-platform collaborative training; how to improve the dynamic adaptability and accuracy of recommendations; and how to ensure dual privacy protection of data during the generation and transmission of recommendation results are all key issues that need to be addressed. Summary of the Invention

[0004] In view of this, the present invention proposes a privacy-preserving recommendation method for digital products and virtual rights users based on federated learning, which can effectively integrate the data features of various platforms, improve the dynamic adaptability and accuracy of recommendations, and ensure dual privacy protection of data during the generation and transmission of recommendation results.

[0005] The technical solution of this invention is implemented as follows:

[0006] A user privacy protection recommendation method for digital products and virtual rights based on federated learning, specifically including:

[0007] Privacy-preserving computation anchors are built locally at each data holder, and cross-platform collaborative training is performed using a federated learning framework to generate a joint feature matrix in the cryptographic state.

[0008] Based on the joint feature matrix of the encrypted state, and combined with the latent semantic mining of user behavior trajectory and the transfer learning mechanism of cross-domain interest graph, a recommendation model with dynamic weight allocation is constructed.

[0009] Based on the recommendation model, recommendations for digital products and virtual benefits are generated;

[0010] A composite encryption strategy is adopted to provide dual privacy protection for the intermediate parameters during the training process of the recommendation model and the final recommendation results.

[0011] As a further optional embodiment of the federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, the privacy computation anchor points include:

[0012] The data anonymization module is used to anonymize the original local user data.

[0013] The feature vectorization module is used to convert the anonymized data into feature vectors.

[0014] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, the step of utilizing a federated learning framework for cross-platform collaborative training to generate a cryptographic joint feature matrix specifically includes:

[0015] Utilize a federated learning framework for cross-platform collaborative training to generate global model parameters;

[0016] Each data holder node performs a linear combination operation on its local feature vectors based on the global model parameters to generate a local partial joint feature vector;

[0017] Each data holder node will homomorphically encrypt a portion of its local joint feature vectors and then upload them to the federated learning coordination node.

[0018] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted joint feature vectors uploaded by each data holder node, generating an encrypted joint feature matrix.

[0019] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting user privacy in digital products and virtual rights, the step of utilizing a federated learning framework for cross-platform collaborative training to generate global model parameters specifically includes:

[0020] The federated learning coordinating node generates initial global model parameters and encrypts them using a homomorphic encryption algorithm. The encrypted initial model parameters are then distributed to each data holder node.

[0021] After receiving the encrypted initial model parameters, each data holder node trains the model locally using the local feature vector. After training, each data holder node performs homomorphic encryption on the updated local model parameters and uploads the encrypted local model parameters to the federated learning coordinating node.

[0022] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted model parameters uploaded by each data holder node. After the aggregation calculation is completed, new global model parameters are generated, and the new global model parameters are homomorphically encrypted before being distributed back to each data holder node.

[0023] Repeat the above steps until the global model converges and meets the preset convergence conditions.

[0024] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, the recommendation model constructed based on the encrypted joint feature matrix, combined with latent semantic mining of user behavior trajectories and transfer learning mechanisms of cross-domain interest graphs, specifically includes:

[0025] Based on the transfer learning mechanism of cross-domain interest graphs, the interest correlations of users across different platforms are identified, and cross-domain interest weight coefficients are generated.

[0026] Based on the latent semantic mining of user behavior trajectories, we extract the characteristics of users' potential needs and generate behavior trajectory weight coefficients.

[0027] A dynamic weight allocation function is designed based on the joint feature matrix of the encrypted state, the cross-domain interest weight coefficient, and the behavioral trajectory weight coefficient.

[0028] Based on the dynamic weight allocation function, combined with real-time user behavior data and the current recommendation scenario, the weight allocation strategy for each feature dimension is learned.

[0029] Based on the learned weight allocation strategy for each feature dimension, a recommendation model with dynamic weight allocation is constructed using a deep learning recommendation model.

[0030] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, a composite encryption strategy is adopted to provide dual privacy protection for the intermediate parameters during the recommendation model training process and the final recommendation results, specifically including:

[0031] During the model training phase, differential privacy technology is used to add controllable noise to the intermediate parameters;

[0032] During the recommendation result generation stage, homomorphic encryption technology is used to encrypt the transmission of the final recommendation list.

[0033] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, the step of adding controllable noise to intermediate parameters during the model training phase using differential privacy technology specifically includes:

[0034] Pre-set the privacy budget value and the intermediate parameters for which noise needs to be added;

[0035] The noise generation mechanism in differential privacy technology is adopted to generate controllable noise that meets the differential privacy conditions according to a preset privacy budget.

[0036] The generated noise is added to the intermediate parameters to obtain the intermediate parameters after adding noise.

[0037] As a further optional solution to the aforementioned federated learning-based recommendation method for protecting the privacy of digital products and virtual rights users, the step of using homomorphic encryption technology to encrypt and transmit the final recommendation list during the recommendation result generation stage specifically includes:

[0038] Based on the homomorphic encryption algorithm, generate public and private key pairs for encryption and decryption;

[0039] Using the generated public key, each recommendation item in the final recommendation list is homomorphically encrypted, transforming the recommendation item into ciphertext, thus obtaining the encrypted recommendation list;

[0040] The encrypted recommendation list is transmitted to the user's device through a secure network channel, and the user's device decrypts the encrypted recommendation list based on the private key.

[0041] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the steps of the above-described federated learning-based digital product and virtual rights user privacy protection recommendation method.

[0042] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described federated learning-based digital product and virtual rights user privacy protection recommendation methods.

[0043] The beneficial effects of this invention are as follows: Under the framework of federated learning, each data holder constructs privacy-preserving computation anchors locally and trains the model to generate an encrypted joint feature matrix. Since federated learning allows platforms to share model parameters and feature information without sharing the original data, the data features extracted by different platforms based on their unique user groups and behavioral patterns can be integrated in an encrypted state. This avoids the privacy leakage risk caused by direct data sharing. Simultaneously, by utilizing the encrypted joint feature matrix, the advantages of each platform's data can be integrated, fully exploring the correlation and complementarity between different platform data, thereby achieving effective fusion of data features from various platforms. Secondly, combining latent semantic mining of user behavior trajectories with the transfer learning mechanism of cross-domain interest graphs to construct a recommendation model with dynamic weight allocation is key to improving the dynamic adaptability and accuracy of recommendations. Latent semantic mining of user behavior trajectories can deeply analyze the underlying potential of user behavior. In terms of intent and interest patterns, subtle changes in user interests are captured. The transfer learning mechanism of cross-domain interest graphs can transfer users' interest knowledge in other fields to the current recommendation scenario, broadening the recommendation perspective and discovering potential interest associations. Through dynamic weight allocation, the weights of each feature dimension are flexibly adjusted based on users' real-time behavior data and the current recommendation scenario, enabling the recommendation model to respond promptly to dynamic changes in user interests and more accurately match user needs, thereby improving the dynamic adaptability and accuracy of recommendations. In addition, during the recommendation result generation stage, a composite encryption strategy is used to encrypt and protect the intermediate parameters during the recommendation model training process. The composite encryption strategy combines the advantages of multiple encryption technologies, increasing the difficulty of data cracking, ensuring the security of intermediate parameters during training, preventing the inference of user's original data through intermediate parameters, comprehensively protecting the privacy and security of user data and recommendation results, and effectively reducing the risk of privacy leakage. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0045] Figure 1 This is a flowchart illustrating a method for detecting the quality of Gastrodia elata seeds based on hyperspectral microscopic imaging technology according to the present invention.

[0046] Figure 2 This is a schematic diagram of the composition of a computing device according to the present invention. Detailed Implementation

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] refer to Figures 1 to 2 A recommendation method for protecting user privacy in digital products and virtual rights based on federated learning, specifically including:

[0049] Privacy-preserving computation anchors are constructed locally at each data holder, and cross-platform collaborative training is performed using a federated learning framework to generate a cryptographic joint feature matrix; in some embodiments, the privacy-preserving computation anchors include:

[0050] The data anonymization module is used to anonymize the original local user data.

[0051] The feature vectorization module is used to convert anonymized data into feature vectors, providing a basic data format for cross-platform collaborative training under the federated learning framework.

[0052] Thus, by de-identifying the original local user data—for example, encrypting, obfuscating, or replacing sensitive content such as user identity information and contact details—the data loses its direct sensitive relevance locally. Even if the data is accessed during subsequent federated learning, attackers cannot directly obtain the user's true private information, greatly reducing the risk of user privacy leakage and meeting users' stringent requirements for data privacy protection. Secondly, the feature vectorization module further abstracts and protects the data during the process of converting the de-identified data into feature vectors. The feature vectors represent the data in a simpler and more secure form, reducing the possibility of exposing sensitive details and providing a secure data foundation for cross-platform collaborative training under the federated learning framework.

[0053] In some embodiments, the generation of a joint feature matrix of the encrypted state through cross-platform collaborative training using a federated learning framework specifically includes:

[0054] Utilize a federated learning framework for cross-platform collaborative training to generate global model parameters;

[0055] Each data holder node performs a linear combination operation on its local feature vectors based on the global model parameters to generate a local partial joint feature vector;

[0056] Each data holder node will homomorphically encrypt a portion of its local joint feature vectors and then upload them to the federated learning coordination node.

[0057] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted joint feature vectors uploaded by each data holder node, generating an encrypted joint feature matrix.

[0058] Specifically, after the federated learning framework starts, it first initializes a global model, whose parameters are set to... In each round of training, each node uses local data to train the currently received global model parameters, with the node... For example, it is based on local data Calculate the gradient of the model parameters The calculation formula is:

[0059] ;

[0060] in, It is a loss function. These are the currently received global model parameters; similarly, the nodes... and nodes Calculate the gradient separately and Then, each node sends its gradient information to the federated learning coordinating node, which updates the global model parameters through average aggregation.

[0061] ;

[0062] in, Indicates the number of training rounds. It is the learning rate;

[0063] Each data holder node, based on the updated global model parameters Perform a linear combination operation on the local feature vectors, and set the node... The set of local feature vectors is Then the local partial joint feature vector The calculation formula is:

[0064] ;

[0065] in, These are linear combination coefficients that can be dynamically adjusted based on model parameters and the importance of features. (Nodes) and nodes Local partial joint feature vectors are also generated in a similar manner. and ;

[0066] Each node performs homomorphic encryption on a portion of its local joint feature vector, using the node... For example, using the Paillier homomorphic encryption algorithm, let the plaintext be... The public key is The encrypted ciphertext The calculation formula is:

[0067] ;

[0068] in, It is a homomorphic encryption function, node Encrypted Uploaded to the Federated Learning Coordination Node, Node and nodes They also uploaded the encrypted versions separately. and ;

[0069] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted joint feature vectors uploaded by each node. Let the aggregated encrypted joint feature vector be... The calculation formula is:

[0070] ;

[0071] in, This represents the aggregation operation under homomorphic encryption, which combines multiple joint eigenvectors of the encryption to generate the joint feature matrix of the encrypted state. .

[0072] Thus, throughout the process, each data holder node processes data locally, uploading only the encrypted portion of the joint feature vector. Homomorphic encryption ensures computation is performed in an encrypted state. Even if the federated learning coordinating node receives ciphertext, it cannot obtain the original local feature vector information. The secure aggregation protocol further ensures the privacy of each node's data during the aggregation process, avoiding the risk of data leakage during transmission and computation, and meeting the stringent data privacy protection requirements of each data holder. Secondly, through the federated learning framework, each data holder node contributes information from its local data in the form of encrypted partial joint feature vectors while protecting privacy. The federated learning coordinating node aggregates this encrypted information through the secure aggregation protocol to generate an encrypted joint feature matrix. This approach breaks down data silos, achieves cross-platform data fusion, and fully utilizes the data resources of each platform.

[0073] In some embodiments, the step of generating global model parameters through cross-platform collaborative training using a federated learning framework specifically includes:

[0074] The federated learning coordinating node generates initial global model parameters and encrypts them using a homomorphic encryption algorithm. The encrypted initial model parameters are then distributed to each data holder node.

[0075] After receiving the encrypted initial model parameters, each data holder node trains the model locally using the local feature vector. After training, each data holder node performs homomorphic encryption on the updated local model parameters and uploads the encrypted local model parameters to the federated learning coordinating node.

[0076] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted model parameters uploaded by each data holder node. After the aggregation calculation is completed, new global model parameters are generated, and the new global model parameters are homomorphically encrypted before being distributed back to each data holder node.

[0077] Repeat the above steps until the global model converges and meets the preset convergence conditions.

[0078] Specifically, the federated learning coordinating node first generates the initial global model parameters. The Paillier homomorphic encryption algorithm is used to encrypt it, and the public key is set to . The private key is Encrypted initial model parameters The calculation formula is:

[0079] ;

[0080] in, It is a generator. , and They are two large prime numbers. It is a random number; the coordinating node will use the encrypted initial model parameters. Distribute to each data holder node;

[0081] After receiving the encrypted initial model parameters, each data holder node uses its local data. , , Model training is performed separately, with nodes For example, it calculates the gradient of model parameters based on local data. The received encrypted model parameters are used to perform certain calculations (since it is homomorphic encryption, some calculations can be performed on the ciphertext), and then the local model parameters are updated (here, the model update process is simplified, assuming a simple update method). Let the local learning rate be... Updated local model parameters The relevant computational logic is embodied in the ciphertext. After training is complete, the node... Homomorphic encryption is performed on the updated local model parameters to obtain And upload it to the Federated Learning Coordination Node. and nodes The encrypted local model parameters were also processed and uploaded in the same way. and ;

[0082] The federated learning coordination node uses a secure aggregation protocol to aggregate and calculate the encrypted model parameters uploaded by each node. Let the aggregated encrypted model parameters be... The calculation formula is:

[0083] ;

[0084] in, This represents an addition operation under homomorphic encryption. After the aggregation calculation is completed, the coordinating node performs homomorphic encryption on the new global model parameters (the aggregated ciphertext corresponds to the information related to the new global model parameters). (In fact, it is already in an encrypted state; this step can be understood as further ensuring security). Then, the private key is used. Perform partial decryption and verification operations (to ensure the correctness of the aggregation), then re-encrypt and distribute the data back to each data holder node;

[0085] Repeating the above steps, in each round, each node performs local training and encrypted upload based on the received encrypted global model parameters, and the coordinating node performs secure aggregation and encrypted processing for distribution. Suppose that after T rounds of training, the global model converges and meets the preset convergence condition, such as the change in model parameters being less than a threshold. ,Right now .

[0086] Thus, throughout the entire federated learning process, data-holding nodes always perform data processing and model training locally, uploading only encrypted model parameters. Homomorphic encryption ensures that even if model parameters exist in ciphertext form during transmission and aggregation, effective calculation and updates are still possible. The secure aggregation protocol further ensures the privacy of data at each node during aggregation, preventing the federated learning coordinating node from obtaining the original data and specific model parameter details of each node, greatly reducing the risk of data leakage and meeting the stringent data privacy requirements of all participating parties. Secondly, through the federated learning framework and secure aggregation mechanism, data-holding nodes can jointly participate in model training without sharing the original data. Each node performs local optimization based on encrypted global model parameters and contributes the encrypted local model parameters to the global model update. This approach breaks down data silos, enables cross-platform data collaboration, fully utilizes data resources from various platforms, and improves the model's generalization ability and accuracy. Furthermore, iterative training is employed, with each round updating based on the results of the previous round. Preset convergence conditions are used to determine whether the model has converged. This training method ensures that the model can be gradually optimized to adapt to the data characteristics of different platforms. At the same time, homomorphic encryption and secure aggregation technologies do not affect the normal training of the model while ensuring privacy, enabling the model to converge stably in a privacy-protected environment.

[0087] Based on the joint feature matrix of the encrypted state, and combined with the latent semantic mining of user behavior trajectories and the transfer learning mechanism of cross-domain interest graphs, a recommendation model with dynamic weight allocation is constructed, specifically including:

[0088] Based on the transfer learning mechanism of cross-domain interest graphs, the interest correlations of users across different platforms are identified, and cross-domain interest weight coefficients are generated.

[0089] Based on the latent semantic mining of user behavior trajectories, we extract the characteristics of users' potential needs and generate behavior trajectory weight coefficients.

[0090] A dynamic weight allocation function is designed based on the joint feature matrix of the encrypted state, the cross-domain interest weight coefficient, and the behavioral trajectory weight coefficient.

[0091] Based on the dynamic weight allocation function, combined with real-time user behavior data and the current recommendation scenario, the weight allocation strategy for each feature dimension is learned.

[0092] Based on the learned weight allocation strategy for each feature dimension, a recommendation model with dynamic weight allocation is constructed using a deep learning recommendation model.

[0093] Specifically, based on the transfer learning mechanism of cross-domain interest graphs, assuming there are three different platforms (corresponding to three nodes), a cross-domain interest graph is constructed by analyzing user behavior data on different platforms. Let's say the user is on the... The interest vectors on each platform are represented as follows: Interest correlations are identified by calculating the similarity between interest vectors from different platforms. Cosine similarity is used for node... and nodes User interest vector and Their similarity The calculation formula is:

[0094] ;

[0095] Based on the similarity calculation results, cross-domain interest weight coefficients are generated. Let the cross-domain interest weight coefficient matrix be... For nodes To the node Interest transfer, weighting coefficient The calculation formula is:

[0096] ;

[0097] in, Traversal Other nodes besides;

[0098] Latent semantic mining based on user behavior trajectories assumes that the user's behavior sequence over a period of time is as follows: Latent semantic features are mined using topic modeling (such as LDA). Let the mined latent semantic topics be... For each latent semantic topic Its importance weight in the behavioral sequence The calculation formula is:

[0099] ;

[0100] in, It is an implicit semantic theme The frequency of occurrence in the behavioral sequence is used to combine the weights of each latent semantic topic into a behavioral trajectory weight coefficient vector. ;

[0101] Joint feature matrix based on cryptographic state Cross-domain interest weight coefficient matrix and behavioral trajectory weight coefficient vector Design a dynamic weight allocation function, and set the encrypted joint feature matrix. The dimension is For the first in the matrix Line number Column elements Its dynamic weight The calculation formula is:

[0102] ;

[0103] in, and It is an adjustment coefficient that satisfies... , These are the weights at corresponding positions in the cross-domain interest weight coefficient matrix. It is the first in the behavior trajectory weight coefficient vector One element;

[0104] Based on a dynamic weight allocation function, this approach combines real-time user behavior data with the current recommendation scenario. Let the real-time user behavior data be represented as... The current recommended scene feature vector is The algorithm learns weight allocation strategies for each feature dimension through a deep learning model (such as a neural network). It takes the encrypted joint feature matrix, dynamic weights, real-time behavioral data, and recommendation scenario features as input. After calculation by the deep learning model, it outputs the recommendation result. Taking a simple neural network as an example, let the weight matrix from the input layer to the hidden layer be... The weight matrix from the hidden layer to the output layer is: Input vector ( (If the relevant features are selected from the joint feature matrix), then the hidden layer output... and final recommendation results The calculation formula is:

[0105] ;

[0106] ;

[0107] in, It is an activation function. and It is a bias term.

[0108] Thus, by employing the transfer learning mechanism of cross-domain interest graphs and the cosine similarity calculation method, the interest correlations between users across different platforms can be accurately identified, generating reasonable cross-domain interest weight coefficients. This allows the recommendation system to fully utilize users' interest information across multiple platforms, uncover potential interest migration patterns, and provide users with more comprehensive recommendations that better match their cross-domain interests. Secondly, the latent semantic mining method based on user behavior trajectories can extract potential latent semantic features from user behavior sequences and generate behavior trajectory weight coefficients. This helps the recommendation system to gain a deeper understanding of users' potential needs and interest change trends, going beyond surface-level behavioral data, thereby providing users with more forward-looking and personalized recommendations. Furthermore, the designed dynamic weight allocation function combines cross-domain interest weights, behavior trajectory weights, and encrypted joint feature matrices, enabling dynamic adjustments based on different recommendation scenarios and real-time user behavior data. By learning the weight allocation strategy through a deep learning model, the recommendation model can adapt to changing environments in real time, improving the accuracy and adaptability of recommendations and providing users with more timely and relevant recommendations.

[0109] Recommendations for digital products and virtual rights are generated based on a recommendation model.

[0110] Specifically, based on the recommendation model built through cross-domain interest graph transfer learning, latent semantic mining of user behavior trajectories, and dynamic weight allocation, the generated digital product and virtual benefit recommendations can more accurately match users' actual needs and interests. The cross-domain interest weight coefficient enables recommendations to cross different platforms and integrate users' interest information across multiple domains; the behavior trajectory weight coefficient deeply captures users' potential needs and interest change trends; and the dynamic weight allocation function combines multiple factors to adjust the recommendation strategy in real time. Therefore, the final recommendation results are more in line with users' preferences, effectively improving users' click-through rate, purchase rate, and significantly enhancing user satisfaction.

[0111] Secondly, because the recommendation model fully considers the user's interest characteristics under different platforms and different behavioral trajectories, the generated recommendation results are no longer limited to a single domain or a single type of digital products and virtual rights. Cross-domain interest mining enables recommendations to cover content that users may be interested in on multiple platforms, while latent semantic mining of behavioral trajectories helps to discover some non-mainstream but potential interests of users. Therefore, the recommendation results have richer diversity and can provide users with more novel choices to meet their diverse needs.

[0112] A composite encryption strategy is adopted to provide dual privacy protection for the intermediate parameters during the training process of the recommendation model and the final recommendation results, specifically including:

[0113] During the model training phase, differential privacy technology is used to add controllable noise to the intermediate parameters;

[0114] During the recommendation result generation stage, homomorphic encryption technology is used to encrypt the transmission of the final recommendation list.

[0115] In some embodiments, the addition of controllable noise to intermediate parameters using differential privacy technology during the model training phase specifically includes:

[0116] Pre-set the privacy budget value and the intermediate parameters for which noise needs to be added;

[0117] The noise generation mechanism in differential privacy technology is adopted to generate controllable noise that meets the differential privacy conditions according to a preset privacy budget.

[0118] The generated noise is added to the intermediate parameters to obtain the intermediate parameters after adding noise.

[0119] Specifically, during the model training phase, differential privacy technology adds controllable noise to ensure that even if the intermediate parameters are maliciously obtained, attackers cannot accurately infer the sensitive information of a single user, provided that a certain privacy budget is met. During the recommendation result generation phase, homomorphic encryption technology ensures that the final recommendation list is transmitted in an encrypted state, avoiding information leakage during transmission. This echoes the privacy protection measures previously implemented in data holding, feature processing, and model building, constructing a full-chain privacy protection system from the data source to the recommendation result output, greatly improving the security of user data.

[0120] Secondly, when adding controllable noise using differential privacy technology during the model training phase, a privacy budget value and intermediate parameters for which noise needs to be added are pre-set, and controllable noise that meets the conditions is generated according to the noise generation mechanism. This approach protects privacy while minimizing the negative impact on model training. A reasonably set privacy budget can balance the relationship between privacy protection and model performance, ensuring that the intermediate parameters after adding noise still retain sufficient effective information for model training. This ensures that the recommendation model can converge normally and learn the inherent patterns of user interests and data, maintaining the accuracy and reliability of the recommendation system.

[0121] Furthermore, homomorphic encryption is used to encrypt the final recommendation list during transmission, ensuring the confidentiality of the recommendation results. The recipient cannot obtain the specific content of the recommendation list without decryption. Only authorized users or systems with the corresponding keys can decrypt and use it. This not only prevents the recommendation results from being stolen or tampered with during transmission, but also works closely with other privacy protection aspects of the entire technical solution, enabling users to receive personalized digital product and virtual benefit recommendations in a secure environment, further enhancing users' trust in the recommendation system.

[0122] In some embodiments, the step of using homomorphic encryption technology to encrypt and transmit the final recommendation list during the recommendation result generation stage specifically includes:

[0123] Based on the homomorphic encryption algorithm, generate public and private key pairs for encryption and decryption;

[0124] Using the generated public key, each recommendation item in the final recommendation list is homomorphically encrypted, transforming the recommendation item into ciphertext, thus obtaining the encrypted recommendation list;

[0125] The encrypted recommendation list is transmitted to the user's device through a secure network channel, and the user's device decrypts the encrypted recommendation list based on the private key.

[0126] Specifically, homomorphic encryption technology is used in the recommendation result generation stage. A public key and private key pair are generated based on the homomorphic encryption algorithm. The public key is used to encrypt each recommendation item in the final recommendation list, converting it into ciphertext. This process ensures that the recommendation list exists in ciphertext during transmission. Even if it is intercepted by a malicious third party, the specific content of the recommendation item cannot be obtained due to the lack of the corresponding private key. This, together with the differential privacy and other privacy protection measures used in the model training stage, forms a multi-layered privacy protection system, which greatly enhances the privacy security of user data and recommendation results and effectively prevents the leakage of sensitive information such as user interests and preferences.

[0127] Secondly, the encrypted recommendation list is transmitted to the user terminal through a secure network channel. The homomorphic encryption ensures the stability and integrity of the transmission in ciphertext. During transmission, no additional errors or data corruption will be introduced due to encryption operations, ensuring that the encrypted recommendation list received by the user terminal is consistent with the content sent by the sender. At the same time, the secure network channel further reduces the risk of data being tampered with or interfered with during transmission, providing a reliable guarantee that users can accurately obtain recommendation results and maintaining the stability and reliability of the recommendation system.

[0128] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the steps of the above-described federated learning-based digital product and virtual rights user privacy protection recommendation method.

[0129] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described federated learning-based digital product and virtual rights user privacy protection recommendation methods.

[0130] 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 within the protection scope of the present invention.

Claims

1. A digital product and virtual rights user privacy protection recommendation method based on federated learning, characterized in that, Specifically comprising: A privacy computing anchor point is constructed locally at each data holder, and a federated learning framework is used to collaboratively train across platforms to generate an encrypted joint feature matrix; Based on the encrypted joint feature matrix, and in combination with the latent semantic mining of user behavior trajectories and the transfer learning mechanism of cross-domain interest graph, a recommendation model with dynamic weight distribution is constructed; The recommendation model is used to generate a recommendation result of digital products and virtual rights and interests; A composite encryption strategy is used to provide double privacy protection for intermediate parameters and final recommendation results in the training process of the recommendation model.

2. The method of claim 1, wherein, The privacy computing anchor point comprises: A data desensitization processing module for performing desensitization operations on local user original data; A feature vectorization module for converting desensitized data into feature vectors.

3. The method of claim 1, wherein, The use of a federated learning framework to collaboratively train across platforms to generate an encrypted joint feature matrix specifically comprises: A global model parameter is generated using a federated learning framework to collaboratively train across platforms; Each data holder node performs linear combination operations on its local feature vectors based on the global model parameter to generate a local partial joint feature vector; Each data holder node uploads the local partial joint feature vector after homomorphic encryption processing to a federated learning coordination node; The federated learning coordination node aggregates and calculates the encrypted partial joint feature vectors uploaded by each data holder node using a secure aggregation protocol to generate an encrypted joint feature matrix.

4. The method of claim 3, wherein, The use of a federated learning framework to collaboratively train across platforms to generate a global model parameter specifically comprises: The federated learning coordination node generates an initial global model parameter and encrypts it using a homomorphic encryption algorithm, and distributes the encrypted initial model parameter to each data holder node; After receiving the encrypted initial model parameter, each data holder node performs model training locally using the local feature vector, and after training, each data holder node encrypts the updated local model parameter using a homomorphic encryption algorithm and uploads the encrypted local model parameter to the federated learning coordination node; The federated learning coordination node aggregates and calculates the encrypted model parameters uploaded by each data holder node using a secure aggregation protocol, and after aggregation and calculation, generates a new global model parameter, encrypts it using a homomorphic encryption algorithm, and then distributes it back to each data holder node; Repeat the above steps until the global model converges and meets the preset convergence condition.

5. The federated learning based user privacy protection recommendation method for digital products and virtual rights according to claim 1, characterized in that, The use of a federated learning framework to collaboratively train across platforms to generate an encrypted joint feature matrix specifically comprises: Based on the transfer learning mechanism of cross-domain interest graph, the interest correlation of users between different platforms is identified to generate cross-domain interest weight coefficients; Based on the latent semantic mining of user behavior trajectories, the user's potential demand features are extracted to generate behavior trajectory weight coefficients; Based on the encrypted joint feature matrix, cross-domain interest weight coefficients, and behavior trajectory weight coefficients, a dynamic weight distribution function is designed; Based on the dynamic weight distribution function, the weight distribution strategy of each feature dimension is learned in combination with the real-time behavior data of users and the current recommendation scenario; Based on the learned weight allocation strategy for each feature dimension, a recommendation model with dynamic weight allocation is constructed using a deep learning recommendation model.

6. The federated learning based user privacy protection recommendation method for digital products and virtual rights and interests according to claim 1, characterized in that, The aforementioned composite encryption strategy provides dual privacy protection for both intermediate parameters during the recommendation model training process and the final recommendation result, specifically including: During the model training phase, differential privacy technology is used to add controllable noise to the intermediate parameters; During the recommendation result generation stage, homomorphic encryption technology is used to encrypt the transmission of the final recommendation list.

7. The federated learning-based digital product and virtual entitlement user privacy-protecting recommendation method of claim 6, wherein, During the model training phase, differential privacy technology is used to add controllable noise to intermediate parameters, specifically including: Pre-set the privacy budget value and the intermediate parameters for which noise needs to be added; The noise generation mechanism in differential privacy technology is adopted to generate controllable noise that meets the differential privacy conditions according to a preset privacy budget. The generated noise is added to the intermediate parameters to obtain the intermediate parameters after adding noise.

8. The federated learning-based digital product and virtual entitlement user privacy-preserving recommendation method of claim 6, wherein, In the recommendation result generation stage, homomorphic encryption technology is used to encrypt and transmit the final recommendation list, specifically including: Based on the homomorphic encryption algorithm, generate public and private key pairs for encryption and decryption; Using the generated public key, each recommendation item in the final recommendation list is homomorphically encrypted, transforming the recommendation item into ciphertext, thus obtaining the encrypted recommendation list; The encrypted recommendation list is transmitted to the user's device through a secure network channel, and the user's device decrypts the encrypted recommendation list based on the private key.

9. A computing device, comprising: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the user privacy protection recommendation method for digital products and virtual rights based on any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the user privacy protection recommendation method for digital products and virtual rights based on federated learning as described in any one of claims 1-8.

Images