Federated learning method of fault-tolerant layered verifiable secure aggregation based on privacy protection
By introducing digital signatures, blinding techniques, and hash chains into federated learning, the problems of privacy leakage and high overhead in federated learning are solved, achieving efficient data transmission and model training, and making it suitable for secure data synchronization from multiple data sources.
Patent Information
- Application Number
- CN202610223130.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-12
- Estimated Expiration
- 2046-02-25
AI Technical Summary
While existing federated learning methods can ensure data privacy and security, they also pose risks of privacy breaches and incur significant computational and communication overhead, and it is difficult to balance training accuracy with system overhead.
We adopt a privacy-preserving, fault-tolerant, layered, verifiable, and secure federated learning method. We set public parameters and identifiers during the initialization phase, use digital signatures and blinding techniques to protect local training results, perform local aggregation and secret sharing on edge servers, and perform global model updates on the aggregation server. We use hash chains to ensure the integrity of model parameters.
It achieves improved training accuracy and reduced computational and communication overhead while ensuring security, enhances data transmission security and privacy, and adapts to network latency issues of edge servers.
Smart Images

Figure CN121727867B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information security and communication technology, and in particular to a federated learning method for privacy-preserving, fault-tolerant, layered, verifiable, and secure aggregation. Background Technology
[0002] With the rapid development of artificial intelligence, prioritizing data privacy and security has become a trend. Traditional distributed machine learning architectures require clients to upload local data to cloud servers for processing. This undoubtedly increases the risk of data leakage for users. Due to data privacy and security constraints, multiple data sources cannot directly exchange data, creating a "data silo" phenomenon. To address the problem of data silos, federated learning has emerged.
[0003] Federated learning is a distributed learning architecture that, unlike traditional distributed systems, does not require clients to upload raw data. In federated learning, users train the model locally using the data and upload the trained results instead of the original data, thus offering privacy protection. However, the risk of privacy breaches still exists. Some attackers can obtain features from the user's uploaded model parameters. Some attackers may even be able to tamper with the user-uploaded training results. Therefore, there is a need to propose more secure federated learning methods that do not affect training accuracy and reduce computational and communication overhead. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure federated learning method that achieves high training accuracy and low computational and communication overhead while ensuring security.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] A privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure aggregation federated learning method is applied to a model training system, which includes an aggregation server. P Edge servers S i ,user U i,j , , Where m is the number of user groups or the total number of edge servers, and n is the number of users in each group, each user group corresponds to one edge server connection. The federated learning method includes the following four stages:
[0007] Initialization phase: The aggregation server sets common parameters and generates initialization parameters for each entity. Common parameters include setting elliptic curves and generators, collision-resistant hash functions, and hash chains to ensure the integrity of the global model. Initialization parameters include unique identifiers for users and edge servers, as well as public-private key pairs. The aggregation server distributes keys between users and edge servers through a secure channel.
[0008] Local model training and signature generation stage: Each user first uses a digital signature algorithm to verify the global model sent by the aggregation server; uses local data to train the model to obtain local model parameters, and uses blinding technology to add a mask to the local model parameters; signs the masked parameters and uploads them to the edge server corresponding to their respective group;
[0009] Local model aggregation and secret sharing stage: The edge server verifies the signature of the parameters uploaded by the user. After the verification is successful, the model is aggregated to obtain a local aggregation result. The local aggregation result is used to generate a digital signature using the private key of the aggregation server. The local aggregation result is then secretly shared to obtain a share. One share is reserved and the remaining shares are transmitted to other edge servers. Each edge server encodes the received share and its reserved share using a super-incrementing sequence. Finally, the digital signature and the encoding result are sent to the aggregation server.
[0010] Global model update phase: The aggregation server decodes the received encoding result to obtain the share; it reconstructs the result according to the secret sharing reconstruction algorithm to obtain the local aggregation result; it verifies whether the result is consistent with the received signature, and if they are inconsistent, the result is discarded; the consistent local aggregation result is globally aggregated to obtain the updated global model parameters, and the new round of global model parameters is added to the hash chain for model parameter integrity check.
[0011] Furthermore, the initialization phase specifically includes:
[0012] Aggregation server based on limited domain Z p Selecting from non-singular elliptic curves q The order generator G; the aggregation server initializes the hash chain and sets the initial hash value. h 0=0 and assign to h old = h 0, h old This represents the latest node value in the hash chain after the previous training round, but during the first training round, it represents the initial state hash node value; the aggregation server calculates the initial model parameters. W 0 The hash value is obtained and added to the hash chain; the nodes in the hash chain are calculated as follows: and , W t+1 Indicates the first t +1 round of global model parameters, h global This represents the hash value of the global model after this round of training. h new This indicates the node values that need to be added to the hash chain after this round of training, used to concatenate the training result of this round with the previous hash node; the aggregation server is for all users. U i,j Edge servers S i Initial model parameters are distributed. W 0 Learning rate And the latest node values added to the hash chain h new ;
[0013] Aggregator server P Generate a unique secure random number as the master seed for each user group. Each user U i,j A user-specific seed is derived based on the main seed and the user's own identity. ,in H For hash functions, Indicates the cascading symbol. id i,j For users' unique identification; the aggregation server is the first i The first group j individual users U i,j Generate public / private key pairs ,in , for the first i One edge server S i Generate public / private key pairs Among them, satisfying The aggregation server distributes keys and unique identifiers between users and edge servers through a secure channel.
[0014] Furthermore, the local model training and signature generation stage specifically includes:
[0015] user U i,j Calculate global model parameters W t hash function value And calculate with the locally stored hash function value. ,Will The hash function value issued by the aggregation server h new The comparison is performed; if they don't match, training ends; if they match, local training continues. U i,j Use the corresponding local data D i,j , No. t Round global model parameters W t and learning rate Perform local training to obtain user data. U i,j The corresponding number t Wheel local model parameters , This represents the loss function trained by the user based on local data;
[0016] user U i,j Blinding the local training results specifically includes: based on the user's... U i,j Exclusive seeds Seed i,j Generate random numbers ,in id i,k It is the first i Group 1 k A unique identifier for each user PRG It is a pseudo-random function generator and satisfies ; in the i Among the users in the group, U i,j Expand the set of random numbers sequentially For each element, convert it to the corresponding l 3D random vector; user U i,j Generate User U i,k No. t Random mask vector during round training ,when j > k State value ;when j = k State value ;when j < k State value Set the opposite mask vector. ,satisfy When, equation Set the model parameters after adding the mask to... The local model parameters are expanded as followsl dimensional vector The random mask vector is expanded as l dimensional vector ;
[0017] Using user private key sk i,j Calculate the hash function value Calculate the coordinates of the points on the elliptic curve. Coordinate point two Generate hash function value and digital signatures , k i,j Indicates the first i Group 1 j The user's identifier value; the model parameters after adding the mask. and digital signatures Send to the edge servers of the group S i .
[0018] Furthermore, the local model aggregation and secret sharing phase specifically includes:
[0019] No. i One edge server S i Based on the local training results received from users within the group Signature of the received document Verification: Calculate the hash function value The coordinates of the elliptic curve are three. and coordinate point four If the calculated hash function value and elliptic curve coordinates do not match the digital signature uploaded by the user, the result is discarded and the training terminates; if they match, the federated averaging algorithm is used for local aggregation to obtain the result. i One edge server S i In the t Local aggregation results of round training: ,in Edge servers use their own private keys. sk i Obtain by signing ;
[0020] Use the secret sharing algorithm to analyze the local aggregation results. To perform secret sharing, the local aggregation result is expanded into an l-dimensional vector: ,Will Write the polynomial middle, ,Right now ,in ;
[0021] Edge server S i The computation needs to be distributed to other edge servers. S k Local aggregation model Shared share ,in x k For edge servers S k The unique identifier value, and send the share To edge servers S k Edge server S i The share-sharing list that needs to be sent is ,in For edge servers S k The list of shares received from other edge servers is compiled as follows: ;
[0022] No. v One edge server S v All shares Encoding is performed using a super-increasing sequence to obtain the encoded value. The encoded result and signature are then sent to the aggregation server.
[0023] Furthermore, the global model update phase specifically includes:
[0024] The aggregation server will receive the encoded results Decoding is performed, and the decoding process is as follows: based on and The encoded result is decoded step by step, where Represents edge server S m Distributed to edge servers S v Local aggregation model share, This indicates the floor function, which ultimately yields the share for each edge server. ;
[0025] The shares are obtained by organizing them, and then reconstructed using the Lagrange interpolation formula based on the share list. as well as ,get polynomial coefficients Extract and obtain the local aggregation result. ;
[0026] Based on the edge server's public key, signature, and reconstruction result, each local aggregation result is verified. If verification fails, training terminates; if verification succeeds, a federated averaging algorithm is used for global aggregation to obtain the result. t +1 round of global model parameters ;
[0027] Update global model parameters W t+1 Write it into the hash chain and calculate the hash value for global updates of the model. And concatenation with the previous hash node And store it on the aggregation server to facilitate subsequent verification of the integrity of global model parameters;
[0028] A new global model is distributed to each user. After receiving the global model, the user performs integrity verification. If the verification is successful, the next round of training begins; otherwise, training is terminated.
[0029] The beneficial effects of this invention are:
[0030] The aggregation server distributes seeds to groups, combined with the user's local identifier, to generate a user-specific mask, enhancing the security and privacy of data transmission. The user's local model training results are not uploaded directly but with an added mask, preventing attackers from obtaining sensitive user information from the transmitted data. Furthermore, the digital signature design is based on elliptic curve cryptography, utilizing computationally efficient doubling operations, enabling a large number of rapid calculations during signing and verification.
[0031] This invention employs a secret sharing technique between edge servers and aggregation servers to secretly share local aggregation results. A super-increasing sequence is used to encode shares, avoiding direct transmission. Digital signature technology is used to sign the encoded results, ensuring the integrity of the transmission process. A hash chain on the aggregation server side primarily ensures that the global model updated by the aggregation server is consistent with the global model to be distributed to users. This invention achieves a balance between efficiency and overhead, while allowing for network latency issues at the edge servers. Compared to privacy-preserving methods in traditional federated learning, this invention not only maintains high training accuracy but also significantly reduces communication and computational overhead. Attached Figure Description
[0032] Figure 1 A flowchart illustrating a privacy-preserving, fault-tolerant, layered, verifiable, and secure federated learning method.
[0033] Figure 2This is a schematic diagram of the model training system structure. Detailed Implementation
[0034] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.
[0035] See Figures 1-2 The present invention provides a technical solution:
[0036] A privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure aggregation federated learning method is applied to a model training system, which includes an aggregation server. P Edge servers S i ,user U i,j , , m is the number of user groups or the total number of edge servers, and n is the number of users in each group. Each user group corresponds to one edge server connection; for example... Figure 1 As shown, the federated learning method includes the following four stages:
[0037] Initialization phase: The aggregation server sets common parameters and generates initialization parameters for each entity. Common parameters include setting elliptic curves and generators, collision-resistant hash functions, and hash chains to ensure the integrity of the global model. Initialization parameters include unique identifiers for users and edge servers, as well as public-private key pairs. The aggregation server distributes keys between users and edge servers through a secure channel.
[0038] Local model training and signature generation stage: Each user first uses a digital signature algorithm to verify the global model sent by the aggregation server; uses local data to train the model to obtain local model parameters, and uses blinding technology to add a mask to the local model parameters; signs the masked parameters and uploads them to the edge server corresponding to their respective group;
[0039] Local model aggregation and secret sharing stage: The edge server verifies the signature of the parameters uploaded by the user. After the verification is successful, the model is aggregated to obtain a local aggregation result. The local aggregation result is used to generate a digital signature using the private key of the aggregation server. The local aggregation result is then secretly shared to obtain a share. One share is reserved and the remaining shares are transmitted to other edge servers. Each edge server encodes the received share and its reserved share using a super-incrementing sequence. Finally, the digital signature and the encoding result are sent to the aggregation server.
[0040] Global model update phase: The aggregation server decodes the received encoding result to obtain the share; it reconstructs the result according to the secret sharing reconstruction algorithm to obtain the local aggregation result; it verifies whether the result is consistent with the received signature, and if they are inconsistent, the result is discarded; the consistent local aggregation result is globally aggregated to obtain the updated global model parameters, and the new round of global model parameters is added to the hash chain for model parameter integrity check.
[0041] Existing federated learning privacy protection techniques struggle to balance training accuracy with overall system overhead. This invention discloses a privacy-preserving, fault-tolerant, layered, verifiable, and secure aggregation federated learning method. The method employs a three-layer architecture for training without introducing complex cryptographic primitives, achieving a balance between privacy protection and system overhead. The first layer, the user layer, primarily performs local training and introduces blinding techniques to protect local training results. The second layer, the edge server layer, primarily implements local aggregation and share sharing, encoding the shared shares to avoid direct transmission. The third layer, the aggregation server layer, primarily performs parameter reconstruction for global model updates.
[0042] In one specific embodiment, the S1 initialization phase specifically includes:
[0043] Aggregation server based on limited domain Z p Selecting from non-singular elliptic curves q The order generator G; the aggregation server initializes the hash chain and sets the initial hash value. h 0=0 and assign to h old = h 0, h old This represents the latest node value in the hash chain after the previous training round, but during the first training round, it represents the initial state hash node value; the aggregation server calculates the initial model parameters. W 0 The hash value is obtained and added to the hash chain; the nodes in the hash chain are calculated as follows: and , W t+1 Indicates the first t +1 round of global model parameters, h global This represents the hash value of the global model after this round of training. h new This indicates the node values that need to be added to the hash chain after this round of training, used to concatenate the training result of this round with the previous hash node; the aggregation server is for all users. U i,j Edge serversS i Initial model parameters are distributed. W 0 Learning rate And the latest node values added to the hash chain h new ;
[0044] Aggregator server P Generate a unique secure random number as the master seed for each user group. Each user U i,j A user-specific seed is derived based on the main seed and the user's own identity. ,in H For hash functions, Indicates the cascading symbol. id i,j For users' unique identification; the aggregation server is the first i The first group j individual users U i,j Generate public / private key pairs ,in , for the first i One edge server S i Generate public / private key pairs Among them, satisfying The aggregation server distributes keys and unique identifiers between users and edge servers via a secure channel. The unique identifier of the edge server is... x i .
[0045] An elliptic curve-based digital signature technology was designed during the initialization phase, while a hash chain was introduced to protect the integrity of the data.
[0046] In one specific embodiment, the S2 local model training and signature generation stage specifically includes:
[0047] user U i,j Calculate global model parameters W t hash function value And calculate with the locally stored hash function value. ,Will The hash function value issued by the aggregation server h new The comparison is performed; if they don't match, training ends; if they match, local training continues. U i,j Use the corresponding local data Di,j , No. t Round global model parameters W t and learning rate Perform local training to obtain user data. U i,j The corresponding number t Wheel local model parameters , This represents the loss function trained by the user based on local data;
[0048] user U i,j Blinding the local training results specifically includes: based on the user's... U i,j Exclusive seeds Seed i,j Generate random numbers ,in id i,k It is the first i Group 1 k A unique identifier for each user PRG It is a pseudo-random function generator and satisfies ; in the i Among the users in the group, U i,j Expand the set of random numbers sequentially For each element, convert it to the corresponding l 3D random vector; user U i,j Generate User U i,k No. t Random mask vector during round training ,when j > k State value ;when j = k State value ;when j < k State value Set the opposite mask vector. ,satisfy When, equation Set the model parameters after adding the mask to... The local model parameters are expanded as follows l dimensional vector The random mask vector is expanded as l dimensional vector ;
[0049] Using user private key sk i,jCalculate the hash function value Calculate the coordinates of the points on the elliptic curve. Coordinate point two Generate hash function value and digital signatures , k i,j Indicates the first i Group 1 j The user's identifier value; the model parameters after adding the mask. and digital signatures Send to the edge servers of the group S i .
[0050] The local model training and signature generation phase primarily involves global model verification and local model training, followed by blinding the training results. The blinded results are then digitally signed, and the signature and blinding results are packaged and sent to the edge server. The aggregation server distributes seeds to the groups, combined with the user's local identifier, to generate a user-specific mask, enhancing the security and privacy of data transmission. The user's local model training results are not uploaded directly but with an added mask, preventing attackers from obtaining sensitive user information from the transmitted data. Furthermore, the digital signature design is based on elliptic curve cryptography, utilizing computationally efficient doubling operations, enabling rapid and extensive computation during the signing and verification process.
[0051] In one specific embodiment, the local model aggregation and secret sharing phase specifically includes:
[0052] No. i One edge server S i Based on the local training results received from users within the group Signature of the received document Verification: Calculate the hash function value The coordinates of the elliptic curve are three. and coordinate point four If the calculated hash function value and elliptic curve coordinates do not match the digital signature uploaded by the user, the result is discarded and the training terminates; if they match, the federated averaging algorithm is used for local aggregation to obtain the result. i One edge server S i In the t Local aggregation results of round training: ,in Edge servers use their own private keys. sk i Obtain by signing ;
[0053] Use the secret sharing algorithm to analyze the local aggregation results. To perform secret sharing, the local aggregation result is expanded into an l-dimensional vector: ,Will Write the polynomial middle, ,Right now ,in ;
[0054] Edge server S i The computation needs to be distributed to other edge servers. S k Local aggregation model Shared share ,in x k For edge servers S k The unique identifier value, and send the share To edge servers S k Edge server S i The share-sharing list that needs to be sent is ,in For edge servers S k The list of shares received from other edge servers is compiled as follows: ;
[0055] No. v One edge server S v All shares Encoding is performed using a super-increasing sequence to obtain the encoded value. The encoded result and signature are then sent to the aggregation server.
[0056] The local aggregation phase and secret sharing phase mainly perform signature verification, deblinded local aggregation, secret sharing, and super-incremental encoding. This invention applies secret sharing technology between the edge server and the aggregation server to secretly share the local aggregation results. Super-incremental sequences are used to encode shares, avoiding direct transmission of shares. Simultaneously, digital signature technology is used to sign the encoded results, ensuring the integrity of the transmission process.
[0057] In one specific embodiment, the global model update phase specifically includes:
[0058] The aggregation server will receive the encoded results Decoding is performed, and the decoding process is as follows: based on and The encoded result is decoded step by step, where Represents edge server S m Distributed to edge servers S v Local aggregation model share, This indicates the floor function, which ultimately yields the share for each edge server. ;
[0059] The shares are obtained by organizing them, and then reconstructed using the Lagrange interpolation formula based on the share list. as well as You can get polynomial coefficients Extract and obtain the local aggregation result. ;
[0060] Based on the edge server's public key, signature, and reconstruction result, each local aggregation result is verified. If verification fails, training terminates; if verification succeeds, a federated averaging algorithm is used for global aggregation to obtain the result. t +1 round of global model parameters ;
[0061] Update global model parameters W t+1 Write it into the hash chain and calculate the hash value for global updates of the model. And concatenation with the previous hash node And store it on the aggregation server to facilitate subsequent verification of the integrity of global model parameters;
[0062] A new global model is distributed to each user. After receiving the global model, the user performs integrity verification. If the verification is successful, the next round of training begins; otherwise, training is terminated.
[0063] In the global model update phase, the received encoded results are decoded and reconstructed to obtain the original local aggregation results. A federated averaging algorithm is used for global updates, and the gradient global update results are written to a hash chain. This hash chain, located on the aggregation server, primarily ensures that the global model updated by the aggregation server is consistent with the global model to be distributed to users. This invention achieves a balance between efficiency and overhead, while allowing for network latency issues at edge servers.
[0064] Next, we need to analyze the correctness of the local training results and the global model update in the federated learning method provided by this invention:
[0065] (1) Correctness of local training results
[0066] Each group has n Individual users, users in a group j Model parameters obtained after local training To perform secure aggregation, a random mask is added to the parameters of the locally trained model: ;
[0067] Random mask satisfies The edge server performs the aggregation operation: ;
[0068] The aggregation server will restructure the obtained shares according to the following formula, because... The polynomial can be reconstructed based on its share: ;
[0069] Parameters for building the first edge server S1:
[0070]
[0071] According to the formula, a polynomial of degree t-1 can be obtained. According to the polynomial, we can obtain Based on the general formula, the local aggregation results for each group can be obtained.
[0072] (2) Correctness of global model update
[0073] Throughout the entire scheme, the assumed entities are all honest but curious, so the correctness of the global model update only requires proving that the data has not been tampered with during transmission.
[0074] The transmission between the aggregation server and users primarily uses hash chains for verification. If an attacker attempts to tamper with data... ,but This would cause the hash node to change. An attacker would then have to find a value that makes... This creates a hash collision, violating the collision resistance property of hashes. Therefore, it ensures that the transmission between the aggregation server and the user is complete and has not been tampered with.
[0075] The communication between users and edge servers is achieved through a signature algorithm. Users sign the blinded result and send the signature and result to the edge server.
[0076] Signature process:
[0077]
[0078] in Send signature and blinding model parameters Send the data to the edge server. The edge server performs the verification operation.
[0079] Verification process:
[0080]
[0081] By checking whether the formulas are equal, we can determine whether the parameters of the blinded model have been tampered with during transmission.
[0082] Data transmission between edge servers and aggregation servers is also guaranteed to be intact through signatures.
[0083] This invention employs a blinding technique, specifically adding a random mask, to protect local training results. Since the uploaded training results contain other data, even if an attacker obtains this data, they cannot derive the user's data features from it. Edge servers can remove the blinding through local aggregation, and the added masks cancel each other out during the aggregation process. That is, the local aggregation result on the edge server is an aggregation of the user's local training results. After aggregation, the edge server uses a secret sharing technique to send the results to other edge servers. The uploaded share is not directly uploaded but encoded before being sent, protecting data privacy during transmission. Compared to privacy protection methods in traditional federated learning, this invention not only maintains high training accuracy but also significantly reduces communication and computational overhead.
[0084] This invention can be widely applied in scenarios where disease diagnosis models are trained using data from multiple hospitals. Since patient information cannot be directly uploaded, each hospital can train its model based on its own data and upload the training results to edge nodes set up for that region. The edge nodes perform local aggregation, but they cannot determine which result corresponds to which hospital. After aggregation, the edge nodes upload the results to the final aggregation node, which also cannot determine which result corresponds to which region. This achieves synchronous data training while protecting data privacy.
[0085] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A federated learning method based on privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure aggregation, characterized in that: Applied to a model training system, the model training system including an aggregation server P Edge servers S i ,user U i,j , , Where m is the number of user groups or the total number of edge servers, and n is the number of users in each group, each user group corresponds to one edge server connection. The federated learning method includes the following four stages: Initialization phase: The aggregation server sets common parameters and generates initialization parameters for each entity. Common parameters include setting elliptic curves and generators, collision-resistant hash functions, and hash chains to ensure the integrity of the global model. Initialization parameters include unique identifiers for users and edge servers, as well as public-private key pairs. The aggregation server distributes keys between users and edge servers through a secure channel. Local model training and signature generation stage: Each user first uses a digital signature algorithm to verify the global model sent by the aggregation server; uses local data to train the model to obtain local model parameters, and uses blinding technology to add a mask to the local model parameters; signs the masked parameters and uploads them to the edge server corresponding to their respective group; Local model aggregation and secret sharing stage: The edge server verifies the signature of the parameters uploaded by the user. After the verification is successful, the model is aggregated to obtain a local aggregation result. The local aggregation result is used to generate a digital signature using the private key of the aggregation server. The local aggregation result is then secretly shared to obtain a share. One share is reserved and the remaining shares are transmitted to other edge servers. Each edge server encodes the received share and its reserved share using a super-incrementing sequence. Finally, the digital signature and the encoding result are sent to the aggregation server. Global model update phase: The aggregation server decodes the received encoding result to obtain the share; it reconstructs the result according to the secret sharing reconstruction algorithm to obtain the local aggregation result; it verifies whether the result is consistent with the received signature, and if they are inconsistent, the result is discarded; the consistent local aggregation result is globally aggregated to obtain the updated global model parameters, and the new round of global model parameters is added to the hash chain for model parameter integrity check.
2. The federated learning method for privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure aggregation based on layered structures as described in claim 1, characterized in that: The initialization phase specifically includes: Aggregation server based on limited domain Z p Selecting from non-singular elliptic curves q The order generator G; the aggregation server initializes the hash chain and sets the initial hash value. h 0=0 and assign to h old = h 0, h old This represents the latest node value in the hash chain after the previous training round, but during the first training round, it represents the initial state hash node value; the aggregation server calculates the initial model parameters. W 0 The hash value is obtained and added to the hash chain; the nodes in the hash chain are calculated as follows: and , W t+1 Indicates the first t +1 round of global model parameters, h global This represents the hash value of the global model after this round of training. h new This indicates the node values that need to be added to the hash chain after this round of training, used to concatenate the training result of this round with the previous hash node; the aggregation server is for all users. U i,j Edge servers S i Initial model parameters are distributed. W 0 Learning rate And the latest node values added to the hash chain. h new ; Aggregator server P Generate a unique secure random number as the master seed for each user group. Each user U i,j A user-specific seed is derived based on the main seed and the user's own identity. ,in H For hash functions, Indicates the cascading symbol. id i,j For users' unique identification; the aggregation server is the first i The first group j individual users U i,j Generate public / private key pairs ,in , for the first i Edge servers S i Generate public / private key pairs Among them, satisfying The aggregation server distributes keys and unique identifiers between users and edge servers through a secure channel.
3. The federated learning method based on privacy protection, fault-tolerant hierarchical verifiable secure aggregation according to claim 2, characterized in that: The local model training and signature generation stage specifically includes: user U i,j Calculate global model parameters W t hash function value And calculate with the locally stored hash function value. ,Will The hash function value issued by the aggregation server h new The comparison is performed; if they don't match, training ends; if they match, local training continues. U i,j Use the corresponding local data D i,j , No. t Round global model parameters W t and learning rate Perform local training to obtain user data. U i,j The corresponding number t Wheel local model parameters , This represents the loss function trained by the user based on local data; user U i,j Blinding the local training results specifically includes: based on the user's... U i,j Exclusive seeds Seed i,j Generate random numbers ,in id i,k It is the first i Group 1 k A unique identifier for each user PRG It is a pseudo-random function generator and satisfies ; in the i Among the users in the group, U i,j Expand the set of random numbers sequentially For each element, convert it to the corresponding l 3D random vector; user U i,j Generate User U i,k No. t Random mask vector during round training ,when j > k State value ;when j = k State value ;when j < k State value Set the opposite mask vector. ,satisfy When, equation Set the model parameters after adding the mask to... The local model parameters are expanded as follows l dimensional vector The random mask vector is expanded as follows l dimensional vector ; Using user private key sk i,j Calculate the hash function value Calculate the coordinates of the points on the elliptic curve. Coordinate point two Generate hash function value and digital signatures , k i,j Indicates the first i Group 1 j The user's identifier value; the model parameters after adding the mask. and digital signatures Send to the edge servers of the group S i .
4. The federated learning method based on privacy protection, fault-tolerant hierarchical verifiable secure aggregation according to claim 3, characterized in that: The local model aggregation and secret sharing phase specifically includes: No. i Edge servers S i Based on the local training results received from users within the group Signature of the received document Verification: Calculate the hash function value The coordinates of the elliptic curve are three. and coordinate point four If the calculated hash function value and elliptic curve coordinates do not match the digital signature uploaded by the user, the result is discarded and the training terminates; if they match, the federated averaging algorithm is used for local aggregation to obtain the result. i Edge servers S i In the t Local aggregation results of round training: ,in Edge servers use their own private keys. sk i Obtain by signing ; Use the secret sharing algorithm to analyze the local aggregation results. To perform secret sharing, the local aggregation result is expanded into an l-dimensional vector: ,Will Write the polynomial middle, ,Right now ,in ; Edge server S i The computation needs to be distributed to other edge servers. S k Local aggregation model Shared share ,in x k For edge servers S k The unique identifier value, and send the share To edge servers S k Edge server S i The share-sharing list that needs to be sent is ,in For edge servers S k The list of shares received from other edge servers is compiled as follows: ; No. v Edge servers S v All shares Encoding is performed using a super-increasing sequence to obtain the encoded value. The encoded result and signature are then sent to the aggregation server.
5. The federated learning method for privacy-preserving, fault-tolerant, hierarchical, verifiable, and secure aggregation based on layered structures, as described in claim 4, is characterized in that: The global model update phase specifically includes: The aggregation server will receive the encoded results Decoding is performed, and the decoding process is as follows: based on and The encoded result is decoded step by step, where Represents edge server S m Distributed to edge servers S v Local aggregation model share, This indicates the floor function, which ultimately yields the share for each edge server. ; The shares are obtained by organizing them, and then reconstructed using the Lagrange interpolation formula based on the share list. as well as ,get polynomial coefficients Extract and obtain the local aggregation result. ; Based on the edge server's public key, signature, and reconstruction result, each local aggregation result is verified. If verification fails, training terminates; if verification succeeds, a federated averaging algorithm is used for global aggregation to obtain the result. t +1 round of global model parameters ; Update global model parameters W t+1 Write it into the hash chain and calculate the hash value for global updates of the model. And concatenation with the previous hash node And store it on the aggregation server to facilitate subsequent verification of the integrity of global model parameters; A new global model is distributed to each user. After receiving the global model, the user performs integrity verification. If the verification is successful, the next round of training begins; otherwise, training is terminated.
