CN115775025A - Lightweight federated learning method and system for space-time data heterogeneous scene - Google Patents

Abstract

The invention discloses a lightweight federated learning method and system oriented to spatio-temporal data heterogeneous scenarios, which can assign different communication probabilities to clients according to changes in client data distribution, so that the degree of data distribution changes greatly and is not affected by Clients perceived by the federated global model participate in federated communication with a higher probability, instead of increasing the communication traffic of all clients equally. By tilting communication resources, it is possible to sense global data distribution changes with less communication traffic, and improve the communication efficiency of federated learning in spatiotemporal heterogeneous scenarios. By using the method and system proposed by the present invention, each image collection and analysis terminal uses the change information of its private case data distribution and the gradient information of the private detection model of the image collection and analysis terminal to dynamically adjust the participation of image collection and analysis terminals in each round of federation process. The probability of communication can achieve the same federated learning performance with less federated communication cost, effectively reducing the federated learning traffic.

Description

A lightweight federated learning method and system for spatiotemporal data heterogeneous scenarios

technical field

The invention relates to the technical field of federated learning, in particular to a lightweight federated learning method and system for heterogeneous spatiotemporal data scenarios.

Background technique

In recent years, deep learning technology has significantly improved the performance of tasks such as computer vision and natural language processing. By training a deep neural network (DNN) with millions of parameters on massive data, the performance of deep learning models has gradually surpassed traditional machine learning and has become the mainstream of the artificial intelligence technology industry. Due to its dependence on massive training data, traditional deep learning usually adopts a centralized model of collecting data on edge devices and uploading it to cloud servers for model training and inference services.

However, in recent years, with the gradual promulgation of data security and personal privacy protection bills, the process of uploading private data collected from mobile phones, tablets and other client devices to the central server faces many restrictions. This trend makes the previous mode of centralized model training and service of end device collection full of challenges. To address this problem, federated learning (FL) is an attractive solution that enables collaborative learning across various devices without sharing private data. Specifically, FL is a decentralized collaborative training framework that transfers encrypted models or gradients with the help of federated servers to collaboratively train each client model. In the FL process, the model training work based on private data is offloaded to the client for local execution, rather than uploaded to the federated server for centralized training, thereby protecting the security of private data.

However, the application of federated learning in practical scenarios needs to face the spatiotemporal heterogeneous data environment. On the one hand, federated learning faces the challenge of spatial heterogeneity, that is, the client's private data may not satisfy the independent and identical distribution assumption. For example, the distribution of detected causes is different due to the characteristics of cases admitted to each hospital. Specifically, the distribution of causes of patients in orthopedic hospitals and infectious disease hospitals will be significantly different. Applying federated learning on such data will cause the global model update direction to be inconsistent with the real data distribution, resulting in slow model convergence or poor performance, that is, spatial heterogeneity in the data. On the other hand, federated learning faces the challenge of time heterogeneity. The distribution of client private data will change over time. From the perspective of the federated system, the distribution of global data is unstable. For example, due to time-related factors such as seasons and epidemics, the proportion of allergic diseases detected in each hospital regularly increases from April to May every year. That is, data storage time heterogeneity. The above two phenomena are common in the application scenarios of federated learning. In the face of heterogeneous data distribution in time and space, the federated server needs to communicate frequently with the client to maintain the consistency between the target distribution of the federated global model and the real distribution.

While the performance of the deep model continues to improve, the demand for multi-modal information processing and the demand for model representation and reasoning capabilities make the structure of the model larger and larger. For example, the VGG deep network commonly used in medical image recognition tasks in intelligent medical scenarios has reached 100 million parameters. The GPT-3 deep model, which has outstanding performance and is often used for medical assistant tasks in intelligent medical scenarios, even has an astonishing 100 billion parameters. Frequent communication of large models will lead to heavy communication costs and power consumption of mobile smart devices, which seriously hinders the application of federated learning. Furthermore, the above contradictions give rise to the challenge of low-cost client collaborative training under the premise of not exposing client private data in the spatio-temporal data heterogeneous scenario.

In response to the lightweight challenge of the federated model, existing work is carried out from two perspectives: reducing the number of federated communications and reducing the amount of federated communications. Among them, in terms of reducing the number of federated communications, the Fed average algorithm controls the communication cost of clients by adjusting the proportion of randomly selected clients in each round of federated learning. However, this method lacks the consideration of the differentiation of client value. The FedPNS method performs probability-based communication control on the client from the perspective of the gradient consistency between the client and the federation global model. This method uses the similarity between the client gradient update direction and the federation global update gradient to measure the importance of the client, and based on this important The property measures the communication probability of the distribution federation. In addition to the gradient-based method, methods such as FedMCCS use a multi-standard client selection mechanism based on client performance to tilt the communication demand to clients with relatively abundant resources, thereby balancing the resource load of the client and reducing the communication demand of weak resource clients. Reduce the communication efficiency of federated learning. On the other hand, in terms of reducing the amount of federated communication, existing work proposes a variety of small-model communication methods from the perspective of compressing the amount of communication models. The Hermes method proposes that in a federated system with inconsistent client data distribution, the federated global model generates a supernetwork for learning global knowledge, and uses the neural network pruning method to select a subnetwork with fewer parameters for each client. Client's private knowledge. In the process of federated communication, the client and server only interact with the intersection part of the model, thereby reducing the amount of communication parameters in a single federated process.

However, most of the existing works assume that the distribution of clients is stable over time. Therefore, when the data distribution changes over time, the existing methods can only increase the communication probability of all clients indiscriminately so that the federated global model can perceive the change of data distribution as soon as possible. This undoubtedly further exacerbated the communication burden on the federal system.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a lightweight federated learning method and system for spatio-temporal data heterogeneous scenarios, which can assign different communication probabilities to clients according to changes in client data distribution, thereby allowing data distribution to change Clients with a large degree and not perceived by the federated global model participate in federated communication with a higher probability, instead of increasing the communication traffic of all clients equally. By slanting communication resources, it is possible to perceive global data distribution changes with less communication traffic, thereby improving the communication efficiency of federated learning in spatiotemporal heterogeneous scenarios.

In order to achieve the above object, the present invention provides the following technical solutions:

On the one hand, the present invention provides a lightweight federated learning method for heterogeneous spatiotemporal data scenarios, including the following steps:

S1. Client local update: After receiving the latest federated global model, use the local private data of the client to update the federated global model to a localized model that is more in line with the distribution of local private data;

S2. Client importance evaluation: use the degree of change in the distribution of private data and the consistency between the gradient information of the client model and the gradient information of the global model to measure the importance of the client's participation in federated communication to the federated global model;

S3. Client communication probability control: according to the client importance obtained in step S2, dynamically adjust the client communication probability, judge whether to upload according to the probability, if so, enter step S4, if not, send the federation global model to all clients;

S4. Federation server aggregation: use the gradient information uploaded by the clients participating in the federation update in this round to update the federation global model, and deliver the updated federation global model to all clients, and enter the next round of federation process until Reach the preset federation round.

继续训练，得到本地模型

户端的优化目标函数为

其中，n为

中样本量，f为客户端i的损失函数。在分类任务中通常为交叉熵损失。Further, the specific process of step S1 is: locally on the client, on the basis of the federated global model ω ^t , use the private data set of client i in round t through the gradient descent iterative algorithm

Continue training to get the local model

The optimization objective function of the client is

Among them, n is

Medium sample size, f is the loss function of client i. In classification tasks it is usually cross-entropy loss.

Further, the specific process of step S2 is:

首先计算模型

在最新数据分篇上的Top-1准确率

其次计算模型

在最近n轮数据分片上的准确率

最后计算两个不同时间端数据在同模型上的准确率差异的绝对值为

反映客户端本地数据绝对分布变化，记做

S21. The model obtained by client i after the tth round of local update is

First calculate the model

Top-1 accuracy rate on the latest data subsection

Second calculation model

Accuracy on the last n rounds of data sharding

Finally, calculate the absolute value of the accuracy difference between two different time-end data on the same model

Reflect the absolute distribution change of the client's local data, record as

对比其与本轮的本地模型

的差异，得到未上传客户端梯度信息

对比客户端计算最近一次上传后接收的联邦全局模型ω_s与最新联邦全局模型ω_t的差异，得到联邦全局模型已感知梯度信息

通过计算客户端梯度信息与联邦全局模型梯度信息的余弦相似度

衡量客户端待上传梯度与联邦全局模型已感知梯度的相对差异，记做

S22. The client caches the latest local model uploaded to the federated server

Compare it with the local model for this round

The difference, get the gradient information of the client that has not been uploaded

Compare the difference between the federated global model ω _s received by the client after the last upload and the latest federated global model ω _t , and obtain the perceived gradient information of the federated global model

By calculating the cosine similarity between the gradient information of the client and the gradient information of the federation global model

Measure the relative difference between the gradient to be uploaded by the client and the perceived gradient of the federated global model, and record as

S23. Combining the absolute change AC and the relative change RC in the form of multiplication as the importance measure of client i in round t:

Further, the specific process of step S3 is:

其中tau控制联邦系统为数据分布的时间变化分配的通信资源，反应对分布变化的敏感程度，较大的tau将使联邦系统更快的适应客户端的分布变化，但也会产生额外的较高的通信开销；P_floor控制客户端的基础通信概率；因此客户端S的重要性较高则上行概率较大，反之客户端重要性较低则上行概率较小。S31. Input the client importance measure value into the communication probability mapping module, and the communication probability of client i is updated as

Among them, tau controls the communication resources allocated by the federated system for time changes in data distribution, reflecting the sensitivity to distribution changes. A larger tau will enable the federated system to adapt to client distribution changes faster, but it will also generate additional higher Communication overhead; P _floor controls the basic communication probability of the client; therefore, the higher the importance of the client S, the higher the uplink probability, and the lower the importance of the client, the lower the uplink probability.

S32. The communication probability control module generates a one-hot vector of the client to be uploaded according to the probability in each round of federation process, that is, a OneHot vector, in which the value of one dimension is 1, and the value of other dimensions is 0, and requires the corresponding client to upload the local model gradient information.

Further, the specific process of step S4 is:

S41. The selected client i uploads the cumulative gradient of the gradient from time s to time t in round t

S42. The client set uploaded in this round is C ^t , and the historical federated global model is ω _t . Update the cumulative gradient information of uploaded clients to the federated global model at a learning rate η

S43. Deliver the updated federated global model to all clients.

On the other hand, the present invention also provides a lightweight federated learning system for spatiotemporal data heterogeneous scenarios, including the following modules to implement any of the methods described above:

The client local update module is used to update the federated global model to a localized model that is more in line with the distribution of local private data by using the local private data of the client after receiving the latest federated global model;

The client importance evaluation module is used to measure the importance of the client's participation in federated communication to the federated global model by using the degree of distribution change of private data and the consistency between the gradient information of the client model and the gradient information of the global model;

The client communication probability control module dynamically adjusts the client communication probability according to the client importance obtained by the client importance evaluation module, and judges whether to upload according to the probability. sent to all clients;

The federation server aggregation module is used to update the federation global model by using the gradient information uploaded by the clients participating in the federation update in this round, and deliver the updated federated global model to all clients, and enter the next round of federation process, Until the preset federation round is reached.

Compared with prior art, the beneficial effect of the present invention is:

The lightweight federated learning method and system for spatio-temporal data heterogeneous scenarios proposed by the present invention can assign different communication probabilities to the client according to the change of client data distribution, so that the data distribution changes greatly and is not affected by the federated global Model-aware clients participate in federated communication with a higher probability, rather than increasing the traffic of all clients equally. By slanting communication resources, it is possible to perceive global data distribution changes with less communication traffic, thereby improving the communication efficiency of federated learning in spatiotemporal heterogeneous scenarios.

By using the lightweight federated learning method and system oriented to heterogeneous spatiotemporal data scenarios proposed by the present invention, each image collection and analysis terminal uses the change information of its private case data distribution and the gradient information of the private detection model of the image collection and analysis terminal, in each Its importance is calculated during a round of federation. And according to the important measurement results of the image collection and analysis terminal, dynamically adjust the probability of each round of federation process image collection and analysis terminal participating in the federation communication. By allocating communication needs to more important image acquisition and analysis terminals, reducing the communication traffic of low-importance image acquisition and analysis terminals, achieving the same federated learning performance with less federated communication costs, and effectively reducing federated learning traffic.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings that are required in the embodiments. Obviously, the accompanying drawings in the following description are only described in the present invention For some embodiments of the present invention, those skilled in the art can also obtain other drawings according to these drawings.

Fig. 1 is a flow chart of a lightweight federated learning method for spatiotemporal data heterogeneous scenarios provided by an embodiment of the present invention.

Fig. 2 is an architecture diagram of a lightweight federated learning system oriented to spatio-temporal data heterogeneous scenarios provided by an embodiment of the present invention.

Detailed ways

In order to better understand the technical solution, the method of the present invention will be described in detail below in conjunction with the accompanying drawings.

In the scenario of spatio-temporal data heterogeneity, the present invention utilizes the distribution change degree of the client private data and the gradient information of the client model to dynamically adjust the uplink communication probability of the client during the federation process, reducing the communication overhead of low-importance clients . This effectively solves the communication efficiency bottleneck problem of the federated system in heterogeneous spatio-temporal data scenarios, and realizes a lightweight federated learning system.

In an environment with heterogeneous spatiotemporal data, the federated global model needs to continuously fit the changing data distribution. Due to data privacy protection, the fitting process of the distribution change needs to be completed by uploading the federated client model information. However, since the data distribution of each client changes with time and the timing is different, when fitting the global data distribution change, how to select the client participating in the communication will greatly affect the fitting effect and the required communication volume. In order to lightly realize the fitting of data distribution changes, the lightweight federated learning method and system for spatio-temporal data heterogeneous scenarios proposed by the present invention are shown in Fig. 1 and Fig. 2 . The system consists of four main modules, which are client local update module, client importance evaluation module, federated communication probability control module, and federated server aggregation module.

The local update module of the client first uses the private data locally on the client to change the federated global model into a private model. Then, the present invention introduces a client importance detection module oriented to the change degree of client distribution to quantify the contribution of each client communication to the global model distribution change perception in each round of federated learning.

First, if the client's local data distribution is stable, its upload communication cannot contribute to the fitting of global distribution changes. Therefore, the importance detection module uses the difference between the accuracy rate of the latest client local model on the current round of incremental samples and the accuracy rate on historical samples to reflect the distribution difference between the current round of local private incremental data and historical data. The measure is called the client absolute distribution change measure.

Secondly, since federated learning involves multiple clients, if a client's local distribution change information has been perceived by the federated global model through the communication process with other clients, the importance of the client's communication for the perception of distribution changes is more important than that of distribution Clients whose changes are not perceived by the federated global model are lower. Therefore, the cosine similarity between the gradient information not uploaded by the client and the gradient information of the global model corresponding to the time window is further introduced to reflect the relative change degree of the distribution change carried by the incremental data of the client and the perceived distribution of the global model. If the client's model change is in the same direction as the federated global model's change, that is, the similarity is high, the part of the client's incremental information that is not perceived by the global model is small, and the corresponding importance is low. Finally, by combining the absolute importance and the relative importance in a multiplicative manner as weights, the degree of distribution change of the client's private data and the similarity between the change of the client's private data and the perceived change of the global model are comprehensively considered. It reflects the importance of communicating with the client to the federated global model to perceive changes in data distribution.

On the basis of measuring the importance of clients, it is natural to allocate more federated communication requirements to clients that contribute more to the changes in the distribution of federated global model-perceived data, which can improve communication efficiency. Therefore, after the importance of the client is updated in each round of federated learning, the client communication control module adjusts the probability of the client uploading the model according to the importance of the client. By monotonically mapping client importance to federated communication probability, high-importance clients will have a higher probability of uploading models to participate in federated aggregation, and correspondingly reduce the model uploading of low-importance clients. Finally, the federation server aggregation module is responsible for updating the model gradient information uploaded by the client according to the probability to the historical federated global model, and sends the updated federated global model to the client.

Example 1

mission details:

The federated learning system contains 10 clients, and the federated learning process lasts for 10 rounds in the experiment. The simulation method of simulating the spatial heterogeneity of data randomly allocates 20% of the category-corresponding data to each client during the federated learning initialization process, and the incremental data of subsequent clients comes from the category allocation during the initialization process. The simulation method of client private data distribution changing over time is to change the data categories of 4 clients starting from the fifth round. The federation process of client 6 in round 5 is described in detail in the case:

Step 1: Client local update

其中i是客户端id属于[1,10]。Step 1.1: The client receives the federated global model ω ⁵ delivered by the server. Each client of the federated system uses local data to update the local model

where i is the client id belonging to [1,10].

Step 2: Client Importance Evaluation

测试在第5轮本地私有增量数据集

上的准确率0.8，与在历史2轮本地私有增量数据集的并集

与

上的准确率0.85，进一步，使用两数据集准确率的差异的绝对值abs(0.80.85)反应当前轮次数据集与历史2轮数据集的差异为0.05，即客户端6本地数据在第5轮数据分布的绝对变化度量值为0.05。Step 2.1: First, perform absolute gravity detection. Taking client 6 as an example, its local private data distribution changes in the fifth round, using the local model

Tested on round 5 local private incremental dataset

Accuracy rate 0.8 on , combined with local private incremental datasets in historical 2 rounds

and

The accuracy rate above is 0.85. Further, the absolute value abs(0.80.85) of the difference between the accuracy rates of the two data sets reflects that the difference between the current round data set and the historical 2 round data sets is 0.05, that is, the local data of client 6 is at The absolute measure of change in the data distribution over 5 rounds was 0.05.

记做

同样的，计算客户端上一次上传时点的联邦全局模型ω⁵与ω³的累计梯度ω⁵-ω³记做

进一步的，计算

与

的余弦相似度为0.4，相对重要性度量为1-0.4＝0.6。Step 2.2: Calculate the relative distribution change degree of client 6. The client model 6 uploaded the local model in the third round before the fifth round, so the cumulative gradient from the third round to the fifth round of the client is calculated as

remember to do

Similarly, calculate the cumulative gradient ω ⁵ -ω ³ of the federated global model ω ⁵ and ω ³ at the last upload time of the client.

Further, calculate

and

The cosine similarity of is 0.4, and the relative importance measure is 1-0.4=0.6.

Step 2.3: The importance of client 6 in round 5 is the product of relative importance and absolute importance, namely

Step 3: Client communication probability control

其中超参数tau设置为2，P_floor设置为5％。则客户端6在第5轮的通信概率为6％。Step 3.1: Based on the measurement results of the importance detection module of the client, in each round of the federation process, update the communication probability of each client j as

The hyperparameter tau is set to 2, and the P _floor is set to 5%. Then the communication probability of client 6 in the fifth round is 6%.

Step 3.2: Client 6 makes a random number of [0,1]. If the random value is less than 6%, it uploads the gradient information of the local model to the federation server, otherwise it does not participate in this round of uploading.

Step 4: Federated Server Aggregation

则新一轮联邦全局模型

其中η为联邦全局模型的更新学习率，在实验中设置为0.01。4.1 The federation server updates the client gradient information uploaded in this round to the federated global model, assuming that the client set C ∈ [6,7,9] uploaded in this round, remember the uploaded gradients of the client [6,7,9] respectively for

Then a new round of federal global model

where η is the update learning rate of the federated global model, which is set to 0.01 in the experiment.

4.2 The federation server sends the federated global model ω ⁶ to all clients and enters the next round of federation process.

It has been proved by experiments that in the scenario where 20% of the client distribution changes with time, the system of the present invention compares the baseline methods such as Fed-average and FedPNS, and can reach the Fed-average and FedPNS method with an average communication cost of 30% of client uploads. On average, 40% of the clients participate in the federation to upload the same performance, that is, the uplink communication overhead is reduced by 10% under the same performance. It effectively improves the efficiency of federated communication and realizes lightweight federated learning in spatiotemporal heterogeneous scenarios.

To sum up, the present invention proposes a lightweight federated learning system and method for spatio-temporal data heterogeneous scenarios. In IoT scenarios such as smart medical care, the client’s private data distribution change information is used to measure the importance of the client. In the federated process Dynamically adjust the communication probability of the client according to the importance, and reduce the communication overhead of the federation system. Each image collection and analysis terminal uses the change information of its private case data distribution and the gradient information of the private detection model of the image collection and analysis terminal, in each round of the federation process calculate its importance. And according to the important measurement results of the image collection and analysis terminal, dynamically adjust the probability of each round of federation process image collection and analysis terminal participating in the federation communication. By allocating communication needs to more important image acquisition and analysis terminals, reducing the communication traffic of low-importance image acquisition and analysis terminals, achieving the same federated learning performance with less federated communication costs, and effectively reducing federated learning traffic.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be described in the foregoing embodiments The recorded technical solutions are modified, or some of the technical features are equivalently replaced, but these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (6)

1. A lightweight federated learning method for space-time data heterogeneous scenes is characterized by comprising the following steps:

s1, local updating of a client: after receiving the latest federated global model, updating the federated global model into a localization model which is more consistent with local private data distribution by using local private data of the client;

s2, client importance evaluation: measuring the importance of the client participating in the federal communication on the global model of the federation by using the distribution change degree of private data and the consistency of the gradient information of the client model and the gradient information of the global model;

s3, controlling the communication probability of the client: dynamically adjusting the communication probability of the client according to the importance of the client obtained in the step S2, judging whether to upload according to the probability, if so, entering the step S4, and if not, issuing the federal global model to all the clients;

s4, federal service side aggregation: updating the global federal model by using the gradient information uploaded by the clients participating in the federal update in the current round, issuing the updated global federal model to all the clients, and entering the next federal process until the preset federal turn is reached.

2. The lightweight federal learning method for spatio-temporal data heterogeneous scenes as claimed in claim 1, wherein the specific process of the step S1 is as follows: local to the client, in the federal global model ω ^t On the basis, by a gradient descent iterative algorithm, using a private data set of the client i in t turns

Continue training to obtain the localModel (model)

The optimization objective function of the client is

Wherein n is

And f is a loss function of the client i.

3. The lightweight federal learning method for spatio-temporal data heterogeneous scenes as claimed in claim 1, wherein the specific process of step S2 is as follows:

s21, the model obtained after the local update of the t round of the client i is

First, a model is calculated

Top-1 accuracy on recent data fragmentation

Second calculation model

Accuracy on the last n rounds of data slicing

Finally, calculating the absolute value of the difference of the accuracy rates of the data at two different time ends on the same model as

Reflect the absolute distribution change of the local data of the client, and record it as

S22, caching the local model uploaded to the federal server last time by the client

Compare it with local model of the current round

Obtaining gradient information of the client end which is not uploaded

Comparing the received federal global model omega which is calculated by the client after last uploading _s With the latest federal global model omega _t Obtaining the perceptual gradient information of the global model of the federation

By calculating the cosine similarity of the gradient information of the client and the gradient information of the global federated model

The relative difference between the gradient to be uploaded at the client and the perceived gradient of the global federal model is measured and recorded as

S23, combining the absolute change AC and the relative change RC in a multiplication mode, and taking the combination as the importance measurement of the client i in the round t:

4. the lightweight federal learning method for spatio-temporal data heterogeneous scenes as claimed in claim 1, wherein the specific process of step S3 is:

s31, customerThe importance metric of the client is input into a communication probability mapping module, and the communication probability of the client i is updated to

The tau control federal system distributes communication resources for the time change of data distribution, and reflects the sensitivity degree of the distribution change; p is _floor The basic communication probability of the client is controlled, the uplink probability of the high-importance client is dynamically increased, and the uplink probability of the low-importance client is reduced;

and S32, the communication probability control module generates a client-side unique heat vector to be uploaded according to the probability in each round of federal process, wherein one dimension value is 1, and the other dimension values are 0, and the corresponding client side is required to upload the gradient information of the local model.

5. The lightweight federal learning method for spatio-temporal data heterogeneous scenes as claimed in claim 1, wherein the specific process of step S4 is:

s41, uploading the gradient accumulated gradient from the S moment to the t moment in the t round by the selected client i

S42, the client set uploaded in the current round is C ^t The historical federal global model is ω _t Updating the accumulated gradient information uploaded to the client to a federal global model by a learning rate eta

And S43, issuing the updated global federated model to all clients.

6. A lightweight federated learning system oriented to spatio-temporal data heterogeneous scenarios, comprising the following modules to implement the method of any one of claims 1-5:

the client local updating module is used for updating the federal global model into a localization model which is more in line with the distribution of local private data by using local private data of the client after receiving the latest federal global model;

the client importance evaluation module is used for measuring the importance of the client participating in the federal communication on the federal global model by utilizing the distribution change degree of private data and the consistency of the gradient information of the client model and the gradient information of the global model;

the client communication probability control module dynamically adjusts the client communication probability according to the client importance obtained by the client importance evaluation module, judges whether to upload according to the probability, enters the federal service end aggregation module if the client communication probability is uploaded, and issues the federal global model to all clients if the client communication probability control module does not upload the client communication probability;

and the federal server side aggregation module is used for updating the federal global model by using the gradient information uploaded by the clients participating in the federal update in the current round, issuing the updated federal global model to all the clients and entering the next federal process until the preset federal turn is reached.

Images