Split federal learning method and device, equipment and medium

Abstract

The invention provides a split federal learning method and device, equipment and a medium. The method comprises the following steps: receiving a second sub-model and activation data; training a second sub-model according to the activation data to obtain a second sub-model parameter gradient and a gradient of the activation data, feeding back the second sub-model parameter gradient to the server, and feeding back the gradient of the activation data to the client, so that the client carries out back propagation on the first sub-model according to the gradient of the activation data to obtain a first sub-model parameter gradient, feeding back the first sub-model parameter gradient to the server; and the server updates the global model parameters according to the first sub-model parameter gradient and the second sub-model parameter gradient. According to the invention, the risk of insufficient computing resources of the client and the server is reduced, and the training effect of the global model is ensured.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a split federated learning method, apparatus, device, and medium. Background Technology

[0002] Federated learning achieves its privacy goals by maintaining the local processing of data among participating parties. In this model, all clients involved in training can collaboratively build a global model without disclosing the original data. This architecture places high demands on the computing power of the clients. While split federated learning, by shifting the training tasks from the client to the server (allowing the client to train only a portion of the model while the other portion is trained on the server), can address the issue of limited client resources, this also increases the server load. Moreover, in real-world applications, situations may arise where both the client and server experience insufficient computing resources simultaneously. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method, apparatus, device, and medium for split federated learning.

[0004] This invention provides a split federated learning method applied to model training hosting nodes, comprising: The system receives a second sub-model and activation data. The second sub-model and the first sub-model received by the client together constitute a global model trained by the server. The activation data is the data output by the client after training the first sub-model. The second sub-model is trained based on the activation data to obtain the gradients of the second sub-model parameters and the activation data. The gradients of the second sub-model parameters are fed back to the server, and the gradients of the activation data are fed back to the client, so that the client backpropagates the first sub-model based on the gradients of the activation data to obtain the gradients of the first sub-model parameters, and feeds back the gradients of the first sub-model parameters to the server. The server updates the global model parameters based on the gradients of the first and second sub-model parameters.

[0005] According to a split federated learning method provided by the present invention, the method further includes: upon receiving the second sub-model, receiving preset first verification information, and interacting with the client based on the first verification information to complete a trusted verification operation of the second sub-model; wherein the client records the trusted verification time of the second sub-model.

[0006] According to a split federated learning method provided by the present invention, the method further includes: after receiving the activation data, receiving preset second verification information, and interacting with the server according to the second verification information to complete the trusted verification operation of the activation data; wherein, the server records the trusted verification time of the activation data.

[0007] According to a split federated learning method provided by the present invention, the server is configured to: record the time of sending the first sub-model and the second sub-model; and record the first time of receiving the gradient of the parameters of the second sub-model. The client is configured to: record the sending time of sending activation data to the model training hosting node; and record the second receiving time of the gradient of the received activation data. Accordingly, the server or the client is configured to: calculate the time consumption information of the model training hosting node in the preparation and computation phases of training the model based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time.

[0008] According to a split federated learning method provided by the present invention, the method further includes: feeding back the forward propagation result obtained from training the second sub-model to the client, so that the client calculates the loss value based on the forward propagation result and sends the loss value to the model training hosting node; Backpropagation is performed based on the loss value to obtain the data preparation time of the model training hosting node; The client is configured to: record the time of receiving the forward propagation result and the time of sending the loss value to the model training hosting node; and determine the forward propagation time and backward propagation time of the model training hosting node in the process of training the second sub-model based on the recorded time of receiving the forward propagation result, the time of sending the loss value to the model training hosting node, the start and end times of the second sub-model training calculation, and the theoretical network time between the client and the model training hosting node.

[0009] This invention also provides a split federated learning method applied to a server, comprising: The global model is split into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. The first sub-model is sent to the client so that the client backpropagates the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feeds the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model. The second sub-model is sent to the model training hosting node so that the model training hosting node trains the second sub-model based on the activation data, obtains the gradient of the second sub-model parameters and the gradient of the activation data, feeds back the gradient of the second sub-model parameters to the server, and feeds back the gradient of the activation data to the client. Update the global model parameters based on the gradients of the first and second sub-model parameters.

[0010] According to a split federated learning method provided by the present invention, the method further includes: The preset second verification information sent by the model training hosting node is determined, the trusted verification operation of the activation data by the model training hosting node is completed, and the trusted verification time of the activation data is recorded.

[0011] According to a split federated learning method provided by the present invention, the method further includes: receiving preset first verification information sent by the client to determine the model training hosting node, completing the trusted verification operation of the model training hosting node on the second sub-model, and recording the trusted verification time of the second sub-model. The server is configured to: record the time of sending the first sub-model and the second sub-model; and record the first time of receiving the gradient of the second sub-model parameters. The client is configured to: record the sending time of sending activation data to the model training hosting node; and record the second receiving time of the gradient of the received activation data. Accordingly, based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time, the time consumption information of the model training hosting node in the training preparation stage and the training computation stage of the training model is calculated.

[0012] According to a split federated learning method provided by the present invention, the training phase includes a training preparation phase, a model preparation phase, and a data preparation phase. The trusted verification time for the second sub-model includes the time when the client receives the legality verification request for the second sub-model and the time when the client returns the verification result to the model training hosting node. The trusted verification time for the activation data includes the time when the server receives the legality verification request for the activation data and the time when the server returns the verification result to the model training hosting node. Accordingly, the step of calculating the time consumption information of the model training hosting node in the preparation and computation phases of training the model based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time includes: The training preparation phase time of the model training hosting node is determined based on the distribution time, the sending time, the theoretical network time between the client and the model training hosting node, the theoretical network time between the server and the model training hosting node, and the start time of the second sub-model training calculation. The model preparation phase time of the model training hosting node is determined based on the time when the client receives the second sub-model validity verification request, the theoretical network latency between the client and the model training hosting node, and the server's delivery time of the second sub-model. The data preparation time of the model training hosting node is determined based on the time when the server receives the activation data validity verification request, the theoretical network latency between the server and the model training hosting node, and the time when the activation data is sent to the model training hosting node. The start time of the second sub-model training computation is jointly determined by the time when the client returns the verification result to the model training hosting node, the time when the server returns the verification result to the model training hosting node, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node.

[0013] According to a split federated learning method provided by the present invention, the step of calculating the time consumption information of the model training hosting node in the preparation and computation phases of training the model based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time includes: The start time of the second sub-model training computation is determined based on the time it takes for the client to return the verification result to the model training hosting node, the time it takes for the server to return the verification result to the model training hosting node, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node. The end time of the second sub-model training computation is determined based on the first receiving time, the second receiving time, the theoretical network time between the client and the model training hosting node, and the theoretical network time between the server and the model training hosting node. The data computation time of the model training hosting node is determined based on the start time and end time of the second sub-model training computation.

[0014] According to a split federated learning method provided by the present invention, the method further includes: The trustworthiness of each hosting node is determined based on its historical latency, resource usage, and network communication quality indicators. Based on the trustworthiness of each hosting node, determine the required hosting nodes for model training.

[0015] This invention also provides a split federated learning method, applied to a client, comprising: Receive the gradient of the first sub-model sent by the server and the activation data sent by the model training hosting node; Backpropagation is performed on the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and the gradient of the first sub-model parameters is fed back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data. The server is configured to update global model parameters based on the gradients of the first sub-model parameters and the second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0016] The present invention also provides a split federated learning apparatus, comprising: The first receiving module is used to receive the second sub-model and activation data. The second sub-model and the first sub-model received by the client together constitute the global model trained by the server. The activation data is the data output by the client after training the first sub-model. The second processing module is used to train the second sub-model based on the activation data, obtain the gradient of the second sub-model parameters and the gradient of the activation data, feed back the gradient of the second sub-model parameters to the server, and feed back the gradient of the activation data to the client, so that the client backpropagates the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feeds back the gradient of the first sub-model parameters to the server; wherein, the server updates the global model parameters based on the gradient of the first sub-model parameters and the gradient of the second sub-model parameters.

[0017] The present invention also provides a split federated learning apparatus, comprising: The splitting module is used to split the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. The first distribution module is used to distribute the first sub-model to the client, so that the client backpropagates the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feeds back the gradient of the first sub-model parameters to the server; the activation data is the data output by the client after training the first sub-model. The second distribution module is used to distribute the second sub-model to the model training hosting node, so that the model training hosting node trains the second sub-model according to the activation data, obtains the gradient of the second sub-model parameters and the gradient of the activation data, feeds back the gradient of the second sub-model parameters to the server, and feeds back the gradient of the activation data to the client. The update module is used to update the global model parameters based on the gradients of the first sub-model parameters and the second sub-model parameters.

[0018] The present invention also provides a split federated learning apparatus, comprising: The second receiving module is used to receive the gradient of the first sub-model and the activation data sent by the model training hosting node from the server. The second processing module is used to backpropagate the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the process of training the second sub-model according to the activation data. The server is configured to update global model parameters based on the gradients of the first sub-model parameters and the second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the split federated learning methods described above.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the split federated learning methods described above.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the split federated learning methods described above.

[0022] This invention provides a method, apparatus, device, and medium for split federated learning. By introducing a model training hosting node, the model training hosting node trains a sub-model based on activation data, obtaining the gradients of the sub-model parameters and the activation data. The sub-model parameter gradients are fed back to the server, and the gradients of the activation data are fed back to the client. This allows the client to backpropagate the sub-model based on the gradients of the activation data to obtain the sub-model parameter gradients, which are then fed back to the server. The server updates the global model parameters based on the sub-model parameter gradients, reducing the risk of insufficient computing resources on both the client and server, and ensuring the training effect of the global model. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the split federated learning method provided by the present invention. Figure 1 .

[0025] Figure 2 This is a flowchart illustrating the split federated learning method provided by the present invention. Figure 2 .

[0026] Figure 3 This is a flowchart illustrating the split federated learning method provided by the present invention. Figure 3 .

[0027] Figure 4 This is a schematic diagram of the structure of the split federated learning device provided by the present invention. Figure 1 .

[0028] Figure 5 This is a schematic diagram of the structure of the split federated learning device provided by the present invention. Figure 2 .

[0029] Figure 6 This is a schematic diagram of the structure of the split federated learning device provided by the present invention. Figure 3 .

[0030] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] Federated learning achieves its privacy goals by maintaining the local processing of data among participating parties. In this model, all clients participating in training can collaboratively build a global model without disclosing the raw data. This architecture places high demands on the computing power of the clients. While splitting federated learning to transfer training tasks from the client to the server (allowing the client to train only a portion of the model while the other portion is trained on the server) can address the issue of limited client resources, it also increases the server load. Moreover, in real-world communication business data processing applications, there may be situations where both the client and server experience insufficient computing resources simultaneously. This is even more true in application environments with high demands for business data, such as the Internet of Things, smart manufacturing, and battlefield awareness.

[0033] Therefore, the present invention proposes a split federated learning method, the main idea of ​​which is to introduce a data custodian, that is, to set up a model training hosting node. This model training hosting node is a terminal that provides computing services, thus meeting the requirements of federated learning in terms of computing power, computing environment, and technical support.

[0034] This invention provides a split federated learning method, see [link to relevant documentation]. Figure 1 The method includes the following steps: Step 11: The model training hosting node receives the second sub-model and activation data. The second sub-model and the first sub-model received by the client together constitute the global model trained by the server; the activation data is the data output by the client after training the first sub-model.

[0035] Step 12: The model training hosting node trains the second sub-model based on the activation data, obtains the gradient of the second sub-model parameters and the gradient of the activation data, feeds back the gradient of the second sub-model parameters to the server, and feeds back the gradient of the activation data to the client, so that the client backpropagates the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feeds back the gradient of the first sub-model parameters to the server; wherein, the server updates the global model parameters based on the gradient of the first sub-model parameters and the gradient of the second sub-model parameters.

[0036] Regarding steps 11 and 12, it's important to note that, firstly, in the decomposed federated learning, a model training hosting node is introduced. This node acts as a computing service provider, offering computing power, runtime environment, and other computing services. It completes part of the training tasks on behalf of the data provider (client) and model provider (server), collaborating with the client and server to train the federated learning model. Simultaneously, for cases involving a model training hosting node, a comprehensive security mechanism for the decomposed federated learning process is proposed. This ensures the security of the federated learning process while collaboratively completing it, even in situations where there is a lack of mutual trust among the parties.

[0037] In this invention, before training begins, the server and / or client can assess the trustworthiness of the model training hosting nodes and select highly trustworthy nodes to participate in federated learning. During the training process, the hosting nodes verify the legitimacy of the model and data with both the client and the server through an authentication mechanism. After training, the training results of the verification samples on the client and server are compared with the training results on the hosting nodes to check whether the hosting nodes engaged in fraudulent or malicious behavior during the model training process.

[0038] In this invention, the server is the model provider, the client is the data provider, and the model training hosting node is the computing service provider.

[0039] The data provider has the data, but limited computing resources.

[0040] The model provider has the model, but has limited computing resources.

[0041] The computing service provider has sufficient computing resources, including sufficient computing power, the operating environment required for training, and the required level of technical support, and only runs one model training hosting task at a time.

[0042] In this invention, one or more clients can participate in the federated learning scenario for model training, or one client and multiple model training hosting nodes can participate in the federated learning scenario for model training. Therefore, the client mentioned in the above steps of this invention can be one or more clients, and the model training hosting node can be one or more nodes. Accordingly, the server splits the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. Therefore, the server distributes the first sub-model to the client and the second sub-model to the model training hosting node.

[0043] It should be noted that after the server and client authenticate each other, they can negotiate and determine: select a hosting node for model training based on trustworthiness, continue to determine the model segmentation layer (i.e., the number of sub-models) based on the performance of the client and the hosting node for model training, and the client can randomly select a small number of sample data from the training data as verification samples and authentication mechanisms (such as encrypted tokens or digital signatures, used for legality verification, and thus can also be used as verification information).

[0044] Furthermore, the client performs anonymization processing on the agreed-upon verification sample data and sends the processed verification sample data to the server.

[0045] The client trains the first sub-model, generating shredded data (i.e., the output of the first sub-model training, used to activate the second sub-model training, and thus also as activation data). The client also sends an encrypted token or digital signature to the model training hosting node. Upon receiving the second sub-model, the hosting node also receives pre-defined first verification information, used to verify the legitimacy of the second sub-model. At this point, the hosting node and the client complete the trusted verification operation of the second sub-model using the encrypted token or digital signature. Simultaneously, the client records the trusted verification time of the second sub-model, including the time the client receives the legitimacy verification request for the second sub-model and the time the client returns the verification result to the hosting node.

[0046] Furthermore, after receiving the activation data, the model training hosting node verifies the trustworthiness of the activation data source with the server using an encrypted token or digital signature. Simultaneously, the server records the time of the trustworthy verification of the activation data. The trustworthy verification time includes the time the server receives the activation data validity verification request and the time the server returns the verification result to the model training hosting node.

[0047] In this invention, the model training hosting node trains a second sub-model based on the activation data, obtains the gradient of the second sub-model parameters and the gradient of the activation data, feeds back the gradient of the second sub-model parameters to the server, and feeds back the gradient of the activation data to the client, so that the client backpropagates the first sub-model based on the gradient of the activation data, obtains the gradient of the first sub-model parameters, and feeds back the gradient of the first sub-model parameters to the server.

[0048] In this invention, the server updates the global model parameters based on the gradients of the first and second sub-model parameters. This is predicated on comparing the training results of the sub-models on the server with the training results returned by the model training hosting node and the client. The reliability of the training results is determined by comparing these comparisons, thereby deciding whether to update the global model parameters based on the sub-model parameter gradients.

[0049] In response, the client sends the anonymized validation sample data to the server. The server uses this validation sample data to perform forward and backward propagation on the first sub-model to obtain the training results. The client also sends a small amount of fragmented validation sample data to the server, which uses this fragmented validation sample data to perform forward propagation on the second sub-model to obtain the training results.

[0050] It should also be noted that each time a global model training task is set up, the client and the model training hosting node need to be negotiated and determined. Selecting a suitable model training hosting node relies on the node's trustworthiness, and trustworthiness can be incorporated into the selection process.

[0051] In this invention, the server and client can directly share information, and the server or client can select a suitable model training hosting node.

[0052] In this invention, the server or client determines the credibility of each hosted node based on its historical time consumption information, historical resource usage, and historical network communication quality index information. Based on the credibility of each hosted node, the required hosted node for model training is determined.

[0053] In this invention, in order to better adapt to the computing power of different managed nodes and to select the model for training managed nodes more quickly and reasonably, this invention can use a model to train a trusted selection model for managed nodes as an evaluation method. This model is trained by a neural network model using the historical time consumption information, historical resource usage and historical network communication quality index information of each managed node as training samples, thereby evaluating the trustedness of the managed nodes.

[0054] Based on trustworthiness, a model training hosting node is selected as a computing service provider to participate in the federated learning model training. The model training hosting node provides computing power and a runtime environment, completing part of the federated learning model training tasks on behalf of the client and server.

[0055] In this invention, network communication quality index data includes latency, communication time, packet loss rate, etc. The client and / or server collect network quality index data with the model training hosting node through active dialing.

[0056] The time consumption information of the managed node includes the time consumed in each stage (data preparation stage, model preparation stage, training stage, etc.), the communication time between the managed node and the server, and the communication time between the managed node and the client.

[0057] Furthermore, in order to obtain the time consumption information of the hosting node, the server determines the preset second verification information sent by the model training hosting node, completes the trusted verification operation of the model training hosting node on the activation data, and records the trusted verification time of the activation data.

[0058] The server is configured to record the distribution time T of the first and second sub-models. send2 Record the first reception time T when the gradient of the second sub-model parameters is received. rec2 .

[0059] The client is configured to record the time T when sending activation data to the model training hosting node. send1 Record the second reception time T of the gradient received from the activation data. rec1 .

[0060] The trusted verification time for the second sub-model includes the time T when the client receives the request for verification of the legitimacy of the second sub-model. req1 The time T for the client to return the validation results to the model training hosting node reply1 The time for verifying the authenticity of the activation data includes the time T when the server receives the request to verify the validity of the activation data. req2 The time T for the server to return the validation results to the model training hosting node reply2 .

[0061] Based on the distribution time, sending time, theoretical network latency t1 between the client and the model training hosting node, theoretical network latency t2 between the server and the model training hosting node, and the start time T of the second sub-model training computation. start Determine the time consumed in the training preparation phase of the model training hosting node.

[0062] Training preparation phase time: t pre =T start -Min(T) send1 + t1, T send2 +t2).

[0063] The model preparation phase time of the model training hosting node is determined based on the time when the client receives the second sub-model validity verification request, the theoretical network latency between the client and the model training hosting node, and the server's delivery time of the second sub-model.

[0064] Model preparation phase time: t pre-model =T reply1 + t1-T send2 - (T) reply1 -T req )=T reg1 + t1-T send2 .

[0065] The data preparation time of the model training hosting node is determined based on the time when the server receives the activation data validity verification request, the theoretical network latency between the server and the model training hosting node, and the time when the activation data is sent to the model training hosting node.

[0066] Data preparation phase time: t pre-data =T reply2 +t2-T send1 - (T) reply2 -T req2 )=T req2 + t2-T send1 .

[0067] The start time of the second sub-model training computation is jointly determined by the time when the client returns the verification result to the model training hosting node, the time when the server returns the verification result to the model training hosting node, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node.

[0068] Start time of training computation for the second sub-model: T srart = Max(T) reply1 + t1, T reply2 + t2).

[0069] The end time of the second sub-model training computation is determined based on the first reception time, the second reception time, the theoretical network time between the client and the model training hosting node, and the theoretical network time between the server and the model training hosting node.

[0070] End time of training computation for the second sub-model: T end =Min(T) rec2 -t2,T rec1 -t1).

[0071] The data computation time of the model training hosting node is determined based on the start time and end time of the second sub-model training computation.

[0072] Data computation time: T train =T end -T srart .

[0073] Furthermore, the model training hosting node feeds back the forward propagation results obtained from training the second sub-model to the client, so that the client can calculate the loss value based on the forward propagation results and send the loss value to the model training hosting node; the model training hosting node performs back propagation based on the loss value to obtain the data preparation time of the model training hosting node; The client is configured to record the time T at which the forward propagation result is received. rec-result And the time T for sending the loss value to the model training hosting node. reply-loss Based on the time of receiving the forward propagation result, the time of sending the loss value to the model training hosting node, the start and end times of the second sub-model training calculation, and the theoretical network latency between the client and the model training hosting node, the forward propagation latency and backward propagation latency of the model training hosting node during the training of the second sub-model are determined.

[0074] The time consumed by forward and backward propagation on the managed node is calculated as follows: Forward propagation time: t propagation =T rec-result -T start -t1.

[0075] Backpropagation time: t backpropapgtion =T end ―T reply-loss ―t1.

[0076] Resource usage includes CPU utilization, memory utilization, GPU utilization, network traffic, computing power, network resource allocation, and even whether the hosted nodes are engaging in fraud or attacks.

[0077] On the client and / or server, after the trustworthiness of the model outputs a new trustworthiness value for the managed nodes, the trustworthiness values ​​of each managed node are periodically updated and maintained. During federated learning, the computing service provider with the highest trustworthiness value is selected as the managed node for model training.

[0078] The split federated learning method provided by this invention introduces a model training hosting node. The hosting node trains a sub-model based on the activation data, obtains the gradients of the sub-model parameters and the gradients of the activation data, and feeds back the sub-model parameter gradients to the server and the gradients of the activation data to the client. This allows the client to backpropagate the sub-model based on the gradients of the activation data to obtain the sub-model parameter gradients, which are then fed back to the server. The server updates the global model parameters based on the sub-model parameter gradients, reducing the risk of insufficient computing resources for both the client and the server and ensuring the training effect of the global model.

[0079] In addition, for cases where model training hosting nodes are involved, security mechanisms are integrated into the entire process of splitting federated learning models, which protects data privacy and model security while reducing the investment of computing power, hardware and other resources in security protection.

[0080] When selecting hosting nodes for model training to participate in federated learning, security and trustworthiness factors should be taken into account in addition to performance to reduce the security risks of federated learning.

[0081] Figure 2 This diagram illustrates a flowchart of a split federated learning method provided by the present invention. Figure 2 The method includes the following steps: Step 21: The server splits the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node.

[0082] Step 22: The server sends the first sub-model to the client so that the client can backpropagate the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model.

[0083] Step 23: The server distributes the second sub-model to the model training hosting node, so that the model training hosting node trains the second sub-model based on the activation data, obtains the gradient of the second sub-model parameters and the gradient of the activation data, feeds back the gradient of the second sub-model parameters to the server, and feeds back the gradient of the activation data to the client.

[0084] Step 24: The server updates the global model parameters based on the gradients of the first sub-model parameters and the second sub-model parameters.

[0085] Furthermore, the server determines the preset second verification information sent by the model training hosting node, completes the trusted verification operation of the activation data by the model training hosting node, and records the trusted verification time of the activation data.

[0086] The method further includes: receiving preset first verification information sent by the client to determine the model training hosting node, completing the trusted verification operation of the model training hosting node on the second sub-model, and recording the trusted verification time of the second sub-model. The server is configured to: record the time when the first and second sub-models are sent; and record the first time the gradient of the second sub-model parameters is received. The client is configured to: record the time when activation data is sent to the model training hosting node; and record the second time when the gradient of the received activation data is received. Accordingly, based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time, the time consumption information of the model training hosting node in the training preparation stage and the training computation stage of the training model is calculated.

[0087] Furthermore, the training phase includes the training preparation phase, the model preparation phase, and the data preparation phase. The trusted verification time for the second sub-model includes the time when the client receives the legality verification request for the second sub-model and the time when the client returns the verification result to the model training hosting node. The trusted verification time for the activation data includes the time when the server receives the legality verification request for the activation data and the time when the server returns the verification result to the model training hosting node. Accordingly, based on the trusted verification time of the second sub-model, the trusted verification time of the activation data, the distribution time, the sending time, the first receiving time, and the second receiving time, the time consumption information of the model training hosting node in the preparation and computation phases of training the model is calculated, including: The training preparation phase time of the model training hosting node is determined based on the distribution time, the theoretical network time between the client and the model training hosting node, the theoretical network time between the server and the model training hosting node, and the start time of the second sub-model training calculation. The model preparation phase time of the model training hosting node is determined based on the time when the client receives the second sub-model validity verification request, the theoretical network latency between the client and the model training hosting node, and the server's delivery time of the second sub-model. The data preparation time of the model training hosting node is determined based on the time when the server receives the activation data validity verification request, the theoretical network latency between the server and the model training hosting node, and the time when the server distributes the first sub-model. The start time of the second sub-model training computation is jointly determined by the time when the client returns the verification result to the model training hosting node, the time when the server returns the verification result to the model training hosting node, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node.

[0088] The start time of the second sub-model training computation is determined based on the time it takes for the client to return the verification result to the model training hosting node, the time it takes for the server to return the verification result to the model training hosting node, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node. The end time of the second sub-model training computation is determined based on the first reception time, the second reception time, the theoretical network latency between the client and the model training hosting node, and the theoretical network latency between the server and the model training hosting node. The data computation time of the model training hosting node is determined based on the start time and end time of the second sub-model training computation.

[0089] The method also includes: The trustworthiness of each hosting node is determined based on its historical latency, resource usage, and network communication quality indicators. Based on the trustworthiness of each hosting node, determine the required hosting nodes for model training.

[0090] Since the method in this embodiment of the invention is based on the same principle as the method in the above embodiments, more detailed explanations will not be repeated here.

[0091] The split federated learning method provided by this invention introduces a model training hosting node. The hosting node trains a sub-model based on the activation data, obtains the gradients of the sub-model parameters and the gradients of the activation data, and feeds back the sub-model parameter gradients to the server and the gradients of the activation data to the client. This allows the client to backpropagate the sub-model based on the gradients of the activation data to obtain the sub-model parameter gradients, which are then fed back to the server. The server updates the global model parameters based on the sub-model parameter gradients, reducing the risk of insufficient computing resources for both the client and the server and ensuring the training effect of the global model.

[0092] In addition, for cases where model training hosting nodes are involved, security mechanisms are integrated into the entire process of splitting federated learning models, which protects data privacy and model security while reducing the investment of computing power, hardware and other resources in security protection.

[0093] When selecting hosting nodes for model training to participate in federated learning, security and trustworthiness factors should be taken into account in addition to performance to reduce the security risks of federated learning.

[0094] Figure 3 This diagram illustrates the flowchart of the split federated learning method provided by the present invention. Figure 3 The method includes: Step 31: The client receives the gradient of the first sub-model and the activation data sent by the model training hosting node from the server.

[0095] Step 32: Backpropagate the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data; The server is configured to update the global model parameters based on the gradients of the first and second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0096] Since the method in this embodiment of the invention is based on the same principle as the method in the above embodiments, more detailed explanations will not be repeated here.

[0097] The split federated learning method provided by this invention introduces a model training hosting node. The hosting node trains a sub-model based on the activation data, obtains the gradients of the sub-model parameters and the gradients of the activation data, and feeds back the sub-model parameter gradients to the server and the gradients of the activation data to the client. This allows the client to backpropagate the sub-model based on the gradients of the activation data to obtain the sub-model parameter gradients, which are then fed back to the server. The server updates the global model parameters based on the sub-model parameter gradients, reducing the risk of insufficient computing resources for both the client and the server and ensuring the training effect of the global model.

[0098] In addition, for cases where model training hosting nodes are involved, security mechanisms are integrated into the entire process of splitting federated learning models, which protects data privacy and model security while reducing the investment of computing power, hardware and other resources in security protection.

[0099] When selecting hosting nodes for model training to participate in federated learning, security and trustworthiness factors should be taken into account in addition to performance to reduce the security risks of federated learning.

[0100] The split federated learning apparatus provided by the present invention is described below. The split federated learning apparatus described below can be referred to in correspondence with the split federated learning method described above.

[0101] Figure 4 A schematic diagram of a split federated learning device provided by the present invention is shown below. Figure 4 The device includes a first receiving module 41 and a second processing module 42, wherein: The first receiving module is used to receive the second sub-model and activation data. The second sub-model and the first sub-model received by the client together constitute the global model trained by the server. The activation data is the data output by the client after training the first sub-model. The second processing module is used to train a second sub-model based on the activation data, obtain the gradient of the second sub-model parameters and the gradient of the activation data, feed back the gradient of the second sub-model parameters to the server, and feed back the gradient of the activation data to the client, so that the client backpropagates the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and feeds back the gradient of the first sub-model parameters to the server; wherein, the server updates the global model parameters based on the gradient of the first sub-model parameters and the gradient of the second sub-model parameters.

[0102] Figure 5 A schematic diagram of a split federated learning device provided by the present invention is shown below. Figure 5The device includes a splitting module 51, a first sending module 52, a second sending module 53, and an updating module 54, wherein: The splitting module is used to split the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. The first delivery module is used to deliver the first sub-model to the client, so that the client can backpropagate the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and then feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model; The second distribution module is used to distribute the second sub-model to the model training hosting node, so that the model training hosting node can train the second sub-model based on the activation data, obtain the gradient of the second sub-model parameters and the gradient of the activation data, feed back the gradient of the second sub-model parameters to the server, and feed back the gradient of the activation data to the client. The update module is used to update the global model parameters based on the gradients of the first and second sub-model parameters.

[0103] Figure 6 A schematic diagram of a split federated learning device provided by the present invention is shown below. Figure 6 The device includes a second receiving module 61 and a second processing module 62, wherein: The second receiving module is used to receive the gradient of the first sub-model and the activation data sent by the model training hosting node from the server. The second processing module is used to backpropagate the first sub-model based on the gradient of the activation data to obtain the gradient of the first sub-model parameters, and then feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data. The server is configured to update the global model parameters based on the gradients of the first and second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0104] Since the method in this embodiment of the invention is based on the same principle as the method in the above embodiments, more detailed explanations will not be repeated here.

[0105] It should be noted that, in the embodiments of the present invention, the relevant functional modules can be implemented by a hardware processor.

[0106] The split federated learning device provided by this invention introduces a model training hosting node. The model training hosting node trains a sub-model based on the activation data, obtains the gradient of the sub-model parameters and the gradient of the activation data, feeds back the gradient of the sub-model parameters to the server, and feeds back the gradient of the activation data to the client. The client then backpropagates the sub-model based on the gradient of the activation data to obtain the gradient of the sub-model parameters, feeds back the gradient of the sub-model parameters to the server, and the server updates the global model parameters based on the gradient of the sub-model parameters. This reduces the risk of insufficient computing resources for both the client and the server and ensures the training effect of the global model.

[0107] In addition, for cases where model training hosting nodes are involved, security mechanisms are integrated into the entire process of splitting federated learning models, which protects data privacy and model security while reducing the investment of computing power, hardware and other resources in security protection.

[0108] When selecting hosting nodes for model training to participate in federated learning, security and trustworthiness factors should be taken into account in addition to performance to reduce the security risks of federated learning.

[0109] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 71, a communication interface 72, a memory 73, and a communication bus 74. The processor 71, communication interface 72, and memory 73 communicate with each other via the communication bus 74. The processor 71 can call logical instructions in the memory 73 to execute a split federated learning method. This method includes: receiving a second sub-model and activation data, wherein the second sub-model and the first sub-model received by the client together constitute a global model trained by the server; the activation data is the data output by the client after training the first sub-model. The second sub-model is trained based on the activation data, and the gradients of the second sub-model parameters and the activation data are obtained. The gradients of the second sub-model parameters are fed back to the server, and the gradients of the activation data are fed back to the client, so that the client can backpropagate the first sub-model based on the gradients of the activation data to obtain the gradients of the first sub-model parameters, and feed the gradients of the first sub-model parameters back to the server. The server updates the global model parameters based on the gradients of the first and second sub-model parameters.

[0110] Alternatively: split the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. The first sub-model is sent to the client so that the client can backpropagate the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and then feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model. The second sub-model is sent to the model training hosting node so that the model training hosting node can train the second sub-model based on the activation data, obtain the gradient of the second sub-model parameters and the gradient of the activation data, and feed back the gradient of the second sub-model parameters to the server and the gradient of the activation data to the client. Update the global model parameters based on the gradients of the first and second sub-model parameters.

[0111] Or: Receive the gradient of the first sub-model sent by the server and the activation data sent by the model training hosting node; The gradient of the activation data is backpropagated to the first sub-model to obtain the gradient of the first sub-model parameters, and the gradient of the first sub-model parameters is fed back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data; The server is configured to update the global model parameters based on the gradients of the first and second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0112] Furthermore, the logical instructions in the aforementioned memory 73 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute a split federated learning method, the method including: receiving a second sub-model and activation data, wherein the second sub-model and the first sub-model received by the client together constitute a global model trained by the server; the activation data is the data output by the client after training the first sub-model; The second sub-model is trained based on the activation data, and the gradients of the second sub-model parameters and the activation data are obtained. The gradients of the second sub-model parameters are fed back to the server, and the gradients of the activation data are fed back to the client, so that the client can backpropagate the first sub-model based on the gradients of the activation data to obtain the gradients of the first sub-model parameters, and feed the gradients of the first sub-model parameters back to the server. The server updates the global model parameters based on the gradients of the first and second sub-model parameters.

[0114] Alternatively: split the global model into a first sub-model and a second sub-model. The first sub-model is trained by the client, and the second sub-model is trained by the model training hosting node. The first sub-model is sent to the client so that the client can backpropagate the first sub-model according to the gradient of the activation data to obtain the gradient of the first sub-model parameters, and then feed the gradient of the first sub-model parameters back to the server; the activation data is the data output by the client after training the first sub-model. The second sub-model is sent to the model training hosting node so that the model training hosting node can train the second sub-model based on the activation data, obtain the gradient of the second sub-model parameters and the gradient of the activation data, and feed back the gradient of the second sub-model parameters to the server and the gradient of the activation data to the client. Update the global model parameters based on the gradients of the first and second sub-model parameters.

[0115] Or: Receive the gradient of the first sub-model sent by the server and the activation data sent by the model training hosting node; The gradient of the activation data is backpropagated to the first sub-model to obtain the gradient of the first sub-model parameters, and the gradient of the first sub-model parameters is fed back to the server; the activation data is the data output by the client after training the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data; The server is configured to update the global model parameters based on the gradients of the first and second sub-model parameters; the gradient of the second sub-model parameters is obtained by the model training hosting node during the training of the second sub-model based on the activation data.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for splitting federated learning, the method comprising: The application is applied to a model training hosting node, and comprises the following steps of: receiving a second sub-model and activation data, wherein the second sub-model and a first sub-model received by a client together constitute a global model trained by a server; the activation data is output by the client after the first sub-model is trained; training the second sub-model according to the activation data, obtaining a second sub-model parameter gradient and a gradient of the activation data, feeding back the second sub-model parameter gradient to the server, and feeding back the gradient of the activation data to the client, so that the client back-propagates the first sub-model according to the gradient of the activation data, obtains a first sub-model parameter gradient, and feeds back the first sub-model parameter gradient to the server; wherein the server updates a global model parameter according to the first sub-model parameter gradient and the second sub-model parameter gradient.

2. The split federated learning method of claim 1, wherein, The method further comprises the following steps of: when the second sub-model is received, preset first verification information is also received, and the first verification information is used to interact with the client to complete a trusted verification operation on the second sub-model; wherein the client records a trusted verification time of the second sub-model.

3. The split federated learning method of claim 1 or 2, wherein, The method further comprises the following steps of: when the activation data is received, preset second verification information is also received, and the second verification information is used to interact with the server to complete a trusted verification operation on the activation data; wherein the server records a trusted verification time of the activation data.

4. The split federated learning method according to claim 3, wherein the server is configured to record a time of issuing the first sub-model and the second sub-model; and record a first receiving time of receiving the second sub-model parameter gradient; the client is configured to record a sending time of sending the activation data to the model training hosting node; and record a second receiving time of receiving the gradient of the activation data; correspondingly, the server or the client is configured to calculate time consumption information of the model training hosting node in a preparation stage and a calculation stage of training a model according to the trusted verification time of the second sub-model, the trusted verification time of the activation data, the issuing time, the sending time, the first receiving time and the second receiving time.

5. The split federated learning method of claim 3, wherein, The method further comprises the following steps of: feeding back a forward propagation result obtained by training the second sub-model to the client, so that the client calculates a loss value according to the forward propagation result, and sends the loss value to the model training hosting node; back-propagating according to the loss value to obtain data preparation time consumption of the model training hosting node; wherein the client is configured to record a time of receiving the forward propagation result and a time of sending the loss value to the model training hosting node; and determine forward propagation time consumption and back-propagation time consumption of the model training hosting node in the process of training the second sub-model according to the time of receiving the forward propagation result, the time of sending the loss value to the model training hosting node, a start time and an end time of the second sub-model training calculation, and theoretical network time consumption between the client and the model training hosting node.

6. A method for splitting federated learning, characterized in that, The application is applied to a server, comprising: splitting a global model into a first sub-model and a second sub-model, the first sub-model being responsible for training by a client; and the second sub-model being responsible for training by a model training hosting node; downloading the first sub-model to the client, so that the client back-propagates the first sub-model according to the gradient of the activation data, obtains a first sub-model parameter gradient, and feeds back the first sub-model parameter gradient to the server; the activation data being data output by the client after training the first sub-model; downloading the second sub-model to the model training hosting node, so that the model training hosting node trains the second sub-model according to the activation data, obtains a second sub-model parameter gradient and the gradient of the activation data, feeds back the second sub-model parameter gradient to the server, and feeds back the gradient of the activation data to the client; updating a global model parameter according to the first sub-model parameter gradient and the second sub-model parameter gradient.

7. The split federated learning method of claim 6, wherein, The method further comprises: determining preset second verification information sent by the model training hosting node, completing a trusted verification operation of the model training hosting node on the activation data, and recording a trusted verification time of the activation data.

8. The split federated learning method of claim 7, wherein, The method further comprises: receiving first verification information determined by the client after the model training hosting node completes a trusted verification operation on the second sub-model and records a trusted verification time of the second sub-model; the server is configured to record a downloading time of the first sub-model and the second sub-model, and record a first receiving time of receiving the second sub-model parameter gradient; the client is configured to record a sending time of sending the activation data to the model training hosting node, and record a second receiving time of receiving the gradient of the activation data; Accordingly, according to the trusted verification time of the second sub-model, the trusted verification time of the activation data, the downloading time, the sending time, the first receiving time and the second receiving time, the time consumption information of the model training hosting node in the training preparation stage and the training calculation stage of training the model is calculated.

9. The split federated learning method of claim 8, wherein, The training stage includes a training preparation stage, a model preparation stage and a data preparation stage, the trusted verification time of the second sub-model includes a time when the client receives a second sub-model legality verification request and a time when the client returns a verification result to the model training hosting node; and the trusted verification time of the activation data includes a time when the server receives an activation data legality verification request and a time when the server returns a verification result to the model training hosting node; Accordingly, the time consumption information of the model training hosting node in the preparation stage and the calculation stage of training the model is calculated according to the trusted verification time of the second sub-model, the trusted verification time of the activation data, the downloading time, the sending time, the first receiving time and the second receiving time, comprising: determine the training preparation phase time consumption of the model training hosting node according to the issuing time, the sending time, the theoretical network time consumption between the client and the model training hosting node, the theoretical network time consumption between the server and the model training hosting node, and the start time of the second sub-model training calculation; determine the model preparation phase time consumption of the model training hosting node according to the time when the client receives the second sub-model legitimacy verification request, the theoretical network time consumption between the client and the model training hosting node, and the issuing time of the second sub-model issued by the server; determine the data preparation time consumption of the model training hosting node according to the time when the server receives the activation data legitimacy verification request, the theoretical network time consumption between the server and the model training hosting node, and the sending time of the activation data sent to the model training hosting node; the start time of the second sub-model training calculation is determined by the time when the client returns the verification result to the model training hosting node, the time when the server returns the verification result to the model training hosting node, the theoretical network time consumption between the client and the model training hosting node, and the theoretical network time consumption between the server and the model training hosting node.

10. The split federated learning method of claim 8 or 9, wherein, the method further comprises: determine the credibility of each hosting node according to the historical time consumption information, the historical resource usage and the historical network communication quality index information of each hosting node; determine the required model training hosting node according to the credibility of each hosting node. application to the client, comprising:

11. The split federated learning method of claim 6, wherein, receive the gradient of the first sub-model issued by the server and the activation data sent by the model training hosting node; propagate the first sub-model in reverse according to the gradient of the activation data, obtain the first sub-model parameter gradient, and feed back the first sub-model parameter gradient to the server; the activation data is the data output after the client trains the first sub-model; the gradient of the activation data is obtained by the model training hosting node during the training of the second sub-model based on the activation data; ​ 12. A method for splitting federated learning, comprising: ​ ​ ​ The server is configured to update global model parameters according to the first sub-model parameter gradient and the second sub-model parameter gradient.

13. A split federated learning apparatus, comprising: Comprise: The first receiving module is used for receiving a second sub-model and activation data, the second sub-model and a first sub-model received by a client together constitute a global model trained by a server; the activation data is data output after the client trains the first sub-model; The second processing module is used for training the second sub-model according to the activation data, obtaining a second sub-model parameter gradient and a gradient of the activation data, feeding back the second sub-model parameter gradient to the server, and feeding back the gradient of the activation data to the client, so that the client back propagates the first sub-model according to the gradient of the activation data, obtains a first sub-model parameter gradient, and feeds back the first sub-model parameter gradient to the server; wherein the server updates global model parameters according to the first sub-model parameter gradient and the second sub-model parameter gradient.

14. A split federated learning apparatus, comprising: Comprise: The splitting module is used for splitting a global model into a first sub-model and a second sub-model, the first sub-model is responsible for training by a client; The second sub-model is responsible for training by a model training hosting node; The first issuing module is used for issuing the first sub-model to the client, so that the client back propagates the first sub-model according to the gradient of the activation data, obtains a first sub-model parameter gradient, and feeds back the first sub-model parameter gradient to the server; the activation data is data output after the client trains the first sub-model; The second issuing module is used for issuing the second sub-model to the model training hosting node, so that the model training hosting node trains the second sub-model according to the activation data, obtains a second sub-model parameter gradient and a gradient of the activation data, feeds back the second sub-model parameter gradient to the server, and feeds back the gradient of the activation data to the client; The updating module is used for updating global model parameters according to the first sub-model parameter gradient and the second sub-model parameter gradient.

15. A split federated learning apparatus, comprising: Comprise: The second receiving module is used for receiving a first sub-model issued by a server and a gradient of activation data sent by a model training hosting node; The second processing module is used for back propagating the first sub-model according to the gradient of the activation data, obtaining a first sub-model parameter gradient, and feeding back the first sub-model parameter gradient to the server; the activation data is data output after the client trains the first sub-model; The gradient of the activation data is obtained by the model training hosting node during training the second sub-model according to the activation data; The server is configured to update global model parameters according to the first sub-model parameter gradient and the second sub-model parameter gradient; the second sub-model parameter gradient is obtained by the model training hosting node during training the second sub-model according to the activation data.

16. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor implements the split federated learning method as claimed in any one of claims 1-5, or implements the split federated learning method as claimed in any one of claims 6-11, or implements the split federated learning method as claimed in claim 12 when the program is executed.

17. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the split federated learning method as claimed in any one of claims 1-5, or implements the split federated learning method as claimed in any one of claims 6-11, or implements the split federated learning method as claimed in claim 12 when executed by the processor.

18. A computer program product comprising a computer program, characterized in that, The computer program implements the split federated learning method as claimed in any one of claims 1-5, or implements the split federated learning method as claimed in any one of claims 6-11, or implements the split federated learning method as claimed in claim 12 when executed by the processor.

Images