Model training method and device based on federal splitting learning, electronic equipment and storage medium
By performing user group division and credit score weighting operations on the client in federal split learning, the problem of low security of model training is solved, effective defense against gradient reversal, data pollution and model poisoning is achieved, and the security of model training is improved. sex.
Patent Information
- Application Number
- CN202410803049.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-20
- Publication Date
- 2024-11-12
AI Technical Summary
In the prior art, models based on federated split learning have low security problems when training, especially inability to effectively defend against malicious attacks such as gradient reversal, data pollution and model poisoning.
By dividing the client participating in model training user groups, weighting operations are performed based on the credit score and update weight of each client, information aggregation is carried out in combination with model update information, enhance anonymity and hide the specific contribution of the client, thereby improving the model training Security.
Effectively defend against malicious attacks such as gradient reversal, data pollution and model poisoning, improving the security of federal split learning model training.
Smart Images

Figure CN118940270A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to a model training method, device, electronic device and storage medium based on federated split learning. Background Art
[0002] With the booming development of AI content generation applications such as Metaverse and Chat Generative Pre-trained Transformer (ChatGPT), the complexity of intelligent business models has increased exponentially, and the demand for computing power and data has also increased. Federated split learning, as a new type of collaborative computing solution, has attracted widespread attention because it can achieve model training while protecting data privacy. However, as the number of participants increases, the security issues of the system are becoming increasingly prominent. Malicious attacks such as gradient reversal, data contamination, and model poisoning seriously threaten the security and efficiency of the training process.
[0003] Currently, it is impossible to defend against malicious attacks such as gradient reversal, data pollution, and model poisoning, resulting in low security when training models based on federated split learning. Summary of the invention
[0004] The present application provides a model training method, device, electronic device and storage medium based on federated split learning, so as to solve the defect of low security in model training based on federated split learning in the prior art, and improve the security of model training based on federated split learning.
[0005] The present application provides a model training method based on federated split learning, which includes: dividing the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; determining the credit scores of the corresponding clients according to the first model update information of the local model uploaded by each client in each user group; determining the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit scores of other clients in the same user group; performing weighted operations according to the target update weight of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; and performing information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
[0006] According to a model training method based on federated split learning provided by the present application, when determining the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit scores of other clients in the same user group, the following operations are performed for each client: determine the scaling factor of the current client according to the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client; determine the target update weight of the current client based on the scaling factor of the current client, the current update weight and the loss information corresponding to the current update weight.
[0007] According to a model training method based on federated split learning provided by the present application, the scaling factor of the current client is determined based on the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client, including: respectively determining the influence coefficients of all clients in the user group corresponding to the current client; for each client in the user group corresponding to the current client, multiplying the influence coefficient of each client by its credit score to obtain a first product of the corresponding client; adding the first products of all clients in the user group corresponding to the current client to obtain a sum value; performing a ratio operation on the first product of the current client and the sum value to obtain a first ratio operation value; multiplying the first ratio operation value by a preset normalization factor to obtain the scaling factor of the current client.
[0008] According to a model training method based on federated split learning provided by the present application, the target update weight of the current client is determined based on the scaling factor of the current client, the current update weight and the loss information corresponding to the current update weight, including: determining the loss information corresponding to the current update weight of the current client; performing a ratio operation on a preset learning rate and the scaling factor of the current client to obtain a second ratio operation value; multiplying the second ratio operation value by the loss information corresponding to the current update weight of the current client to obtain a second product; subtracting the current update weight of the current client from the second product to obtain the target update weight of the current client.
[0009] According to a model training method based on federated split learning provided by the present application, after determining the credit scores of the corresponding clients respectively according to the first model update information of the local model uploaded by each client in each user group, it also includes: determining the dynamic score threshold of each user group respectively according to the credit scores of each client in each user group; for each client in each user group, if the credit score of any client is greater than the dynamic score threshold of the corresponding user group, the corresponding client is eliminated.
[0010] According to a model training method based on federated split learning provided in the present application, the first model update information of the local model uploaded by each client is the model update information after the disturbance is introduced.
[0011] According to a model training method based on federated split learning provided in the present application, the local model is at least one of an image recognition model, a speech recognition model, a natural language processing model and an information prediction model.
[0012] The present application also provides a model training device based on federated split learning, including: a division module, used to divide the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; a first determination module, used to determine the credit scores of the corresponding clients according to the first model update information of the local model uploaded by each client in each user group; a second determination module, used to determine the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit scores of other clients in the same user group; an operation module, used to perform weighted operations according to the target update weights of each client in each user group and the first model update information, to obtain the second model update information of the corresponding user group; an aggregation module, used to perform information aggregation based on each second model update information, to obtain the target model update information of the global model corresponding to each local model.
[0013] The present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a model training method based on federated split learning as described in any one of the above is implemented.
[0014] The present application also provides a storage medium, which is a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the model training method based on federated split learning as described in any of the above.
[0015] The model training method, device, electronic device and storage medium based on federated split learning provided by the present application can enhance anonymity by dividing the clients participating in the model training of the same federated split learning into user groups in each round of model training of federated split learning to obtain at least two user groups; further, according to the first model update information of the local model uploaded by each client in each user group, the credit score of the corresponding client is determined respectively; and according to the credit score of each client, the current update weight and the credit score of other clients in the same user group, the target update weight of the corresponding client is determined respectively; and then, according to the target update weight of each client in each user group and the first model update information, a weighted operation can be performed respectively to obtain the second model update information of the corresponding user group, and by aggregating the update weight of each client in combination with the model update information within the group, the specific contribution of each client can be hidden to protect individual privacy; thus, information aggregation can be performed based on each second model update information to obtain the target model update information of the global model corresponding to each local model, and since anonymity is enhanced and the specific contribution of each client is hidden, malicious attacks such as gradient reversal, data pollution, and model poisoning can be effectively defended, thereby improving the security of model training based on federated split learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] In order to more clearly illustrate the technical solutions in the present application or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0017] Figure 1 It is a flowchart of the model training method based on federated split learning provided in this application.
[0018] Figure 2 It is a structural schematic diagram of the electronic device provided by this application. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0020] It should be noted that in the description of the present application, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or position relationship indicated by the terms "upper", "lower" and the like is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. Unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.
[0021] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.
[0022] Combine the following Figure 1-Figure 2 The present application describes a model training method, device, electronic device, and storage medium based on federated split learning.
[0023] Figure 1 This is a flowchart of the model training method based on federated split learning provided in this application. Figure 1 As shown in the figure, in each round of model training of federated split learning, this method includes:
[0024] Step 100: Divide the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups.
[0025] Step 200: Determine the credit score of each client in each user group according to the first model update information of the local model uploaded by each client.
[0026] Step 300: Determine the target update weight of each client according to the credit score of each client, the current update weight, and the credit scores of other clients in the same user group.
[0027] Step 400 , performing weighted calculation according to the target update weight of each client in each user group and the first model update information, to obtain the second model update information of the corresponding user group.
[0028] Step 500: Aggregate information based on each second model update information to obtain target model update information of the global model corresponding to each local model.
[0029] It should be noted that the execution subject of the model training method based on federated split learning provided in the embodiment of the present application can be a server, a computer device, etc., such as a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (Ultra-mobile Personal Computer, UMPC), a netbook or a personal digital assistant (Personal Digital Assistant, PDA), etc. A model training device based on federated split learning may be set or connected to the server or computer device of the present application. In the present application, the model training device based on federated split learning can be controlled to execute the model training method based on federated split learning of the present application by controlling the model training device based on federated split learning.
[0030] The server of the present application can collaborate with multiple clients as participants to train a shared model through federated split learning. The shared model can be, for example, an image recognition model, a speech recognition model, a natural language processing model, and an information prediction model. It can be applied to fields such as medicine, finance, and communications. The image recognition model is a machine learning model used to identify objects, scenes, or patterns in digital images. The speech recognition model is a machine learning model used to convert voice signals into text or commands. The natural language processing model is a machine learning model used to process and understand human language text. Such models can include tasks such as text classification, named entity recognition, sentiment analysis, and machine translation. The information prediction model is a machine learning model used to predict future events, trends, or results. This model can be applied to tasks such as time series analysis, regression analysis, and predicting market trends.
[0031] A multi-way shuffler may be provided in the server of the present application, and the multi-way shuffler may be used to periodically divide the clients participating in the federated split learning into multiple groups, and each group is connected to a shuffler.
[0032] Therefore, in each round of model training of federated split learning, the present application divides the clients participating in the same federated split learning model training into user groups to obtain at least two user groups.
[0033] Specifically, the present application may use multiple shufflers S = {S1, ..., S M}, all training participants, i.e., clients, are periodically divided into M groups, each of which is connected to a shuffler. This grouping method can enhance anonymity and data security because it isolates the data and learning process of each group.
[0034] In each round of training t, the M User Group U t m Contains members {u1,u2,…,u km}, where the number of members in each group is k m It is guaranteed to be at least k0 and less than N, where the specific values of k0 and N can be set according to actual needs. This is done to ensure that each group has enough members to protect individual privacy while avoiding over-centralization.
[0035] Furthermore, each client in each user group performs local model training separately. During the training process of the local model, the clients not only update their model parameters to obtain model update information, but also introduce algorithmic perturbations, such as processing the model update information by adding noise to achieve differential privacy, which helps to enhance the model's ability to resist external attacks and data leakage. After introducing the perturbation, the first model update information of each client is obtained and uploaded to the server. It should be noted that the training data for local model training is determined according to the specific type of the model, and is not specifically limited in this application.
[0036] In one embodiment, the present application may introduce the following local disturbance mechanism:
[0037] R r (w it )=w it +n it ;
[0038] Among them, n it is the original weight w it The additive random noise of the interference; R r (w it ) represents the weight after the perturbation; we sample n from a zero-mean Gaussian distributionit , the distribution is similar to w it In particular, in the federated learning scenario, we use the l2-norm to evaluate the distance of weight updates:
[0039]
[0040] where r is the number of clients in the group of users; the weights uploaded from benign federated learning clients are naturally bounded, which means that a carefully selected threshold can be adaptively used to limit the range of weights without significantly reducing the utility. Therefore, in our framework, we apply the widely used clipping technique to reduce the weight w it Replace with w it / max{1;∥w it}} 2 / C, where C is the median of the l2-norm of the unclipped weights. Therefore, the formula is reformulated as:
[0041]
[0042] Where I is the identity matrix, N(0,s 2 I) represents a value that satisfies a normal distribution.
[0043] The server may receive first model update information of the local model uploaded by each client.
[0044] Furthermore, the present application can determine the credit score of the corresponding client according to the first model update information of the local model uploaded by each client in each user group.
[0045] Specifically, for each user group, the first model update information of each client in the user group can be input into the credit scoring model to obtain the overall credit score of the user group output by the credit scoring model, and the overall credit score of the user group is used as the credit score of each client in the user group. The credit scoring model can be a model that is pre-built and trained and used to predict the overall credit score based on multiple model update information, or it can be a model corresponding to the following expression:
[0046]
[0047] Among them, f (Efficiency) represents the function of learning efficiency, g (Privacy Risk) represents the function of information leakage risk, λ is a trade-off coefficient used to balance efficiency and privacy, S i represents the i-th user group, and P represents the shuffler; both the function of learning efficiency and the function of information leakage risk include a method for processing the input model update information, which is not specifically limited in this application and can be implemented using any existing technology.
[0048] The importance of accuracy and privacy protection can be balanced in this application.
[0049] Furthermore, the present application determines the target update weight of each client according to the credit score of each client, the current update weight, and the credit scores of other clients in the same user group.
[0050] Furthermore, the present application can perform weighted calculations according to the target update weight of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group.
[0051] Furthermore, the present application can aggregate information based on each second model update information to obtain target model update information of the global model corresponding to each local model, and further, the target model update information can be sent back to each client. This process is repeated until the global model reaches the expected training quality or meets a specific stop condition.
[0052] The model training method based on federated split learning provided in the embodiment of the present application can enhance anonymity by dividing the clients participating in the model training of the same federated split learning into user groups in each round of model training of federated split learning to obtain at least two user groups; further, according to the first model update information of the local model uploaded by each client in each user group, the credit score of the corresponding client is determined respectively; and according to the credit score of each client, the current update weight and the credit score of other clients in the same user group, the target update weight of the corresponding client is determined respectively; and then, according to the target update weight of each client in each user group and the first model update information, a weighted operation can be performed respectively to obtain the second model update information of the corresponding user group; by aggregating the update weight of each client in combination with the model update information within the group, the specific contribution of each client can be hidden to protect individual privacy; thus, information aggregation can be performed based on each second model update information to obtain the target model update information of the global model corresponding to each local model, and since anonymity is enhanced and the specific contribution of each client is hidden, malicious attacks such as gradient reversal, data pollution, and model poisoning can be effectively defended, thereby improving the security of model training based on federated split learning.
[0053] In one embodiment, when determining the target update weight of the corresponding client according to the credit score of each client, the current update weight, and the credit scores of other clients in the same user group, the following operations are performed for each client:
[0054] Determine a scaling factor for the current client based on the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client;
[0055] Based on the scaling factor of the current client, the current update weight, and the loss information corresponding to the current update weight, the target update weight of the current client is determined.
[0056] Further, the scaling factor of the current client is determined according to the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client, including:
[0057] Determine the influence coefficients of all clients in the user group corresponding to the current client respectively;
[0058] For each client in the user group corresponding to the current client, multiply the influence coefficient of each client by its credit score to obtain a first product of the corresponding client;
[0059] Add the first products of all clients in the user group corresponding to the current client to obtain a sum value;
[0060] Perform a ratio operation on the first product of the current client and the sum value to obtain a first ratio operation value;
[0061] The first ratio operation value is multiplied by a preset normalization factor to obtain a scaling factor of the current client.
[0062] Further, based on the scaling factor of the current client, the current update weight and the loss information corresponding to the current update weight, determining the target update weight of the current client includes:
[0063] Determine the loss information corresponding to the current updated weight of the current client;
[0064] Performing a ratio operation on the preset learning rate and the scaling factor of the current client to obtain a second ratio operation value;
[0065] Multiplying the second ratio operation value by the loss information corresponding to the current updated weight of the current client to obtain a second product;
[0066] The current update weight of the current client is subtracted from the second product to obtain the target update weight of the current client.
[0067] Specifically, when the target update weight of the corresponding client is determined according to the credit score of each client, the current update weight, and the credit scores of other clients in the same user group, for the client currently being processed, the scaling factor of the client can be determined by combining the credit score of the client and the credit scores of other clients in the user group corresponding to the current client with the following calculation formula:
[0068]
[0069] Among them, n i represents the influence coefficient of the i-th client, B(ui,t) is the dynamic credit score of client ui at time t, B(uj,t) represents the dynamic credit score of client uj at time t; r is a pre-set normalization factor to ensure that the sum of the scaling factors is 1; N is the total number of clients in the user group.
[0070] It should be noted that the above formula represents the scaling factor p of each client. i,t is the ratio of the product of its credit score and its influence coefficient, relative to the sum of the corresponding products of all clients. This approach allows the weight of each client to be dynamically adjusted to reflect the credibility and potential maliciousness of its behavior, thereby effectively reducing the weight of malicious users in system decisions and enhancing the overall security of the system.
[0071] Furthermore, the target update weight of the client can be determined according to the scaling factor of the client, the current update weight, the preset learning rate and the loss information corresponding to the current update weight, combined with the following calculation formula:
[0072]
[0073] in, represents the target update weight of client i at time t, represents the target update weight of client i at time t-1, that is, the current update weight, η represents the preset learning rate, It represents the loss information corresponding to the current updated weight. The present application does not specifically limit the process of obtaining the loss information corresponding to the current updated weight.
[0074] By determining the scaling factor of the client, the present application can gradually minimize the impact of the poisoning attack, and further can determine the target update weight of the client based on the scaling factor based on the client, the current update weight and the loss information corresponding to the current update weight, and then can perform weighted operations according to the target update weight of each client in each user group and the first model update information, to obtain the second model update information of the corresponding user group, and by aggregating the update weight of each client in combination with the model update information within the group, the specific contribution of each client can be hidden to protect individual privacy; thereby, information aggregation can be performed based on each second model update information to obtain the target model update information of the global model corresponding to each local model, and since anonymity is enhanced and the specific contribution of each client is hidden, malicious attacks such as gradient reversal, data pollution, and model poisoning can be effectively defended, thereby improving the security of model training based on federated split learning.
[0075] In one embodiment, after determining the credit score of the corresponding client according to the first model update information of the local model uploaded by each client in each user group, the method further includes:
[0076] Determine a dynamic score threshold for each user group based on the credit score of each client in each user group;
[0077] For each client in each user group, if the credit score of any client is greater than the dynamic score threshold of the corresponding user group, the corresponding client will be eliminated.
[0078] This application can maintain a dynamic scoring threshold Thr to eliminate opponents with abnormal credit scores, i.e., clients. Specifically, the dynamic scoring threshold for each user group can be determined based on the credit scores of each client in each user group, combined with the following calculation formula:
[0079] U * = {U|B(u i ,t)>Thr,Thr=m * +as * +bias,u i ∈U};
[0080] Among them, m * Represents client u at time t i The mean credit score of the user group, s * Represents client u at time t i The standard deviation of the credit scores of the user group in which the user is located; a is an adjustable threshold rate that provides a satisfactory false positive limit. Bias represents the deviation value. In this application, a=2 and bias=s * According to the current network security situation, the specific values of deviation and a can also be further adjusted.
[0081] It should be noted that the present application may also adopt other threshold mechanisms, but the optimal threshold mechanism depends on the selection of neural network models and hyperparameters in specific scenarios.
[0082] Furthermore, the credit score of each client in each user group can be compared with the dynamic score threshold of the corresponding user group. If the credit score of any client is greater than the dynamic score threshold of the corresponding user group, the client will be eliminated, that is, it cannot participate in the next round of model training.
[0083] This application can avoid malicious attacks such as gradient reversal, data pollution, and model poisoning by eliminating abnormal clients so that they cannot participate in subsequent model training. Therefore, it can improve the security of model training based on federated split learning.
[0084] In one embodiment, in the multi-shuffler-based federated split learning mechanism provided in the present application for implementing the above method, the workflow is carefully designed to strengthen data privacy protection and optimize model training efficiency. First, the system uses a set of shufflers S = {S1, ..., S M}Realize dynamic grouping of participants, and reallocate users to M groups every cycle t Ensure that the number of members in each group is between k0 and N to maintain the anonymity and data diversity of each group. At the beginning of each round of learning, users are randomly assigned to different groups through a shuffler and then undergo local model training. During local training, users not only update their model parameters, but also introduce algorithmic perturbations, such as adding noise to achieve differential privacy, which helps enhance the model's ability to resist external attacks and data leakage. To further protect user privacy, the model weights within each group will be mixed and split through a process called model shuffling. This step ensures that the model weight updates within the group do not directly expose the information of any individual user, thereby hiding the individual contribution while keeping the model structure unchanged. After the model update is completed, the defender conducts a comprehensive evaluation based on the instant malicious score and the user's upload time history score to identify and screen out possible malicious participants. The aggregation step uses an unequal weight aggregation strategy to process the model updates of each group based on the aforementioned evaluation results. This strategy not only takes into account the instant evaluation value, but also integrates the user's long-term credit performance, thereby optimizing the learning results of the entire network and the overall performance of the model. Through such multi-layer protection and optimization measures, the system effectively achieves a balance between maintaining efficient learning and strengthening data privacy.
[0085] In this application, random grouping is performed through a shuffler to eliminate high-risk users, and each user is given an instant malicious score for subsequent elimination.
[0086] Users train the model locally and introduce perturbations to achieve differential privacy or increase the difficulty of identifying the adversary model.
[0087] Split and mix model weights within groups, hiding weight updates while keeping the model structure unchanged.
[0088] After receiving each shuffler update, the user is evaluated based on the instant evaluation value and the average of the historical credit score based on the upload time.
[0089] Based on the evaluation values, unequal aggregation is performed and the user updates are made public to ensure the security and effectiveness of the overall model update.
[0090] This application constructs the data allocation and model learning problems as multi-objective optimization problems, so that how to choose the best combination of data and model parameters is transformed into finding the optimal strategy for the multi-objective planning. By comprehensively considering the diversity of data and the generalization ability of the model, the policy optimization problem is further transformed into a reinforcement learning problem, a machine learning quintuple is established, and the newly proposed privacy protection criterion is added to the reward function. This application also transforms the multi-way shuffler, reduces the complexity of raw data processing, and adds a new data allocation mechanism, and obtains a multi-way shuffler federated split learning algorithm, which is superior to existing federated learning solutions in multiple performance and security indicators such as global data diversity, average learning efficiency, and privacy protection success rate.
[0091] The model training device based on federated split learning provided in the present application is described below. The model training device based on federated split learning described below and the model training method based on federated split learning described above can be referenced to each other.
[0092] Furthermore, the present application also provides a model training device based on federated split learning.
[0093] The model training device based on federated split learning includes: a division module, which is used to divide the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; a first determination module, which is used to determine the credit scores of the corresponding clients according to the first model update information of the local model uploaded by each client in each user group; a second determination module, which is used to determine the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit scores of other clients in the same user group; an operation module, which is used to perform weighted operation according to the target update weight of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; an aggregation module, which is used to perform information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
[0094] In one embodiment, the first determination module is further used to: determine a dynamic score threshold for each user group according to the credit score of each client in each user group; for each client in each user group, if the credit score of any client is greater than the dynamic score threshold of the corresponding user group, the corresponding client is eliminated.
[0095] In one embodiment, the second determination module is specifically used to perform the following operations for each client when determining the target update weight of the corresponding client based on the credit score of each client, the current update weight and the credit scores of other clients in the same user group: determining the scaling factor of the current client based on the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client; determining the target update weight of the current client based on the scaling factor of the current client, the current update weight and the loss information corresponding to the current update weight.
[0096] In one embodiment, the second determination module includes a first determination unit, which is used to: respectively determine the influence coefficients of all clients in the user group corresponding to the current client; for each client in the user group corresponding to the current client, multiply the influence coefficient of each client by its credit score to obtain a first product of the corresponding client; add the first products of all clients in the user group corresponding to the current client to obtain a sum value; perform a ratio operation on the first product of the current client and the sum value to obtain a first ratio operation value; multiply the first ratio operation value by a preset normalization factor to obtain a scaling factor of the current client.
[0097] In one embodiment, the second determination module includes a second determination unit, which is used to: determine the loss information corresponding to the current update weight of the current client; perform a ratio operation on the preset learning rate and the scaling factor of the current client to obtain a second ratio operation value; multiply the second ratio operation value by the loss information corresponding to the current update weight of the current client to obtain a second product; subtract the current update weight of the current client from the second product to obtain the target update weight of the current client.
[0098] The model training device based on federated split learning provided by the embodiment of the present application can enhance anonymity by dividing the clients participating in the model training of the same federated split learning into user groups in each round of model training of federated split learning to obtain at least two user groups; further, the credit score of the corresponding client is determined respectively according to the first model update information of the local model uploaded by each client in each user group; and the target update weight of the corresponding client is determined respectively according to the credit score of each client, the current update weight and the credit score of other clients in the same user group; and then the target update weight of each client in each user group and the first model update information can be weighted according to the target update weight of each client in each user group to obtain the second model update information of the corresponding user group, and the specific contribution of each client can be hidden by aggregating the update weight of each client in combination with the model update information within the group to protect individual privacy; thus, information aggregation can be performed based on each second model update information to obtain the target model update information of the global model corresponding to each local model, and since anonymity is enhanced and the specific contribution of each client is hidden, malicious attacks such as gradient reversal, data pollution, and model poisoning can be effectively defended, thereby improving the security of model training based on federated split learning.
[0099] Figure 2 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 2 As shown, the electronic device may include: a processor 210, a communication interface 220, a memory 230 and a communication bus 240, wherein the processor 210, the communication interface 220 and the memory 230 communicate with each other through the communication bus 240. The processor 210 may call the logic instructions in the memory 230 to execute the model training method based on federated split learning, the method comprising: dividing the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; determining the credit score of the corresponding client according to the first model update information of the local model uploaded by each client in each user group; determining the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit score of other clients in the same user group; performing weighted operation according to the target update weight of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; performing information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
[0100] In addition, the logic instructions in the above-mentioned memory 230 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0101] On the other hand, the present application also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the model training method based on federated split learning provided by the above methods, the method including dividing the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; determining the credit score of the corresponding client according to the first model update information of the local model uploaded by each client in each user group; determining the target update weight of the corresponding client according to the credit score of each client, the current update weight and the credit score of other clients in the same user group; performing weighted operations based on the target update weight of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; performing information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
[0102] On the other hand, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the model training method based on federated split learning provided by the above-mentioned methods, the method comprising: dividing the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; determining the credit scores of the corresponding clients according to the first model update information of the local model uploaded by each client in each user group; determining the target update weights of the corresponding clients according to the credit scores of each client, the current update weights and the credit scores of other clients in the same user group; performing weighted operations based on the target update weights of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; performing information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
[0103] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.
[0104] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.
[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A model training method based on federated split learning, characterized in that: In each round of federated split learning model training, it includes: Divide the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; Determine the credit score of each client in each user group according to the first model update information of the local model uploaded by each client; Determine the target update weight of each client according to its credit score, current update weight, and credit scores of other clients in the same user group; Performing weighted calculations according to the target update weights of each client in each user group and the first model update information to obtain the second model update information of the corresponding user group; Information aggregation is performed based on each second model update information to obtain target model update information of the global model corresponding to each local model.
2. The model training method based on federated split learning according to claim 1, characterized in that: When determining the target update weight of the corresponding client according to the credit score of each client, the current update weight, and the credit scores of other clients in the same user group, the following operations are performed for each client: Determine a scaling factor for the current client based on the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client; Based on the scaling factor of the current client, the current update weight, and the loss information corresponding to the current update weight, the target update weight of the current client is determined.
3. The model training method based on federated split learning according to claim 2 is characterized in that: Determining the scaling factor of the current client according to the credit score of the current client and the credit scores of other clients in the user group corresponding to the current client includes: Determine the influence coefficients of all clients in the user group corresponding to the current client respectively; For each client in the user group corresponding to the current client, multiply the influence coefficient of each client by its credit score to obtain a first product of the corresponding client; Add the first products of all clients in the user group corresponding to the current client to obtain a sum value; Perform a ratio operation on the first product of the current client and the sum value to obtain a first ratio operation value; The first ratio operation value is multiplied by a preset normalization factor to obtain a scaling factor of the current client.
4. The model training method based on federated split learning according to claim 2, characterized in that: The step of determining the target update weight of the current client based on the scaling factor of the current client, the current update weight, and the loss information corresponding to the current update weight includes: Determine the loss information corresponding to the current updated weight of the current client; Performing a ratio operation on the preset learning rate and the scaling factor of the current client to obtain a second ratio operation value; Multiplying the second ratio operation value by the loss information corresponding to the current updated weight of the current client to obtain a second product; The current update weight of the current client is subtracted from the second product to obtain the target update weight of the current client.
5. The model training method based on federated split learning according to claim 1, characterized in that: After determining the credit score of the corresponding client according to the first model update information of the local model uploaded by each client in each user group, the method further includes: Determine a dynamic score threshold for each user group based on the credit score of each client in each user group; For each client in each user group, if the credit score of any client is greater than the dynamic score threshold of the corresponding user group, the corresponding client will be eliminated.
6. The model training method based on federated split learning according to claim 1, characterized in that: The first model update information of the local model uploaded by each client is the model update information after the disturbance is introduced.
7. The model training method based on federated split learning according to any one of claims 1 to 6, characterized in that: The local model is at least one of an image recognition model, a speech recognition model, a natural language processing model and an information prediction model.
8. A model training device based on federated split learning, characterized in that: include: A partitioning module, used to partition the clients participating in the model training of the same federated split learning into user groups to obtain at least two user groups; A first determination module, configured to determine the credit score of each client in each user group according to first model update information of the local model uploaded by each client; A second determination module is used to determine the target update weight of each client according to the credit score of each client, the current update weight and the credit scores of other clients in the same user group; A calculation module, used for performing weighted calculation according to the target update weight of each client in each user group and the first model update information, to obtain the second model update information of the corresponding user group; The aggregation module is used to perform information aggregation based on each second model update information to obtain the target model update information of the global model corresponding to each local model.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the model training method based on federated split learning as described in any one of claims 1 to 7 is implemented.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the model training method based on federated split learning as described in any one of claims 1 to 7 is implemented.
