Privacy protection crowdsourcing federal learning method supporting node offline
Through the combination of blockchain and smart contracts, decentralized privacy protection crowdsourcing federated learning is realized, and the model aggregation of nodes is supported, which solves the problems of central server untrustworthiness and network instability, and reduces computing and communication overhead. Ensure data security and model accuracy.
Patent Information
- Application Number
- CN202411140902.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-22
AI Technical Summary
The existing privacy protection crowdsourcing federated learning scheme relies on fully trusted central servers, and in unstable network environments, the model aggregation failure is prone to node disconnection, resulting in high computing and communication costs.
Adopt a decentralized architecture based on blockchain, using identity broadcast encryption schemes and smart contracts for model aggregation, supporting aggregation algorithm with disconnection tolerance, and realizing decentralized security through blockchain storage and processing of local model ciphertexts. polymerization.
Ensure privacy and data security during model training, reduce the computing and communication overhead when nodes are disconnected, provide traceability of malicious nodes and correct model aggregation, and solves the untrustworthy central server and disconnection of nodes. The problem.
Smart Images

Figure CN119011264A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of crowdsourcing federated learning security and provides a privacy-preserving crowdsourcing federated learning method that supports node disconnection. Background Art
[0002] The rapid development of smart mobile devices, mobile Internet, 4G / 5G technology, and the Internet of Things (IoT) has not only promoted the research of mobile crowdsourcing technology (MCS), but also made it the forefront of new data management. MCS motivates community members through the Internet platform, effectively utilizes a large amount of underutilized computing and human resources, and provides cost-effective computing power for individuals and enterprises. With the continuous advancement of machine learning (ML) technology, the application of MCS has expanded from data analysis and text processing to model training tasks, forming ML-based model training crowdsourcing technology (ML-MCS). However, ML-MCS faces the challenge of privacy leakage in practice, because all data needs to be uploaded to the central server, which conflicts with the public's growing awareness of privacy protection, resulting in the "data island" phenomenon and hindering the effective use of data. To solve this problem, researchers proposed a new distributed machine learning paradigm, federated learning (FL), and combined it with MCS to form FLMCS. FLMCS effectively protects data privacy by performing data processing and model training locally, and then uploading model parameters rather than raw data to the central server for aggregation, while promoting data sharing and utilization. FLMCS has demonstrated broad application potential in many fields such as smart healthcare, smart home, and financial services, providing new solutions for data privacy protection and effective utilization.
[0003] FL-MCS solves the "data island" problem in ML-MCS by allowing task publishers to outsource model training tasks to workers with the required data and computing power, and aggregates the trained models through the FL platform to form a shared global model. However, FL-MCS still faces challenges in protecting privacy. Malicious attackers may obtain private information of workers' local data by analyzing plaintext gradients for gradient reconstruction attacks or gradient inference attacks. To address this risk, researchers have proposed privacy-preserving FL-MCS (PPFL-MCS) schemes based on differential privacy (DP), homomorphic encryption (HE) and secure aggregation (SA). However, these schemes have their own limitations: PPFL-MCS based on DP provides privacy protection by adding noise, but may reduce model accuracy; PPFL-MCS based on HE can resist external attacks, but cannot prevent internal attackers from obtaining private information through shared private key decryption; and PPFL-MCS based on SA can protect internal privacy, but external attackers can still obtain global model information by intercepting communications.
[0004] Therefore, the prior art has the following technical problems:
[0005] 1. Most existing PPFL-MCS schemes require a fully trusted central server to aggregate local gradients uploaded by workers. However, in real MCS scenarios, it is difficult to find a fully trusted server due to financial conflicts of interest and potential malicious attackers. At the same time, existing PPFL-MCS schemes are highly dependent on central servers to complete the aggregation operation, which is vulnerable to single point of failure attacks (SPOF);
[0006] 2. The FL-MCS system also often faces unstable communication network problems, especially in harsh industrial IoT environments, where workers may be disconnected due to connection problems, affecting the accuracy of the global model;
[0007] 3. Existing methods are used to enhance the robustness of models in unstable networks, but these methods have high computational and communication costs. Summary of the invention
[0008] In response to the above research issues, the purpose of the present invention is to provide a privacy-preserving crowdsourcing federated learning method that supports offline, so as to solve the problem that the central server is unreliable and the model aggregation is easily affected by node offline.
[0009] In order to achieve the above object, the present invention adopts the following technical solution:
[0010] A privacy-preserving crowdsourcing federated learning method that supports offline communication includes the following steps:
[0011] Step S1. The task issuer publishes its model training task to the blockchain and waits for the participation of the training nodes;
[0012] Step S2. Based on the given model training task, companies and organizations with data (data required for deep learning tasks, which may be image data, text data, audio data or time series according to different training task data) and sufficient computing power join the federated learning system as training nodes;
[0013] Step S3. Based on all training nodes, the task issuer completes the settings of deep learning model hyperparameters (learning rate, batch size, number of training rounds, etc.), deep learning methods (stochastic gradient descent, convolutional neural network, etc.), and security parameters λ (used for encryption algorithms), and initializes a global model W based on these settings. 0, where the global model is in plain text, i.e., unencrypted, and is the aggregated result of all client local models. The key generation center generates a master public-private key pair (mpk, msk) based on the security parameter λ, and generates and distributes private keys for the master public-private key pair, training nodes, and task issuers based on their identity IDs;
[0014] Step S4. If it is the first round of training, train node u i (i∈[1,n]) based on the global model W 0 Use local data to train the deep learning model and obtain the local model i represents the i-th training node, and n represents the total number of training nodes. If the training is t rounds, t is an integer greater than 1, and the training node u i (i∈[1, n]) obtains the global model ciphertext [[W t-1 ]], decrypt it using its own private key to obtain the global model plaintext W t-1 , and train based on the global model and local data to obtain a new local model
[0015] Step S5. For any training round number t, based on the local model plaintext obtained after the local model training is completed (i∈[1,n]), each training node u i (i∈[1,n]) uses the broadcast encryption algorithm to encrypt it and obtain the local model ciphertext And transmit it to the consensus node in the blockchain;
[0016] Step S6. For any training round number t, based on the received local model ciphertext The consensus node completes the consensus chain through the DPoS consensus algorithm, that is, uploading all local model ciphertexts to the blockchain;
[0017] Step S7. For any training round number t, based on the local model ciphertext set stored in the blockchain, the smart contract first determines whether there is a training node offline (based on whether the size of the local model ciphertext set is equal to the total number of training nodes). If there is a node offline, the smart contract executes the offline aggregation algorithm. On the contrary, the smart contract executes the non-offline aggregation algorithm to complete the local model ciphertext aggregation and obtain the global model ciphertext [[W t ]] and stored in the blockchain;
[0018] Step S8. For any training round number t, if the current training round number t is less than or equal to the set total number of training rounds, each training node u i (i∈[1,n]) Download the global model ciphertext of the previous round [[W t-1]], and use your own private key to decrypt and obtain the global model plaintext W t-1 ;
[0019] Step S9. For any training round number t, if the current training round number t is greater than the set total number of training rounds, all training nodes terminate the model training task, and the task issuer obtains the previous round of global model ciphertext [[W t-1 ]], and use your own private key to decrypt and obtain the global model plaintext W t-1 , the current model training is completed and the system is suspended until the next training task starts.
[0020] Further, the specific steps of step S3 are:
[0021] S3.1. Based on the given security parameter λ, obtain large prime numbers p and q, and satisfy p′=2p+1 and q′=2q+1, which are also two large prime numbers;
[0022] S3.2. Obtain the modulus N=p′q′ based on the prime number;
[0023] S3.3. Obtaining the order ord=Npq based on prime numbers and modulus
[0024] S3.4. Obtain a modulus N based on the modulus N and the order 2 , a group of order ord
[0025] S3.5. Group-based Get a generator g;
[0026] S3.6. Obtain triples based on generators, where x 1 , x 2 From the module N 2 The multiplication group of Two random numbers randomly selected from ;
[0027]
[0028] S3.7. Based on multiplication groups Get two hash functions
[0029] S3.8. Group-based Get two n-tuples Wherein i and n have the same meanings as in S4;
[0030] S3.9. Obtain a function about x′ required for decryption based on the modulus N
[0031] S3.10. Based on all the above, construct the master public and private key pair of the key generation center and The master public key is made public and the master private key is kept secret;
[0032] S3.11. Based on the master public-private key pair and training node ID i , the key generation center first calculates h i , 1 =H 1 (ID i ), h i , 2 =H 2 (ID i ), where h i , 1 and h i,2 Corresponding to the hash value of the identity, then calculating s i,1 =p -1 h i,1 and i,2 =p - 1 h i,2 , and finally calculate the private key usk corresponding to the training node i,1 =(s i,1 x 1 +s i,2 x 2 ) -1 (mod N q ), usk i,2 =k i , where i∈[1,n], i and n have the same meanings as in S4;
[0033] S3.12. Based on the master public-private key pair and the task issuer ID, the key generation center first calculates h 1 =H 1 (ID), h 2 =H 2 (ID), where h 1 and h 2 Corresponding to the hash value of the identity, then calculating s 1 =p -1 h 1 and 2 =p -1 h 2 , and finally calculate the private key usk corresponding to the training node TP =(s 1 x 1 +s 2 x 2 ) -1 (mod Nq).
[0034] Further, the specific steps of step S4 are:
[0035] S4.1. If it is the first round of training, all training nodes are based on the global model W 0 , local data, machine learning algorithm and hyperparameters to train the deep learning model, i.e., train the local model, and go to step S5, wherein the machine learning algorithm used in this method is stochastic gradient descent;
[0036] S4.2. If the training is for round t, where t is an integer greater than 1, the training node u i (i∈[1, n]) Download the last round of training from the blockchain to get the global model ciphertext [[W t-1 ]], decrypt it using its own private key, and obtain the global model plaintext W t-1 , and train based on the global model plaintext and local data to obtain a new local model Among them, η represents the learning rate, represents partial derivative, b represents bias, and go to step S5;
[0037] Further, the specific steps of step S5 are:
[0038] S5.1. For any number of training rounds t, each training node u i (I∈[1,n]) from the multiplicative group modulo N Uniformly select a random number r from i ;
[0039] S5.2. For the local model plaintext obtained after the local model training is completed Each training node u i (i∈[1,n]) uses the broadcast encryption algorithm to calculate the local model ciphertext in And meet
[0040] S5.3. Each training node u i (i∈[1,n]) the local model ciphertext calculated in S5.2 Transmitted to the consensus node in the blockchain;
[0041] Further, the specific steps of step S7 are:
[0042] S7.1. For any training round number t, the smart contract first determines whether the number of local model ciphertext sets stored in the blockchain is equal to the total number of training nodes. If they are equal, no node is offline, and the process goes to step S7.2 to execute the non-offline aggregation algorithm; if the number of local model ciphertext sets is less than the total number of training nodes, there is a node offline, and the process goes to step S7.3;
[0043] S7.2. Smart contract executes the non-disconnected aggregation algorithm and calculates the global model ciphertext and store it in the blockchain;
[0044] S7.3. The smart contract executes the offline aggregation algorithm. The smart contract first counts the online node set U′ and goes to step S7.4;
[0045] S7.4. The smart contract calculates for all online nodes i∈U′ in the online node set U′ as well as And go to step S7.5;
[0046] S7.5. Smart Contract Computation as well as Get the global model ciphertext [[W t ]]=(C 1 , C 2 , C′ 3 , C 4 ) and store it in the blockchain;
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] First, the present invention adopts a new identity-based broadcast encryption scheme, and the encryption scheme is aggregate decryptable, that is, the ciphertext sent by each node can only be decrypted by itself, while the aggregated result of the ciphertext sent by all nodes in the broadcast group can be decrypted by all receivers in the broadcast group. This enables the scheme of the present invention to ensure the privacy of the local model submitted by each node during the model training process;
[0049] Second, the present invention conducts federated learning based on blockchain and uses smart contracts to calculate the aggregation results of the ciphertexts of each local model, which provides traceability of malicious nodes that upload malicious data and ensures the correctness of the model aggregation process;
[0050] 3. This invention conducts federated learning based on blockchain, realizes decentralized federated learning security aggregation, and solves the problem of untrustworthy central servers in the aggregation process of traditional centralized federated learning models;
[0051] Fourth, the present invention adopts a new type of offline tolerance aggregation algorithm, which can support the node offline problem caused by network problems during the federated learning training process. Compared with other offline tolerance solutions, this solution greatly reduces the computational overhead and communication overhead of the aggregation process when the node is offline. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 This is the architecture diagram of the privacy-preserving crowdsourcing federated learning system based on blockchain that supports node disconnection;
[0053] Figure 2 It is a flowchart of the privacy-preserving crowdsourcing federated learning system based on blockchain that supports node disconnection. DETAILED DESCRIPTION
[0054] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.
[0055] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.
[0056] 1. Task release: The task publisher releases its model training task to the blockchain and recruits clients to participate in this model training task. The task publisher and the recruited clients form a broadcast group S.
[0057] 2. System initialization: Given a security parameter λ, the key generation center (KGC) selects large prime numbers p, q, and p′=2p+1 and q′=2q+1 are also large prime numbers, and calculates N=p′q′ as the modulus. Then a modulus N with an order of N·(p-1)(q-1) / 2 is selected. 2 Cyclic group of KGC uniform random selection And calculate a triple Then, two secure hash functions are selected KGC uniform random selection and a function L(x) = (x-1) / N about x and calculate Finally, KGC sets the master public key to And the master private key is In addition, the task issuer initializes a global model and uploads it to the blockchain.
[0058] 3. Key distribution: All clients and task issuers in the broadcast group S send their own identities ({ID}i, ID TP ) to the key generation center. The key generation center receives the user u i ID i Then, first calculate h i,1 =H 1 (ID i ), h i,2 =H 2 (ID i ). Then calculate s i,1 =p -1 h i,1 ≡p -1 (s i,1 p+t i,1 Nq)(mod Nq) and s i,2 =p -1 h i,2 ≡p -1 (s i,2 p+t i,2 Nq)(mod Nq). Finally, the key generation center calculates the private key usk i,1 =(s i,1 x 1 +s i,1 x 2 ) -1 (mod Nq) and usk i,2 =k i . And the private key usk i =(usk i,1 ,usk i,2 ) is sent to the corresponding u through a secure channel i .
[0059] 4. Local model training: Take the i-th client u in the t-th round i For example, when obtaining the global model C(W t-1 ), the following steps are performed:
[0060] 4.1. Global model decryption: client u i Use your own private key suk i Decryption calculation
[0061] 4.2. Global model update: Client u i Based on the decrypted global model W of round t-1 t-1 , stochastic gradient descent algorithm and local data set training to obtain the local model of round t
[0062] 5. Consensus on the chain: Client u i Based on local model Master public key mpk, private key usk i And the broadcast user group S, calculate the ciphertext corresponding to the local model ciphertext in And meet Client i Cipher its local model Upload to the consensus node. The consensus node verifies all local model ciphertexts received from the client, then completes the consensus process based on DPoS and uploads all ciphertexts to the blockchain.
[0063] 6. Model aggregation: In the tth round, the smart contract deployed in the blockchain executes the following two algorithms to aggregate the local model ciphertexts of all clients according to whether there is any client offline, and stores the aggregated global model ciphertext in the blockchain.
[0064] 6.1. Non-disconnected aggregation: If no client is disconnected during the current training round, the smart contract calculates the local model ciphertext aggregation
[0065] 6.2 Offline aggregation: If a client is offline in the current training round, the smart contract first counts the online node set U′, and for all online nodes i∈U′ in the online node set U′, calculate as well as Final calculation as well as Get the global model ciphertext [[W t ]]=(C 1 , C 2 , C′ 3 , C 4 ).
[0066] The above three steps of local model training, consensus on the chain, and model aggregation will be repeated continuously until the global model converges or the number of training rounds reaches the maximum.
[0067] 7. Model distribution: When the model training task is completed, the task publisher can download the final global model ciphertext from the blockchain. Then he uses his own private key usk TP , decryption calculation to obtain the final global model plaintext
[0068] In real-world scenarios, it is very difficult to find a completely trustworthy central server. The server may modify the aggregation results for economic interests and other purposes, which ultimately causes the trained model to deviate from the original goal. Therefore, it is necessary to implement decentralized model aggregation. In the actual training process of the federated learning model, it is easy to cause information loss or partial training nodes to drop offline due to network problems. Therefore, it is necessary to implement a model aggregation algorithm that supports offline tolerance to avoid model aggregation failure due to partial training nodes dropping offline due to information loss. The present invention implements a privacy-preserving federated learning system that supports offline, which has the characteristics of multi-function, multi-scenario, openness, security, and efficiency.
[0069] The above are only representative embodiments of the present invention in many specific application scopes, and do not constitute any limitation on the protection scope of the present invention. Any technical solutions formed by transformation or equivalent replacement fall within the protection scope of the present invention.
Claims
1. A privacy-preserving crowdsourcing federated learning method that supports node disconnection, characterized in that: The steps include: Step S1. The task issuer publishes its model training task to the blockchain and waits for the participation of the training nodes; Step S2. Based on the given model training task, companies and organizations with data and sufficient computing power join the federated learning system as training nodes; Step S3. Based on all training nodes, the task issuer completes the settings of the deep learning model hyperparameters, deep learning methods, and security parameters λ used for encryption algorithms, and initializes a global model W0 based on these settings. Here, the global model is in plain text, that is, unencrypted, and is the aggregation result of all client local models. The key generation center generates a master public-private key pair (mpk, msk) based on the security parameter λ, and generates and distributes private keys for the master public-private key pair, training nodes, and task issuers based on their identity IDs; Step S4. If it is the first round of training, train node u i (i∈[1,n]) trains the deep learning model based on the global model W0 and local data to obtain the local model i represents the i-th training node, n represents the total number of training nodes, if it is t rounds of training, t is an integer greater than 1, and the training node obtains the global model ciphertext [[W t-1 ]], decrypt it using its own private key to obtain the global model plaintext W t-1 , and train based on the global model and local data to obtain a new local model Step S5. For any training round number t, based on the local model plaintext obtained after the local model training is completed Each training node u i (i∈[1,n]) uses the broadcast encryption algorithm to encrypt it and obtain the local model ciphertext And transmit it to the consensus node in the blockchain; Step S6. For any training round number t, based on the received local model ciphertext The consensus node completes the consensus chain through the DPoS consensus algorithm, that is, uploading all local model ciphertexts to the blockchain; Step S7. For any training round number t, based on the local model ciphertext set stored in the blockchain, the smart contract first determines whether there is a training node offline according to whether the size of the local model ciphertext set is equal to the total number of training nodes. If there is a node offline, the smart contract executes the offline aggregation algorithm. On the contrary, the smart contract executes the non-offline aggregation algorithm to complete the local model ciphertext aggregation and obtain the global model ciphertext [[W t ]] and stored in the blockchain; Step S8. For any training round number t, if the current training round number t is less than or equal to the set total number of training rounds, each training node u i (i∈[1,n]) Download the global model ciphertext of the previous round [[W t-1 ]], and use your own private key to decrypt and obtain the global model plaintext W t-1 ; Step S9. For any training round number t, if the current training round number t is greater than the set total number of training rounds, all training nodes terminate the model training task, and the task issuer obtains the previous round of global model ciphertext [[W t-1 ]], and use your own private key to decrypt and obtain the global model plaintext W t-1 , the current model training is completed and the system is suspended until the next training task starts.
2. A privacy-preserving crowdsourcing federated learning method supporting node disconnection according to claim 1, characterized in that: The specific steps of step S2 are: S2.
1. Potential training nodes evaluate the requirements of model training tasks and their own data and computing resources; S2.
2. Qualified training nodes register and join the federated learning system through smart contracts; S2.
3. The smart contract updates the training node list and records the information of participating nodes.
3. A privacy-preserving crowdsourcing federated learning method supporting node disconnection according to claim 1, characterized in that: The specific steps of step S3 are: S3.
1. Based on the given security parameter λ, obtain large prime numbers p and q, and satisfy p ′ =2p+1 and q ′ =2q+1 are also two large prime numbers; S3.
2. Obtaining the modulus N=p based on prime numbers ′ q ′ ; S3.
3. Obtaining the order ord=Npq based on prime numbers and modulus S3.
4. Obtain a modulus N based on the modulus N and the order 2 , a group of order ord S3.
5. Group-based Get a generator g; S3.
6. Get triples based on generators, where x1, x2 are from modulo N 2 The multiplication group of Two random numbers randomly selected from ; Among them, P pub Represents a triple, g1, g2, g3 are S3.
7. Based on multiplication groups Get two hash functions S3.
8. Group-based Get two n-tuples wherein i and n have the same meanings as in S4; S3.
9. Obtain a function about x required for decryption based on the modulus N S3.
10. Constructing the master public and private key pair of the key generation center and The master public key is made public and the master private key is kept secret; S3.
11. Based on the master public-private key pair and training node ID i , the key generation center first calculates h i,1 =H1(ID i ),h i,2 =H2(ID i ), where h i,1 and h i,2 Corresponding to the hash value of the identity, then calculating s i,1 =p -1 h i,1 and i,2 =p -1 h i,2 , and finally calculate the private key usk corresponding to the training node i,1 =(s i,1 x1+s i,2 x2) -1 (modNq),usk i,2 =k i , where i∈[1,n], i and n have the same meanings as in S4; S3.
12. Based on the master public-private key pair and the task issuer ID, the key generation center first calculates h1=H1(ID), h2=H2(ID), where h1 and h2 correspond to the hash values of the identity respectively, and then calculates s1=p -1 h1 and s2 = p -1 h2, finally calculate the private key usk corresponding to the training node TP =(s1x1+s2x2) -1 (mod Nq).
4. A privacy-preserving crowdsourcing federated learning method supporting node disconnection according to claim 2, characterized in that: The specific steps of step S4 include: S4.
1. If it is the first round of training, all training nodes train the deep learning model based on the global model W0, local data, machine learning algorithm and hyperparameters, that is, train the local model, and go to step S5, otherwise go to step S4.2, where the machine learning algorithm used in this method is stochastic gradient descent; S4.
2. If the training is for round t, where t is an integer greater than 1, the training node u i (i∈[1,n]) Download the last round of training from the blockchain to get the global model ciphertext [[W t-1 ]], decrypt it using its own private key, and obtain the global model plaintext W t-1 , and train based on the global model plaintext and local data to obtain a new local model Among them, η represents the learning rate, represents partial derivative, b represents bias, and go to step S5.
5. The privacy-preserving crowdsourcing federated learning method supporting node disconnection according to claim 3, characterized in that: The specific steps of step S5 include: S5.
1. For any number of training rounds t, each training node u i (i∈[1,n]) modulo N 2 The multiplication group of Uniformly select a random number r from i ; S5.
2. For the local model plaintext obtained after the local model training is completed Each training node u i (i∈[1,n]) uses the broadcast encryption algorithm to calculate the local model ciphertext in And meet S5.
3. Each training node u i (i∈[1,n]) the local model ciphertext calculated in S5.2 Transmitted to the consensus node in the blockchain.
6. A privacy-preserving crowdsourcing federated learning method supporting node disconnection according to claim 4, characterized in that: The specific steps of step S7 include: S7.
1. For any training round number t, the smart contract first determines whether the number of local model ciphertext sets stored in the blockchain is equal to the total number of training nodes. If they are equal, no node is offline, and the smart contract goes to step S7.2 to execute the non-offline aggregation algorithm; if the number of local model ciphertext sets is less than the total number of training nodes, there is a node offline, and the smart contract goes to step S7.3; S7.
2. Smart contract executes the non-disconnected aggregation algorithm and calculates the global model ciphertext and store it in the blockchain; S7.
3. The smart contract executes the offline aggregation algorithm. The smart contract first counts the online node set U ′ , and go to step S7.4; S7.
4. The smart contract is on the online node set U ′ All online nodes i∈U ′ ,calculate as well as And go to step S7.5; S7.
5. Smart Contract Computation as well as Get the global model ciphertext [[W t ]]=(C1,C2,C3′,C4) and store it in the blockchain.
Citation Information
Patent Citations
Verifiable privacy protection federated learning method and system
CN116467736A
Data privacy protection and high model quality distributed federated learning method
CN117251882A
Privacy protection method and system for financial data sharing based on federated learning
US20230328043A1
Mitigating private data leakage in a federated learning system
US20240169094A1
