Federal recommendation privacy protection method and system based on ring signature and selective aggregation

Abstract

The invention discloses a federal recommendation privacy protection method based on ring signature and selective aggregation, which realizes anonymization privacy protection and selective aggregation in a federal learning process through an innovative double-server architecture and a hierarchical security mechanism. The method comprises the following steps: 1) deploying an aggregation server and a security server in a system initialization stage, and completing parameter initialization and key distribution; 2) the client executes local model training, and anonymization privacy protection uploading is realized by adopting a ring signature technology; 3) the security server performs selective aggregation of credibility verification, and completes grouping and parameter fusion based on user similarity; and 4) ensuring the completeness and source credibility of model updating through a hierarchical security parameter distribution mechanism. While the recommendation precision is ensured, the user identity leakage and privacy tracking are effectively prevented, the comprehensive security guarantee is provided for the federal recommendation system, and the method has the advantages of low calculation overhead and high communication efficiency.

Description

Technical Field

[0001] This invention relates to a federated recommendation privacy protection model and device based on ring signatures, belonging to the fields of artificial intelligence and information security technology. Background Technology

[0002] With the widespread adoption of the internet and mobile devices, recommender systems have become a core component of e-commerce, social networks, and content platforms. Traditional centralized recommender systems require the collection of users' personal data and interaction behavior, posing a serious risk of privacy breaches. Federated recommender systems, through a "data-inactive, model-dynamic" approach, allow users to train models locally and only upload model parameters, thus protecting user privacy to a certain extent.

[0003] However, existing federated recommender systems still face many challenges. On the one hand, model parameters uploaded by clients may be reverse-engineered to infer sensitive user information or participant identities, leading to a double leakage of behavioral and identity privacy. On the other hand, traditional federated recommender aggregation strategies ignore the differences in user interests, making it difficult to maintain personalized recommendations while protecting privacy. More importantly, techniques such as differential privacy and homomorphic encryption used to enhance privacy often severely compromise recommendation accuracy or incur extremely high computational costs.

[0004] Ring signature technology provides strong anonymity guarantees for federated recommendation systems, making it impossible for attackers to determine the source of specific model updates. However, directly applying ring signatures to federated recommendation systems faces problems such as high signature verification overhead and inability to support personalized aggregation. To address these issues, this invention proposes a privacy-preserving model and device for federated recommendation based on ring signatures. Through an innovative dual-server architecture and layered security mechanisms, it provides comprehensive privacy protection while ensuring recommendation quality. Summary of the Invention

[0005] Purpose of the Invention: To address the problems of privacy leakage risks, coarse aggregation strategies, and high computational overhead in existing federated recommendation systems, this invention proposes a privacy-preserving method and system for federated recommendations based on ring signatures and selective aggregation. Ring signatures provide strong anonymity, trusted verification ensures system security, and selective aggregation improves recommendation performance, achieving a balance between privacy protection and recommendation performance.

[0006] Technical solution: The present invention provides a federated recommendation privacy protection model and apparatus based on ring signatures, comprising the following steps:

[0007] 1) System initialization phase: Deploy aggregation server and security server, which are responsible for model fusion and security authentication respectively; initialize global model parameters and complete item clustering; generate and distribute RSA key pairs for encrypted communication, ECDSA key pairs for identity authentication, and public key rings required to build ring signatures.

[0008] 2) Anonymization and privacy protection during the upload stage: The client trains a personalized recommendation model based on local user interaction data; the ChaCha20 algorithm and hybrid encryption mechanism are used to encrypt and protect the model parameters; and ring signature technology is used to anonymize the uploaded messages to ensure that the data source is untraceable.

[0009] 3) Selective aggregation stage based on trusted verification: The security server verifies the ring signature uploaded by the client to ensure the integrity and legality of the message; decrypts the model parameters and forwards them to the aggregation server; the aggregation server divides the clients into similar groups and dissimilar groups based on cosine similarity and the elbow rule; and performs secure aggregation of the model parameters within the similar groups.

[0010] 4) Layered security parameter distribution phase: The security server re-encrypts and digitally signs the aggregated model parameters; similar group clients receive and verify the parameter updates and complete the local model update; dissimilar group clients only receive item clustering information for subsequent optimization.

[0011] Compared with the prior art, the advantages of the present invention are as follows:

[0012] (1) Strong anonymity is provided by ring signature technology, combined with multi-layer encryption mechanism, which effectively prevents the leakage of user identity and sensitive information and meets strict privacy protection requirements.

[0013] (2) A selective aggregation strategy based on user similarity is adopted to ensure that model updates are performed among clients with similar feature distributions, which significantly improves the accuracy and personalization of the recommendation system.

[0014] (3) By using a dual-server architecture, the computing tasks are reasonably divided, and by combining a lightweight encryption algorithm and an optimized signature verification process, the computing and communication overhead is kept low while ensuring security. Attached Figure Description

[0015] Figure 1 This is the processing flow of an embodiment of the present invention.

[0016] Figure 2 This is the overall model of an embodiment of the present invention. Detailed Implementation

[0017] To enhance understanding of the present invention, the invention is further explained in detail below with reference to specific embodiments. The processing flow and overall model of the present invention are as follows: Figure 1 and Figure 2 As shown, the detailed implementation process is as follows.

[0018] Step 1: System initialization, specifically including: Sub-step 1-1: Deploying the aggregation server and security server. The aggregation server is responsible for user clustering, similarity evaluation, and model fusion, while the security server is responsible for parameter exchange and authentication. Unlike the traditional single-server federated learning architecture, this solution adopts a dual-server collaborative mechanism, achieving separation of responsibilities: the aggregation server focuses on model performance optimization, while the security server focuses on privacy and security. This design not only avoids the privacy risks caused by single-point control but also provides an architectural foundation for subsequent anonymization authentication and selective aggregation through separated computation; Sub-step 1-2: The aggregation server first initializes the global model parameters. And use K-Means to embed the item vectors Perform clustering to generate cluster label vectors ,in Represents items The cluster number to which it belongs; unlike the traditional federated recommendation method where all clients participate in aggregation indiscriminately, this scheme divides items into semantic categories through pre-clustering, aligning user grouping with interests, thereby improving the effectiveness of aggregation within similar groups and reducing noise from heterogeneous user updates; Sub-steps 1-3: The security server generates and distributes keys for the system, including: the security server for Generate RSA key pairs Generate your own RSA key pair Used for encrypting and decrypting session keys; for each Generate ECDSA signature key pair Generate your own ECDSA signature key pair These key pairs are used for authentication;

[0019] Receiving client Sending ring signature request and its public key Generate a public key ring And send to the client Client Using the public key ring and its private key Generate ring signature structure And calculate the ring signature link tag ,in This is the initial challenge value. It is a vector of random numbers;

[0020] The key system consists of two layers: transmission encryption (RSA) and identity authentication (ECDSA), and dynamically constructs signature groups using a ring signature request mechanism. Compared to static groups or fixed pseudonyms, this dynamic ring construction supports flexible anonymization of clients in each training round, effectively defending against cross-round association attacks.

[0021] Step 2: Anonymous and privacy-preserving upload, specifically including: Sub-step 2-1: The client trains a local model based on local user interaction logs, and minimizes the objective function by optimizing the scoring function and the local item embedding matrix. ,in Calculate the predicted rating for item i for user u; Sub-step 2-2: The client encrypts the local parameter update, specifically by randomly generating a symmetric key. With random numbers The ChaCha20 algorithm is used to generate model update ciphertext. Use the RSA public key of the secure server. The encryption key is obtained by combining the symmetric key with a random number. Construct upload message ;

[0022] A hybrid encryption mode, employing symmetric encryption to protect the data body and asymmetric encryption to protect the key, significantly reduces computational and communication overhead compared to homomorphic encryption while ensuring transmission confidentiality. Compared to differential privacy, this scheme guarantees that model accuracy is unaffected by noise, making it more suitable for high-dimensional, frequent parameter update scenarios in federated recommendations that require high model accuracy. Sub-steps 2-3: The client uses the described ring signature structure. Regarding the message The process involves signing the message and uploading the signed message to a secure server. Anonymity is achieved through ring signatures, where the signer voluntarily selects a set of public keys to form a ring. The verifier can only confirm that the signature originated from that ring, but cannot determine the specific identity.

[0023] Step 3: Selective aggregation based on trusted verification, specifically including: Sub-step 3-1: The security server processes the received messages. The ring signature verification process is as follows: Initialization For each ,calculate and Where G is the base point of the elliptic curve. This represents the public key mapping value of the i-th client in the public key ring;

[0024] calculate And verify If the signature is valid, then the signature is valid. Forward to the aggregation server, otherwise discard the message; Sub-step 3-2: The security server decrypts the model update using its RSA private key. Decryption And then through Restore local parameter updates And send it to the aggregation server;

[0025] Sub-step 3-3: The aggregation server is based on all clients. Users are grouped, and for each core client c and target item category k, the similarity is calculated. Based on cosine similarity ranking and the elbow rule, clients are divided into similar and dissimilar groups.

[0026] Traditional federated aggregation averages updates from all clients, ignoring the heterogeneity of user interests and potentially reducing personalization performance. This solution dynamically divides users into similar groups based on item clustering and user embedding similarity, aggregating only highly relevant updates within each group. This improves the model's relevance to specific interest groups after aggregation, reduces interference from dissimilar clients, and optimizes recommendation quality while protecting privacy.

[0027] Sub-steps 3-4: Aggregate client parameters within similar groups and calculate the average parameter. .

[0028] Step 4: Layered security parameter distribution, specifically including: Sub-step 4-1: The security server encrypts and signs the aggregate parameter V_s. A new session key is generated. Calculate ciphertext The client's RSA public key is used to encrypt the session key and a random number to obtain... Use the ECDSA private key of the security server to sign the ciphertext and generate... Sub-step 4-2: Clients within similar groups receive triples. Then, the signature is verified using the ECDSA public key of the security server. After successful verification, use your own RSA private key. Decryption Thus, the aggregation parameters are restored. Sub-step 4-3: Clients within the similar group set their local parameters. Updated to Clients in dissimilar groups only receive item clustering labels M for comparative reinforcement of subsequent training, without directly updating model parameters. The hierarchical distribution strategy achieves a differentiated balance between privacy and utility. Clients in similar groups obtain the complete aggregated model to quickly improve performance, while clients in dissimilar groups only receive item clustering labels for comparative learning supervision of the local model. This avoids the negative impact of low-quality updates and provides cross-group semantic knowledge through clustering labels.

[0029] This invention constructs a federated recommendation framework that provides privacy protection at both the identity and behavior levels by integrating ring signatures, a dual-server architecture, selective aggregation, and a hierarchical distribution mechanism. Compared with existing solutions that only focus on parameter encryption or noise injection, this solution achieves a dynamic balance between privacy protection strength and recommendation accuracy through collaborative innovation in four levels: system architecture, anonymous authentication, aggregation strategy, and knowledge transfer. It is particularly suitable for cross-platform recommendation scenarios that are sensitive to privacy and have diverse user interests.

[0030] Based on the same inventive concept, the present invention provides a ring signature-based federated recommendation privacy protection model and apparatus, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the aforementioned ring signature-based federated recommendation privacy protection model and apparatus.

[0031] It should be noted that the above embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. A federated recommendation privacy protection method based on ring signatures and selective aggregation, characterized in that, The method includes the following steps: Step 1: During the system initialization phase, deploy the aggregation server and the security server to complete parameter initialization and key distribution; Step 2: The client executes local model training and uses ring signature technology to achieve anonymous and privacy-preserving upload; Step 3: The security server performs selective aggregation for trusted verification and completes grouping and parameter fusion based on user similarity; Step 4: Through a hierarchical secure parameter distribution mechanism, ensure the integrity and source credibility of model updates.

2. The federated recommendation privacy protection method based on ring signatures and selective aggregation according to claim 1, characterized in that, Step 1: System initialization, specifically includes: Sub-step 1-1: Deploying the aggregation server and security server, wherein the aggregation server is responsible for user clustering, similarity assessment, and model fusion, and the security server is responsible for parameter exchange and authentication; Sub-step 1-2: The aggregation server first initializes the global model parameters. And use K-Means to embed the item vectors Perform clustering to generate cluster label vectors ,in Represents items The cluster number it belongs to; Sub-steps 1-3: The security server generates and distributes keys to the system, including: the security server for Generate RSA key pairs Generate your own RSA key pair Used for encrypting and decrypting session keys; for each Generate ECDSA signature key pair Generate your own ECDSA signature key pair These key pairs are used for authentication; Receiving client Sending ring signature request and its public key Generate a public key ring And send to the client Client Using the public key ring and its private key Generate ring signature structure And calculate the ring signature link tag ,in This is the initial challenge value. It is a vector of random numbers.

3. The federated recommendation privacy protection method based on ring signatures and selective aggregation according to claim 1, characterized in that, Step 2: Anonymous and privacy-preserving upload, specifically including: Sub-step 2-1: The client trains a local model based on local user interaction logs, and minimizes the objective function by optimizing the scoring function and the local item embedding matrix. Sub-step 2-2: The client encrypts local parameter updates, specifically by randomly generating a symmetric key. With random numbers The ChaCha20 algorithm is used to generate model update ciphertext. ; Use the RSA public key of the security server The encryption key is obtained by combining the symmetric key with a random number. Construct upload message Sub-steps 2-3: The client uses the ring signature structure. Regarding the message Sign the message and upload the signed message to the secure server.

4. The federated recommendation privacy protection method based on ring signatures and selective aggregation according to claim 1, characterized in that, Step 3: Selective aggregation based on trusted verification specifically includes: Sub-step 3-1: The security server processes the received messages. The ring signature verification process is as follows: Initialization For each ,calculate and ;calculate ;verify If the signature is valid, then the signature is valid. Forward to the aggregation server, otherwise discard the message; Sub-step 3-2: The security server decrypts the model update using its RSA private key. Decryption And then through Restore local parameter updates And send it to the aggregation server; Sub-step 3-3: The aggregation server, based on all clients'... Users are grouped, and for each core client c and target item category k, the similarity is calculated. Based on cosine similarity ranking and the elbow rule, clients are divided into similar and dissimilar groups; sub-steps 3-4: aggregate client parameters within the similar groups and calculate the average parameter. .

5. The federated recommendation privacy protection method based on ring signatures and selective aggregation according to claim 1, characterized in that, Step 4: Layered security parameter distribution, specifically includes: Sub-step 4-1: The security server encrypts and signs the aggregate parameter V_s, specifically by generating a new session key. Calculate ciphertext The client's RSA public key is used to encrypt the session key and a random number to obtain... Use the ECDSA private key of the security server to sign the ciphertext and generate... Sub-step 4-2: Clients within similar groups receive triples. Then, the signature is verified using the ECDSA public key of the security server. After successful verification, use your own RSA private key. Decryption Thus, the aggregation parameters are restored. Sub-step 4-3: Clients within the similar group set their local parameters. Updated to Clients in dissimilar groups only receive item clustering labels M, which are used for supervised contrast reinforcement during subsequent training, rather than directly updating model parameters.

6. A federated recommendation privacy protection system based on ring signatures and selective aggregation, characterized in that, To implement the federated recommendation privacy protection method based on ring signature and selective aggregation as described in any one of claims 1-5, the system comprises: module 1: system initialization module; module 2: anonymized privacy protection upload module; module 3: selective aggregation module based on trusted verification; and module 4: hierarchical security parameter distribution module. Module 1: System Initialization Module, which is used to: deploy the aggregation server and security server and clarify their functional division; initialize global model parameters and complete the clustering of item embedding vectors; generate and distribute various key pairs for encryption authentication, and assist clients in building ring signature capabilities; Module 2: Anonymized and privacy-preserving upload module, used to: guide the client to train the model locally and generate parameter updates; perform multi-layer encryption on the parameter updates to generate secure transmission messages; and use ring signature technology to anonymize and sign the messages before uploading. Module 3, a selective aggregation module based on trusted verification, is used for: performing ring signature verification and decryption on messages uploaded by clients on the secure server side; dynamically grouping clients based on their similarity on the aggregation server side; and securely aggregating client parameter updates within similar groups. Module 4: Layered security parameter distribution module, used to: encrypt and sign the aggregated model parameters and securely distribute them to the target client; guide the client to verify and decrypt the received messages; and execute differentiated parameter update strategies based on the client's group (similar group or dissimilar group).

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a federal recommendation privacy protection method based on ring signature and selective aggregation as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by a processor, the computer instructions implement a federally recommended privacy protection method based on ring signatures and selective aggregation as described in any one of claims 1-5.

Images