A secure and efficient non-interactive federated learning framework, its construction method, and learning method

Abstract

This paper presents a secure and efficient non-interactive federated learning framework, its construction method, and its learning method, relating to the field of federated learning modeling technology. It addresses the problems of high communication overhead and privacy leaks caused by model attacks in existing technologies. In the learning framework, multiple non-interactive distributed nodes jointly designate a server as a trainer; multiple trainers and interactive distributed computing nodes designate a server as a central aggregator, and each trainer is associated with at least one non-interactive node. The construction method involves the interactive distributed computing nodes and multiple trainers first downloading the global model from the selected central aggregator. After each interactive distributed computing node and each trainer trains a sub-model, it uploads it as a local model to the central aggregator. The central aggregator receives the local models, updates the global model, and returns the results to the interactive distributed computing nodes and multiple trainers. This process is repeated multiple times until the global model reaches the expected accuracy, completing the construction. This framework is applicable to the field of connected vehicle technology.

Description

Technical Field

[0001] This invention relates to the field of federated learning modeling technology, and to a secure and efficient non-interactive federated learning framework in which an edge server accesses a large number of encrypted training sample model parameters, label data, and security between the client and the edge server. Background Technology

[0002] Most machine learning (ML) solutions rely on a central platform that collects data from various clients and trains a global model. However, this approach is unsuitable for scenarios where client behavior involves confidential data, potentially severely impacting environmental security and privacy. Federated learning (FL) has become highly relevant as a collaborative ML paradigm in recent years, where data privacy and integrity are paramount.

[0003] In traditional FL (Flexible Training) setups, a subset of clients is randomly selected for training each round. Local updates are then performed on these selected clients over several epochs, and finally, the updated local models are aggregated on the server to obtain the global model. It's easy to see that this training paradigm requires significant communication overhead and limits the scalability of multi-node training. The situation worsens when clients are resource-constrained mobile devices. Clients with poor wireless channel conditions may also face higher latency, lower throughput, and intermittent disconnections. Therefore, reducing the additional communication burden of model training will improve the overall efficiency of FL. Furthermore, the extra processing required for disconnected clients also incurs computational and communication overhead, and system performance decreases as the number of client disconnections increases.

[0004] Traditional client-cloud two-tier architecture-based FL models struggle to balance high communication efficiency and high model performance. Cloud servers can handle massive amounts of data, but transmission time is directly proportional to the data volume. This means that outsourcing data processing to the cloud can lead to response delays. Edge computing can address this issue. Integrating edge servers into a traditional FL model allows it to leverage edge computing and offload more computational tasks to edge devices. By adding local training epochs or more computations on edge nodes before global aggregation on the server, the local model becomes more accurate, reducing communication costs and the number of FL training epochs. Therefore, a layered FL model based on a client-edge-cloud three-tier architecture can effectively reduce communication costs and improve communication efficiency. Using FL model training in edge computing scenarios can also address data privacy concerns related to the client.

[0005] The emergence of the Internet of Vehicles (IoV) aims to improve user experience by offering more diverse services, ranging from ensuring customer safety to enhancing customer comfort. With the deployment of 5G, vehicle slicing, and edge computing, the development of the human-vehicle-road-edge server architecture will truly achieve synergy. Although edge computing-based IoV frameworks are gradually maturing, vehicle data in IoV still involves privacy-sensitive data such as driver route information and personal information. Therefore, a privacy-protecting inappropriate behavior detection system for IoV based on the traditional FL (Flexible Link) model with a Client-Cloud two-layer architecture is proposed. Vehicles frequently need to communicate in real-time with cloud servers located remotely. This process not only introduces high latency issues but also puts significant pressure on the cloud servers. The combination of FL and edge computing in IoV can ensure secure and efficient modeling.

[0006] Layered FL (Client-Edge-Cloud) architecture still faces two challenges. The first is that layered FL remains interactive, imposing significant communication overhead between the client and edge servers, which is difficult to manage in practice, requiring the client to iterate intermediate results on its training samples. The second challenge is the risk of model attacks during training due to iterative updates of plaintext model parameters, leading to privacy breaches of training samples. Training samples typically contain a large amount of statistical data from an organization, and disclosing this information could cause financial losses. Summary of the Invention

[0007] This invention solves the problems of high communication overhead and privacy leakage caused by model attacks in the prior art.

[0008] The purpose of this invention is to provide a non-interactive regression training framework, AuthENIFL, which enables non-interactive regression training and model updates while protecting participants' local data, effectively reducing computational costs and ultimately achieving a well-trained model.

[0009] To achieve the objectives of this invention, the specific technical solution provided is as follows:

[0010] Technical Solution 1: A secure and efficient non-interactive federated learning framework, comprising the following components:

[0011] Includes multiple servers and multiple non-interactive distributed computing nodes (UEs). * and multiple interactive distributed computing nodes (UEs) i Among them, non-interactive distributed computing nodes (UEs) * No powerful local computing server; multiple non-interactive distributed computing nodes (UEs) *A single server is designated as the trainer MEC; multiple trainer MECs and interactive distributed computing nodes (UEs) are involved. i A server is designated as the central aggregator (CSP), and each trainer MEC is paired with at least one non-interactive distributed computing node (UE). * It is associated with a central aggregator (CSP) that contains multiple participant users (UEs).

[0012] This invention also provides a method for constructing the aforementioned secure and efficient non-interactive federated learning framework, the method being as follows:

[0013] 1) Each non-interactive distributed computing node (UE) * The encrypted data is uploaded to the designated trainer MEC, and then the interactive distributed computing node UE... i Multiple trainer MECs download the global model from the designated central aggregator CSP, and each interactive distributed computing node (UE) downloads the model. i After training the sub-model with each trainer MEC, it is uploaded as a local model to the central aggregator CSP.

[0014] 2) After receiving the local model, the central aggregator (CSP) updates the global model and returns the global model to the interactive distributed computing node (UE). i And each trainer's MEC;

[0015] 3) Repeat steps 1) and 2) above repeatedly, updating and iterating multiple times until the global model reaches the expected accuracy, thus completing the construction.

[0016] This invention also provides a non-interactive federated learning method, which is implemented based on the non-interactive federated learning framework described in claim 1, and the method includes the following steps:

[0017] Step 1: Initialization. A central aggregator (CSP) is set up. A homomorphic encryption scheme is selected based on additive homomorphism. A public key and a private key are generated based on the selected homomorphic encryption scheme.

[0018] Step 2: There are m participants (UEs), each interactive node (UE) i We have a dataset, 1 ≤ i ≤ m, which includes feature data X. i and label data Y i Let m be an example. * Each participant corresponds to a non-interactive distributed computing node (UE*), and each... We have a dataset, 1≤j≤m * The dataset includes feature data X j and label data Y jAll participants' corresponding non-interactive distributed computing nodes (UEs) * Each participant is assigned a trainer MEC, and each participant corresponds to a non-interactive distributed computing node UE. * and interactive node UE i The dataset is encrypted using a public key and then sent to the designated trainer MEC and central aggregator CSP, respectively.

[0019] Step 3: The designated trainer MEC replaces UE. * Non-interactive model training is performed between the central aggregator (CSP) and the participants (UEs) and the CSP. Interactive model training is then conducted between the CSP and the designated trainer (MEC) and each participant (UE). After multiple iterations, the training continues. i Obtain a local model M mec and M i ;

[0020] Step 4: The designated trainer MEC and each participant UE i Local model M mec and M i Update to the central aggregator (CSP); after receiving the local model, the central aggregator updates the global model and returns it to the trainer MEC and each participant UE on the mobile edge server. i This process is repeated until the global model update reaches the expected accuracy.

[0021] Preferably, a preferred embodiment is provided, wherein the encryption scheme in step 1 is specifically as follows:

[0022] The homomorphic encryption scheme employs Paillier encryption, ECC encryption, and secret sharing technology. Enc is defined to represent encrypting a data record, and Dec is defined to represent decrypting a ciphertext message.

[0023] Preferably, each participant UE mentioned in step 2 * The dataset is encrypted using a public key as follows:

[0024] Step A1: Randomly select two non-singular integer matrices A j and B j Given a random number R, calculate A j X j Y j RB and B -1 X j ;

[0025] Step A2: Pass B and A through a secure passage where there are no eavesdroppers. j X j and Enc(Y jRB) is sent to MEC, along with A and B. -1 X j Send to the central aggregator (CSP).

[0026] Preferably, the global model update in step 4 specifically refers to:

[0027] Step B1: The central aggregator (CSP) sends the model parameters RW corresponding to the global model to the interactive distributed computing node (UE). i The model parameters RWA corresponding to the global model j Give MEC to the trainer;

[0028] Step B2: Interactive Distributed Computing Node (UE) i Obtain the global model parameter RW, and then, based on the global model parameter RW, interact with the distributed computing node UE. i For X i and Y i Perform gradient descent algorithm to update the local model and obtain the local model parameters W. i Finally, UE i Set local model parameters W i The corresponding model M i Send to the central aggregator (CSP).

[0029] Step B3: The trainer's MEC downloads the global model and obtains the global model parameters RWA. j Then, based on the global model parameters RWA j , for A j X j 、Enc(Y j RB) performs local model updates to obtain the encrypted local gradient Enc(L) mec Finally, MEC will convert the encrypted ground gradient Enc(L) into a single gradient. mec The corresponding model M) mec Send to the central aggregator (CSP).

[0030] Step B4: The central aggregator CSP, according to M i Get RW i According to M mec Get Enc(L) mec Then CSP according to RW i 、Enc(L mec ) and B -1 X j Perform aggregated average update to obtain global model parameters, thereby obtaining the global model.

[0031] Step B5: Iterate through steps B1-B4 until the model converges to a certain threshold or reaches the maximum number of iterations.

[0032] Preferably, step B2 involves the interactive distributed computing node (UE) based on the global model parameters RW. i For X i and Y i The process of performing local model updates using the gradient descent algorithm is as follows:

[0033] Through calculation Where α represents the learning rate, which is typically set to 0.001. Let L(x) be the first derivative with respect to W. i ,y i ) = Y i log(h(WX i ))+(1-Y i )log(1-h(WX i h() represents a logical function.

[0034] Preferably, step B3 involves using the global model parameter RWA. j , for A j X j 、Enc(Y j The specific steps for RB to perform a local model are as follows:

[0035] By calculating WX j =R -1 RW1A j -1 A j X j To eliminate the blindness caused by the random mask R and A, MEC can compute the gradient Enc(L). mec )=Enc(R(h(WX j )-Y j B j =Enc(Rh(WX) j B j Enc(Y j RB j ) -1 .

[0036] Preferably, the CSP in step B4 is based on RW i 、Enc(L mec ) and B -1 X j The specific steps for performing aggregate averaging are as follows:

[0037] First, calculate R(h(WX) j )-Y j B j =Dec(Enc(L) mec )),

[0038] Then calculate (L(x) j ,y j ))=R(h(WX j )-Y j B j B j -1 X j and

[0039] Final calculation

[0040] Preferably, the gradient descent described in steps B2 and B3 specifically involves updating using full gradient descent, mini-batch gradient descent, or stochastic gradient descent.

[0041] The beneficial effects of this invention are as follows:

[0042] In this invention, AuthENIFL, NPMML

[21] , and VAN

[21] are all non-interactive FLs, and UE * All tasks are completed offline. Therefore, UE * The communication cost is only one round of interaction, and no computational overhead or communication is generated during the training process, thereby reducing computational overhead and communication data volume.

[0043] The computational overhead of UE* in the non-interactive framework was measured by uploading each dataset n = 1000, 2000, 3000, 4000, 5000, 5000, 1000, each participating client m* = 100, 200, 300, 400, 500, 1000, and each processing feature dimension d = 50, 100, 150, 200, 250, 300.

[0044] Secondly, the method described in this invention significantly reduces communication overhead.

[0045] Compared to edge-based federated learning (FL), AuthENIFL complements efficient client-side edge updates, ensuring higher communication efficiency, scalability in the face of change, and enhanced security. AuthENIFL allows some clients to perform federated learning without interacting with other clients. This invention constructs a non-interactive global regression training protocol and designs a non-interactive privacy training protocol, enhancing identity privacy protection and traceability, and establishing a stable and secure connection at the network edge layer. Attached Figure Description

[0046] Figure 1 This is a flowchart of a non-interactive federated learning method described in Embodiment 3.

[0047] Figure 2It is an architecture diagram of a secure and efficient non - interactive federated learning framework graph described in Embodiment 1.

[0048] Figure 3 It is a comparison graph of the loss convergence rate of the regression model. In the graph, ACCURACY EVALUATION OF LOGISTICREGRESSION is the accuracy evaluation of Logistic regression, where n is the number of samples; d is the dimension; a is the learning rate; L is the number of iterations; TND: training data set; TTD: test data set.

[0049] UCID represents the US Census Income data set, which is a public data set obtained from the website https: / / archive.ics.uci.edu / ml / datasets / census+income.

[0050] DD is the diabetes data set, which is a public data set obtained from the website https: / / datahub.io / machine - learning / diabetes.

[0051] BMD is the bank marketing data set, which is a public data set in marketing obtained from the website https: / / archive.ics.uci.edu / ml / datasets / Bank+.

[0052] The MNIST data set is public data obtained from the website http: / / yann.lecun.com / exdb / mnist / .

[0053] Figure 4 It is a comparison graph of the computational overheads experimented in the AuthENIFL framework and NPMML

[21] . VAN

[21] increases significantly with the increase of computational time and feature dimension; NPMML

[21] increases slowly with the increase of computational time and feature dimension; while AuthENIFL gradually tends to a straight line with the increase of computational time and feature dimension; in the graph, the Chinese translation of Computation Time(s) is "计算时间(s)", and the Chinese translation of features is "特征".

[0054] Figure 5 、 6 and 7 are comparison graphs of the communication overheads between the AuthENIFL framework and the client - edge cloud hierarchical FL[7] respectively. Figure 5 、 6 In and 7, the Chinese translation of the ordinate Communication(MB) is "通信(MB)", Figure 7 and the Chinese translation of the abscissa samples is "样本". Detailed Implementation

[0055] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0056] Implementation Method 1, see [link] Figure 2 This embodiment describes a secure and efficient non-interactive federated learning framework, which includes the following components:

[0057] Includes multiple servers and multiple non-interactive distributed computing nodes (UEs). * and multiple interactive distributed computing nodes (UEs) i Among them, non-interactive distributed computing nodes (UEs) * No powerful local computing server; multiple non-interactive distributed computing nodes (UEs) * A single server is designated as the trainer MEC; multiple trainer MECs and interactive distributed computing nodes (UEs) are involved. i A server is designated as the central aggregator (CSP), and each trainer MEC is paired with at least one non-interactive distributed computing node (UE). * It is associated with a central aggregator (CSP) that contains multiple participant users (UEs).

[0058] Implementation Method Two: This implementation method describes a method for constructing a secure and efficient non-interactive federated learning framework as described in Implementation Method One. The method is as follows:

[0059] 1) Each non-interactive distributed computing node (UE) * The encrypted data is uploaded to the designated trainer MEC, and then the interactive distributed computing node UE... i Multiple trainer MECs download the global model from the designated central aggregator CSP, and each interactive distributed computing node (UE) downloads the model. i After training the sub-model with each trainer MEC, it is uploaded as a local model to the central aggregator CSP.

[0060] 2) After receiving the local model, the central aggregator (CSP) updates the global model and returns the global model to the interactive distributed computing node (UE). i And each trainer's MEC;

[0061] 3) Repeat steps 1) and 2) above repeatedly, updating and iterating multiple times until the global model reaches the expected accuracy, thus completing the construction.

[0062] Implementation Method 3, see below Figure 1 This embodiment describes a non-interactive federated learning method, which includes the following steps:

[0063] Step 1: Initialization. A central aggregator (CSP) is set up. A homomorphic encryption scheme is selected based on additive homomorphism. A public key and a private key are generated based on the selected homomorphic encryption scheme.

[0064] Step 2: There are m participants (UEs), each interactive node (UE) i We have a dataset, 1 ≤ i ≤ m, which includes feature data X. i and label data Y i Let m be an example. * Each participant corresponds to a non-interactive distributed computing node (UE). * Each We have a dataset, 1≤j≤m * The dataset includes feature data X j and label data Y j All participants' corresponding non-interactive distributed computing nodes (UEs) * Each participant is assigned a trainer MEC, and each participant corresponds to a non-interactive distributed computing node UE. * and interactive node UE i The dataset is encrypted using a public key and then sent to the designated trainer MEC and central aggregator CSP, respectively.

[0065] Step 3: The designated trainer MEC replaces UE. * Non-interactive model training is performed between the central aggregator (CSP) and the participants (UEs) and the CSP. Interactive model training is then conducted between the CSP and the designated trainer (MEC) and each participant (UE). After multiple iterations, the training continues. i Obtain a local model M mec and M i ;

[0066] Step 4: The designated trainer MEC and each participant UE i Local model M mec and M i Update to the central aggregator (CSP); after receiving the local model, the central aggregator updates the global model and returns it to the trainer MEC and each participant UE on the mobile edge server. i This process is repeated until the global model update reaches the expected accuracy.

[0067] Implementation Method 4: This implementation method is a non-interactive federated learning method described in Implementation Method 3. The homomorphic encryption scheme in step 1 simultaneously employs the Paillier encryption scheme, the ECC encryption scheme, and secret sharing technology. Enc is defined to represent encrypting a data record, and Dec is defined to represent decrypting a ciphertext message.

[0068] Implementation Method Five: This implementation method is a non-interactive federated learning method described in Implementation Method Three, where each participant (UE) is described in step 2. * An example of using a public key to encrypt a dataset is as follows:

[0069] Step A1: Randomly select two non-singular integer matrices A j and B j Given a random number R, calculate A j X j Y j RB and B -1 X j ;

[0070] Step A2: Pass B and A through a secure passage where there are no eavesdroppers. j X j and Enc(Y j RB) is sent to MEC, along with A and B. -1 X j Send to the central aggregator (CSP).

[0071] Implementation Method Six: This implementation method is an example illustrating the global model update in step 4 of a non-interactive federated learning method described in Implementation Method Three. Specifically:

[0072] Step B1: The central aggregator (CSP) sends the model parameters RW corresponding to the global model to the interactive distributed computing node (UE). i The model parameters RWA corresponding to the global model j Give MEC to the trainer;

[0073] Step B2: Interactive Distributed Computing Node (UE) i Obtain the global model parameter RW, and then, based on the global model parameter RW, interact with the distributed computing node UE. i For X i and Y i Perform gradient descent algorithm to update the local model and obtain the local model parameters W. i Finally, UE i Set local model parameters W i The corresponding model M i Send to the central aggregator (CSP).

[0074] Step B3: The trainer's MEC downloads the global model and obtains the global model parameters RWA. j Then, based on the global model parameters RWA j , for A j X j 、Enc(Y jRB) performs local model updates to obtain the encrypted local gradient Enc(L) mec Finally, MEC will convert the encrypted ground gradient Enc(L) into a single gradient. mec The corresponding model M) mec Send to the central aggregator (CSP).

[0075] Step B4: The central aggregator CSP, according to M i Get RW i According to M mec Get Enc(L) mec Then CSP according to RW i 、Enc(L mec ) and B -1 X j Perform aggregated average update to obtain global model parameters, thereby obtaining the global model.

[0076] Step B5: Iterate through steps B1-B4 until the model converges to a certain threshold or reaches the maximum number of iterations.

[0077] Implementation Method Seven: This implementation method is a non-interactive federated learning method described in Implementation Method Six, specifically step B2, which involves adjusting X based on the global model parameters RW. i and Y i An example of performing local model updates using the gradient descent algorithm is provided below:

[0078] Through calculation Where α represents the learning rate, which is typically set to 0.001. Let L(x) be the first derivative with respect to W. i ,y i ) = Y i log(h(WX i ))+(1-Y i )log(1-h(WX i h() represents a logical function.

[0079] Implementation Method Eight: This implementation method is a non-interactive federated learning method described in Implementation Method Six, specifically step B3, which involves calculating the global model parameters RWA. j , for A j X j 、Enc(Y j RB) provides an example of a local model, specifically:

[0080] By calculating WX j =R -1 RW1A j -1 A j Xj To eliminate the blindness caused by the random mask R and A, MEC can compute the gradient Enc(L). mec )=Enc(R(h(WX j )-Y j B j =Enc(Rh(WX) j B j Enc(Y j RB j ) -1 .

[0081] Implementation Method Nine: This implementation method is a non-interactive federated learning method described in Implementation Method Six, specifically step B4, which involves using RW... i 、Enc(L mec ) and B -1 X j An example of performing aggregated averaging is as follows:

[0082] First, calculate R(h(WX) j )-Y j B j =Dec(Enc(L) mec Then calculate (L(x)) j ,y j ))=R(h(WX j )-Y j B j B j -1 X j and Final calculation

[0083] Implementation Method 10: This implementation method is an example of the gradient processing described in steps B2 and B3 of the non-interactive federated learning method described in Implementation Method 6. Specifically, the update adopts full gradient descent, mini-batch gradient descent, or stochastic gradient descent.

[0084] Implementation Method Eleven: This implementation method uses a trainer MEC to connect multiple non-interactive distributed computing nodes (UEs). * For example, we will illustrate a method for constructing a safe and efficient non-interactive federated learning framework:

[0085] First, load the client UE that acts as the data provider. * The number of parameters for training a non-interactive federated model is equal. The central aggregator (CSP) selects an encryption scheme based on homomorphic encryption, and the designated trainer (MEC) collects private versions of the user experience (UE). * Training dataset X j ,Y j, j ∈ [1, m * , m * < m; specifically, in the data processing stage, the MEC will perform the following operations, select a random R ∈ Z n , i ∈ [1, m] and its inverse R -1 ∈ Z n , and send R, R -1 to the UE i and secretly store R, R -1 , for the encrypted data sets X j , Y j .

[0086] In the model training stage, local model updates, gradient calculations, and global model updates will be continuously iteratively executed. In the gradient calculation stage, the designated trainer MEC communicates with the central aggregator CSP to update the current gradient using the SGD algorithm. The central aggregator CSP aggregates the local model parameters to obtain the global model and updates it.

[0087] The method of the present invention was compared with existing methods in an experiment. The experimental results show that the present invention can reduce the communication overhead.

[0088] In AuthENIFL and client-edge cloud hierarchical FL[7], the present invention observes the experimental results according to the number of iterations l, the feature dimension d, and the sample size n. In each experiment, l is set to 500, 1000, 1500, 2000, 2500, 3000, d is set to 50, 100, 150, 200, 250, 300, and n is set to 1000, 2000, 3000, 4000, 5000, 6000. The comparison of the communication overhead between AuthENIFL and client-edge cloud hierarchical FL[7] is shown in Figure 5 、 6 and 7. Similarly, the communication overhead of AuthENIFL is also much lower than that of client-edge cloud hierarchical FL[7].

[0089] The computational overhead of each experiment is as Figure 4 shown, based on Figure 4A comparison of the communication overhead of the three curves shows that, compared to NPMML

[21] , VAN

[21] increases significantly with increasing computation time and feature dimension, while NPMML

[21] increases slowly with increasing computation time and feature dimension; AuthENIFL gradually approaches a straight line with increasing computation time and feature dimension. Clearly, AuthENIFL's computational overhead is superior to both NPMML

[21] and VAN

[21] . Therefore, the AuthENIFL framework constructed in this invention is highly efficient.

Claims

1. A method for constructing a secure and efficient non-interactive federated learning framework, characterized in that, The non-interactive federated learning framework includes multiple servers and multiple non-interactive distributed computing nodes (UEs). * and multiple interactive distributed computing nodes (UEs) i Among them, non-interactive distributed computing nodes (UEs) * It features a powerful local computing server and multiple non-interactive distributed computing nodes (UEs). * They jointly designate a server as the trainer MEC; Multiple trainer MECs and interactive distributed computing nodes UE i A server is designated as the central aggregator (CSP), and each trainer MEC is paired with at least one non-interactive distributed computing node (UE). * It is associated with a central aggregator (CSP) that contains multiple participant users (UEs). The method is as follows: Step 1: Each non-interactive distributed computing node (UE) * The encrypted data is uploaded to the designated trainer MEC, and then the interactive distributed computing node UE... i Multiple trainer MECs download the global model from the designated central aggregator CSP, and each interactive distributed computing node (UE) downloads the model. i After training the sub-model with each trainer MEC, it is uploaded as a local model to the central aggregator CSP. Step 2: After receiving the local model, the Central Aggregator (CSP) updates the global model and returns the global model to the Interactive Distributed Computing Node (UE). i And each trainer's MEC; Step 3: Repeat steps 1 and 2 above repeatedly, updating and iterating multiple times until the global model reaches the expected accuracy, thus completing the construction.

2. A non-interactive federated learning method, characterized in that, The learning method is implemented based on the construction method described in claim 1, and the method includes the following steps: Step 1: Initialization. A central aggregator (CSP) is set up. A homomorphic encryption scheme is selected based on subtractive homomorphism. A public key and a private key are generated based on the selected homomorphic encryption scheme. Step 2: There are m participants (UEs), each interactive node (UE) i Having a dataset The dataset includes feature data X i and label data Y i Let m be an example. * Each participant corresponds to a non-interactive distributed computing node (UE). * each Having a dataset The dataset includes feature data X j and label data Y j All participants' corresponding non-interactive distributed computing nodes (UEs) * Each participant is assigned a trainer MEC, and each participant corresponds to a non-interactive distributed computing node UE. * and interactive node UE i The dataset is encrypted using a public key and then sent to the designated trainer MEC and central aggregator CSP, respectively. Step 3: The designated trainer MEC replaces UE. * Non-interactive model training is performed between the central aggregator (CSP) and the participants (UEs) and the CSP. Interactive model training is then conducted between the CSP and the designated trainer (MEC) and each participant (UE). After multiple iterations, the training continues. i Obtain a local model M mec and M i ; Step 4: The designated trainer MEC and each participant UE i Local model M mec and M i Updated to the central aggregator (CSP); after receiving the local model, the central aggregator updates the global model and returns it to the trainer MEC and each participant UE on the mobile edge server. i This process is repeated until the global model update reaches the expected accuracy.

3. The non-interactive federated learning method according to claim 2, characterized in that, The encryption scheme described in step 1 is as follows: The homomorphic encryption scheme employs Paillier encryption, ECC encryption, and secret sharing technology. Enc is defined to represent encrypting a data record, and Dec is defined to represent decrypting a ciphertext message.

4. A non-interactive federated learning method according to claim 2, characterized in that, Each participant UE described in step 2 * The dataset is encrypted using a public key as follows: Step A1: Randomly select two non-singular integer matrices A j and B j Given a random number R, calculate A j X j Y j RB and B - 1 X j ; Step A2: Pass B and A through a secure passage where there are no eavesdroppers. j X j and Enc(Y j RB) is sent to MEC, along with A and B. -1 X j Send to the central aggregator (CSP).

5. A non-interactive federated learning method according to claim 2, characterized in that, The global model update described in step 4 specifically refers to: Step B1: The central aggregator (CSP) sends the model parameters RW corresponding to the global model to the interactive distributed computing node (UE). i The model parameters RWA corresponding to the global model j Give MEC to the trainer; Step B2: Interactive Distributed Computing Node (UE) i Obtain the global model parameter RW, and then, based on the global model parameter RW, interact with the distributed computing node UE. i For X i and Y i Perform gradient descent algorithm to update the local model and obtain the local model parameters W. i Finally, UE i Set local model parameters W i The corresponding model M i Send to the central aggregator CSP; Step B3: The trainer's MEC downloads the global model and obtains the global model parameters RWA. j Then, based on the global model parameters RWA j , for A j X j 、Enc(Y j RB) performs local model updates to obtain the encrypted local gradient Enc(L) mec Finally, MEC will convert the encrypted ground gradient Enc(L) into a single gradient. mec The corresponding model M) mec Send to the central aggregator CSP; Step B4: The central aggregator CSP, according to M i Get RW i According to M mec Get Enc(L) mec Then CSP according to RW i 、Enc(L mec ) and B -1 X j Perform aggregated average update to obtain global model parameters, thereby obtaining the global model; Step B5: Iterate through steps B1-B4 until the model converges to a certain threshold or reaches the maximum number of iterations.

6. A non-interactive federated learning method according to claim 5, characterized in that, Step B2 describes the interactive distributed computing node UE based on the global model parameters RW. i For X i and Y i The process of performing local model updates using the gradient descent algorithm is as follows: Through calculation ,in This represents the learning rate, which is typically set to 0.

001. This represents the first derivative with respect to W. , Represents a logical function. = .

7. A non-interactive federated learning method according to claim 5, characterized in that, Step B3 describes the process of using global model parameters RWA. j , for A j X j 、Enc(Y j The specific steps for local model updates (RB) are as follows: Through calculation To eliminate the blindness caused by random masks R and A, MEC can compute gradients. .

8. A non-interactive federated learning method according to claim 5, characterized in that, The central aggregator CSP mentioned in step B4 is based on RW i 、Enc(L mec ) and B -1 X j The specific steps for performing aggregated average updates are as follows: First calculate , Then calculate and , Final calculation .

9. A non-interactive federated learning method according to claim 5, characterized in that, The gradient calculation in steps B2 and B3 specifically involves updating using full gradient descent, mini-batch gradient descent, or stochastic gradient descent.

Images