CN115115066A - Contrastive Learning-Based Federated Learning Personalization Approach - Google Patents

Abstract

The invention discloses a comparative learning-based federal learning personalization method, which realizes personalization by balancing a global model trained on a whole data set and a local model trained on a local subset. And correcting local updating through the consistency loss of the characteristics learned by the current local model and the characteristics learned by the global model, thereby realizing the individuation of federal learning. The characteristics of the global model, the characteristics of the model obtained by local data training and the characteristics of the local model trained by the current model are calculated at the same time, and then the distances among the three characteristics are normalized through personalized parameters, namely, comparative learning is realized on a model level, so that the global model is corrected to adapt to tasks on different clients. The method and the system can adapt to different tasks of the client, and meet the requirement of federal learning personalized customization.

Description

Contrastive Learning-Based Federated Learning Personalization Approach

technical field

The invention belongs to the field of personalized federation learning model, and in particular relates to a federated learning personalization method based on contrastive learning.

Background technique

Federated learning arises because of the problem of data silos, and its goal is to collaboratively train data that is generated by many remote devices or local clients and cannot be shared due to privacy concerns. Federated learning enables multiple parties to jointly learn a machine learning model without exchanging their respective local data. In each round of federated learning training, the updated local models of each party are transmitted to the server, which further aggregates the local models to update the global model. Raw data is not exchanged during the learning process. Federated learning has become an important field of machine learning, attracting a lot of research interest. In addition, it is also used in many fields, such as medical imaging, smart furniture, etc.

Since federated learning focuses on learning all client data to obtain a high-quality global model, it is impossible to capture the personal information of all clients. For clients with high data quality and high contribution, the global model can be well adapted to the client's task, but for clients with low contribution, the global model may not have good adaptability . When each client updates its local model, its local target may be far from the global target. Therefore, federated learning faces the problem of heterogeneous data distribution in real-world deployment, and the global model obtained by training may not be suitable for working on all clients participating in the training.

A key challenge of federated learning is dealing with the heterogeneity of local data distribution across parties. The global model trained by federated learning often ignores the personal information of low-contribution clients, resulting in a low generalization ability of the global model. To address this challenge, a simple and effective approach is to personalize the device or model. For the global model obtained by broadcasting, the client performs corresponding personalized training to ensure that the final model can adapt to the task.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a personalized method for federated learning based on contrastive learning in view of the deficiencies of the prior art. The present invention can ensure that the global model can be revised according to the data distribution of the client.

The object of the present invention is achieved through the following technical solutions: a method for personalized federated learning based on contrastive learning, comprising the following steps:

(1) Train an independent model locally:

(1.1) The local client trains an independent model according to the data it holds, and the independent model only contains the data information of a single client and adapts to the task of the client;

(1.2) Build a contrastive learning network, and use the contrastive learning network to extract representations from independent models;

(2) The client participates in federated learning training:

(2.1) The client starts normal federated learning training, and gets local updates according to the global model and local data published by the server;

(2.2) After the training is completed, each client uploads the locally updated local model to the server;

(2.3) The server aggregates to obtain a new global model and broadcasts it globally, and the client extracts the global representation from the new global model after receiving it;

(3) The client performs personalized correction training:

(3.1) After the client receives the latest global model, the client starts personalized correction training, and obtains three representations in the training phase: independent model representation, global model representation and the local model representation being updated;

(3.2) Calculate the model consistency loss through the three representations obtained in step (3.1), and update the local model together with the training loss;

(3.3) Repeat steps (3.1) to (3.2), perform iterative correction training, and finally obtain a personalized local model.

Further, step (1.1) includes:

There are a total of N clients, denoted as P ₁ ,...,P _N , and the client P _i has a local data set D ⁱ . For local individual training, the training objectives of the independent model are as follows:

为P_i的经验损失；

表示客户端P_i的独立模型权重；x表示Dⁱ中的一个训练数据，y表示x对应的标签；l_i表示客户端P_i本地的损失函数；E表示期望。in,

is the experience loss of _Pi ;

represents the independent model weight of the client _Pi ; x represents a training data in ^Di , y represents the label corresponding to x; _li represents the local loss function of the client _Pi ; E represents the expectation.

Further, step (1.2) includes:

用

表示独立模型输出层之前的网络；将训练完成的独立模型权重放入对比学习网络中，便得到独立模型的表征

The independent model weights are

use

Represents the network before the output layer of the independent model; put the weights of the independent model after training into the comparative learning network to obtain the representation of the independent model

Further, step (2.1) includes:

For the local client, the training objectives of the federated learning local model are as follows:

为P_i的第t轮的本地经验损失L_i,t；w_i表示客户端P_i的本地模型权重；x表示Dⁱ中的一个训练数据，y表示x对应的标签；l_i,t表示客户端P_i第t轮本地的损失函数；E表示期望。in,

is the local experience loss _Li,t of the t-th round of _Pi ; _wi represents the local model weight of the client _Pi ; x represents a training data in ^Di , y represents the label corresponding to x; li _,t represents The loss function local to the t-th round of the client P _i ; E represents the expectation.

Further, step (2.3) includes:

The server receives the client's model update and uses the FedAvg aggregation algorithm to form a new global model; FedAvg calculates the average of the client's local model updates as the global model update, where each client is weighted according to the number of its training examples; using D = ∪ _i∈[N] D ⁱ represents all federated training datasets, and the training objective of the global model in round t is expressed as:

Among them, L _i,t represents the empirical loss of the t- _th round defined by Pi; L _g,t represents the loss of the global model in the t-th round; w _g is the weight of the global model; || represents the modulo;

后，服务器将全局模型

进行广播，发送至各个客户端；客户端根据最新的全局模型计算得到全局模型的表征

Get the new global model after the aggregation is complete

After the server will global model

Broadcast and send to each client; the client calculates the representation of the global model according to the latest global model

Further, step (3.1) includes:

计算，第二部分为模型对比损失函数计算，模型对比损失函数定义为

对于每一个输入x以及本地模型

从独立模型中提取出表征

从全局模型中提取出表征

以及从正在训练的本地模型提取出的表征

则模型对比损耗函数定义为：The loss of personalized correction training consists of two parts, the first part is composed of the loss function

Calculation, the second part is the model comparison loss function calculation, the model comparison loss function is defined as

for each input x and the local model

Extract representations from independent models

Extract representations from global models

and representations extracted from the local model being trained

Then the model contrast loss function is defined as:

Further, in step (3.1):

The server decides whether to perform personalized correction training for the corresponding client according to whether it receives the personalized application from the client. When the server receives the personalized application from the client, it determines that the client needs to undergo personalized correction training, and then the client starts personalized training. Correct the training; if the personalized application from the client is not received, jump to step (2.1) for the next round of federated learning training.

Further, step (3.2) includes:

The goals of the updated local model are:

表示独立模型的权重，x表示Dⁱ中的一个训练数据，y表示x对应的标签。Among them, μ controls the personalized hyperparameter of the model contrast loss weight, which is set by the client; w _g represents the weight of the global model, w _i represents the weight of the local model,

represents the weight of the independent model, x represents a training data in ^Di , and y represents the label corresponding to x.

The beneficial effects of the present invention are as follows:

(1) The revised training of the present invention is independent of the normal federated training, and the individualization process will not affect the global model, that is, the convergence process of the global model will not be affected;

(2) The present invention can efficiently map the features through the comparative learning and calculation representation at the model level;

(3) The present invention can control the degree of personalization of the model through consistency loss, and can perform fine-grained customization.

Description of drawings

FIG. 1 is a schematic diagram of a federated learning personalization method based on contrastive learning of the method of the present invention.

FIG. 2 is a specific flow chart of the method of the present invention for the individualized method of federated learning based on contrastive learning.

Detailed ways

The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

The present invention achieves personalization by weighing a global model trained on the entire dataset versus a local model trained on a local subset. Based on this point of view, the present invention is a method for personalized federated learning based on contrastive learning, which corrects the local update through the loss of consistency between the representation learned by the current local model and the representation learned by the global model, thereby realizing the personalization of federated learning. . Specifically, the method will simultaneously calculate the representation of the global model, the representation of the model trained on local data, and the representation of the local model trained by the current model, and then normalize the distance between these three representations through personalized parameters, that is, in Contrastive learning is implemented at the model level to modify the global model to adapt to tasks on different clients.

As shown in Figure 1, the present invention specifically comprises the following steps:

(1) An independent model is trained locally.

Each client in federated learning has different training data, which is stored locally and cannot be transmitted. Before starting federated learning training, the present invention requires each client to use local data to train an independent model as one of the subsequent personalized representations. Due to the heterogeneity of data between clients, there may be cases where clients lack data and cannot train a good independent model. However, this does not affect subsequent personalized customization, because the goal of personalized customization is to adapt the global model to a specific task, and the independent model plays a guiding role.

Therefore, step (1) includes the following sub-steps:

(1.1) The local client trains an independent model according to the data it holds. The weight of the independent model is expressed as _wind . The independent model only contains the data information of a single client and adapts to the task of the client.

Suppose there are N clients in federated learning, denoted as P ₁ ,...,P _N , and the client P _i owns the local dataset D ⁱ . For local individual training, the training objectives of the independent model are as follows:

为P_i的经验损失。

表示客户端P_i的独立模型权重；x表示Dⁱ中的一个训练数据，y表示x对应的标签；l_i表示客户端P_i本地常规的损失函数；E表示期望。in,

is the experience loss of _Pi .

represents the independent model weight of the client _Pi ; x represents a training data in ^Di , y represents the label corresponding to x; _li represents the local conventional loss function of the client _Pi ; E represents the expectation.

(1.2) Build a contrastive learning network and use the contrastive learning network to extract representations from independent models.

用

表示独立模型整个网络，用

表示独立模型输出层之前的网络。将训练完成的独立模型权重放入对比学习网络中，便可得到独立模型的表征

其中输入x表示一个训练数据。Similar to the typical contrastive learning framework SimCLR, this contrastive learning network has three components: the base encoder, the neural network projection head, and the output layer. The base encoder is used to extract the representation vector from the input, and the neural network projection head is used to map the representation vector to a space with fixed dimensionality. Finally, the output layer will be used to produce predictions for each class. Similar to the above, assume that the independent model weights are

use

represents the entire network of independent models, using

Represents the network before the output layer of the independent model. Put the weights of the trained independent models into the comparative learning network to obtain the representation of the independent models

where input x represents a training data.

(2) The client participates in federated learning training.

After the independent model is trained locally and the representation is obtained, the formal federated training process will begin. In order to ensure that the global model can be converged, the present invention separates the training of the global model from the correction training of the individualized model, and maintains two tasks for each other. That is, before the client performs personalized training, a report needs to be sent to the server. After receiving the personalized application, the server will mark the client to ensure that the client does not participate in the global model aggregation. Similarly, the client will only perform correction training after the round of federated learning training is completed, so as to ensure that the client's personalization will not affect the global model operation. Therefore, in step (2), only the normal federated learning training process is involved.

Therefore, step (2) includes the following substeps:

(2.1) The client starts the normal federated learning training, and gets local updates according to the global model and local data published by the server.

The server initializes the global model and broadcasts the global model to various clients. The client receives the global model, utilizes the local data and performs one or more iterations using stochastic gradient descent to update the local model.

For the local client, the training objectives of the federated learning local model are as follows:

为P_i的第t轮的本地经验损失L_i,t。w_i表示客户端P_i的本地模型权重；x表示Dⁱ中的一个训练数据，y表示x对应的标签；l_i,t表示客户端P_i第t轮本地常规的损失函数；E表示期望。本地训练完成后，客户端将本地模型更新上传给服务器。in,

is the local experience loss _Li,t for the t- _th round of Pi. w _i represents the local model weight of the client _Pi ; x represents a training data in Di, y represents the label corresponding to x; l ⁱ _{, t} represents the local conventional loss function of the t- _th round of the client Pi; E represents the expectation . After the local training is completed, the client uploads the local model update to the server.

(2.2) After training is completed, each client sends its local model and uploads local updates to the server to ensure that the server can receive local updates without personalization.

(2.3) The server aggregates the new global model and broadcasts it globally, and the client extracts the global representation from the new global model after receiving it.

The server receives the model update from the client and uses the FedAvg aggregation algorithm to form a new global model. FedAvg computes the average of the client's local model updates as the global model update, where each client is weighted according to the number of its training examples. Using D = ∪ _i∈[N] D ⁱ to denote all federated training datasets, the training objective of the global model at round t can be expressed as:

Among them, L _i,t represents the empirical loss of the t- _th round defined by Pi; L _g,t represents the loss of the global model; w _g is the weight of the global model; || represents the modulo.

后，服务器将全局模型

进行广播，发送至各个客户端。客户端根据最新的全局模型可以计算得到全局模型的表征

其中x表示一个训练数据。Get the new global model after the aggregation is complete

After the server will global model

Broadcast and send to each client. The client can calculate the representation of the global model according to the latest global model

where x represents a training data.

(3) The client performs personalized correction training.

After the client obtains the latest global model, it will perform personalized correction training of the local model. The goal of personalized correction training is to correct the distance between local model representations and global model representations as well as independent model representations. For example, if the client wants the local model to adapt to the local task, the goal of its correction training is to reduce the distance between the representation learned by the local model and the representation learned by the global model, and increase the representation learned by the local model and the independent model. the distance between the representations.

Therefore, as shown in Figure 2, step (3) includes the following sub-steps:

(3.1) The client receives the latest global model and sends an application for personalized training to the server. The server decides whether to perform personalized correction training for the corresponding client according to whether it receives the personalized application from the client. When the server receives the personalized application from the client, it determines that the client needs to undergo personalized correction training, and then the client starts personalized training. Corrected training. If the personalized application from the client is not received, jump to step (2.1) for the next round of federated learning training. Among them, three representations are obtained in the personalized correction training stage: independent model representation, global model representation, and local model representation being updated.

计算，第二部分为模型对比损失函数计算，模型对比损失函数定义为

对于每一个输入x以及本地模型

可以从独立模型中提取出表征

从全局模型中提取出表征

以及从正在训练的本地模型提取出的表征

则模型对比损耗函数可以定义为：The loss for correction training consists of two parts, the first part is composed of the regular loss function

Calculation, the second part is the model comparison loss function calculation, the model comparison loss function is defined as

for each input x and the local model

Representations can be extracted from independent models

Extract representations from global models

and representations extracted from the local model being trained

Then the model contrast loss function can be defined as:

(3.2) The model consistency loss is calculated through three representations, and the local model is updated together with the traditional training loss. The goal is to minimize the following formula:

表示独立模型的权重，x表示Dⁱ中的一个训练数据，y表示x对应的标签。Among them, μ controls the personalized hyperparameter of the model contrast loss weight, and controls the degree of personalization of the local model, which is set by the client. w _g represents the weight of the global model, w _i represents the weight of the local model,

represents the weight of the independent model, x represents a training data in ^Di , and y represents the label corresponding to x.

(3.3) Repeat steps (3.1) to (3.2), and correct the training through multiple iterations until the overall loss update value of the corrected training is less than 0.0001 or when the preset number of times is reached, and the training is ended, and finally a personalized federated learning local model is obtained.

It should be noted that the correction training in the present invention is separated from the federated learning training, and the update of the global model is the same as the traditional federated learning method, so it will not interfere with the normal federated learning training. For the local model, whether to perform revision training does not affect the update of the global model, but only affects the degree of personalization of the local model.

In an embodiment of the present invention, compared with the federated average learning algorithm in the experiment on the CIFAR-10 decadal data set, the test accuracy of the model corrected by the present invention is improved by an average of 3.1% (±0.4%), where μ is set as 5.

The goal of federated learning is to learn all the data features involved in training. However, due to the existence of data heterogeneity, some relatively hidden data features on the client are often ignored, resulting in differences in the performance of the global model on each client. In order to solve this problem, the present invention is a method for personalized federated learning based on contrastive learning. Through the contrastive learning technology, the step of correction training is added after normal local training. By calculating the representation of the independent model, the global model and the local model being updated, the local model is modified to ensure that it can adapt to different tasks of the client and meet the requirements of personalized customization of federated learning.

The content described in the embodiments of the present specification is only an enumeration of the realization forms of the inventive concept, and the protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments, and the protection scope of the present invention also extends to those skilled in the art. Equivalent technical means that can be conceived by a person based on the inventive concept.

Claims (8)

1. A federal learning individualization method based on comparative learning is characterized by comprising the following steps:

(1) training out an independent model locally:

(1.1) training an independent model by the local client according to data held by the local client, wherein the independent model only contains data information of a single client and adapts to tasks of the client;

(1.2) building a contrast learning network, and extracting the representation from the independent model by using the contrast learning network;

(2) the client participates in the federal learning training:

(2.1) the client starts normal federal learning training and obtains local update according to a global model and local data issued by the server;

(2.2) after the training is finished, each client uploads the locally updated local model to the server;

(2.3) aggregating the server to obtain a new global model and carrying out global broadcasting, and extracting global representations from the new global model after the client receives the new global model;

(3) the client carries out personalized correction training:

(3.1) after the client receives the latest global model, the client starts personalized correction training, and three representations are obtained in a training stage respectively: an independent model representation, a global model representation and a local model representation being updated;

(3.2) calculating to obtain model consistency loss through the three representations obtained in the step (3.1), and updating the local model together with the training loss;

and (3.3) repeating the steps (3.1) to (3.2), and performing iterative correction training to finally obtain the personalized local model.

2. The comparative learning-based federal learning personalization method of claim 1, wherein step (1.1) comprises:

there are N clients, denoted P ₁ ,...,P _N Client P _i Having local data sets D ⁱ For local individual training, the training objectives of the independent model are as follows:

wherein,

is P _i (ii) a loss of experience;

representing a client P _i The independent model weights of (a); x represents D ⁱ Y represents the label to which x corresponds; l _i Representing a client P _i A local penalty function; e represents expectation.

3. The comparative learning-based federal learning personalization method of claim 1, wherein step (1.2) comprises:

the independent model weight is

By using

Representing the network before the independent model output layer; putting the trained independent model weight into a comparative learning network to obtain the representation of the independent model

4. The comparative learning-based federal learning personalization method of claim 1, wherein step (2.1) comprises:

for the local client, the training objectives of the federated learning local model are as follows:

wherein,

is P _i Local experience loss L of the t-th round _i,t ；w _i Representing a client P _i Local model weight of (2); x represents D ⁱ Y represents the label to which x corresponds; l _i,t Representing a client P _i Local loss function of the t round; e represents expectation.

5. The comparative learning-based federal learning personalization method of claim 1, wherein step (2.3) comprises:

the server receives the model update of the client and forms a new global model by adopting a FedAvg aggregation algorithm; the FedAvg calculates the average of the local model updates of the clients as global model updates, wherein each client is weighted according to the number of training examples thereof; using D ═ U _i∈[N] D ⁱ Representing all the data sets of federal training, the training targets of the t-th round global model are represented as:

wherein L is _i,t Represents P _i Defining the experience loss of the t-th round; l is a radical of an alcohol _g,t Represents the loss of the global model for the t-th round; w is a _g Is the weight of the global model; | represents modulo;

to obtain new compounds at the completion of the polymerizationOffice model

The server then maps the global model

Broadcasting and sending the broadcast to each client; the client calculates the representation of the global model according to the latest global model

6. The comparative learning-based federal learning personalization method of claim 1, wherein step (3.1) comprises:

the loss of the personalized correction training is composed of two parts, wherein the first part is a loss function

The second part is the calculation of a model contrast loss function defined as

For each input x and the local model

Extracting tokens from independent models

Extracting tokens from a global model

And tokens extracted from the local model being trained

Then model contrast loss functionThe number is defined as:

7. the comparative learning-based federal learning personalization method of claim 1, wherein in step (3.1):

the server determines whether the corresponding client performs the personalized correction training according to whether the personalized application of the client is received or not, judges that the client needs to perform the personalized correction training after the server receives the personalized application of the client, and then starts the personalized correction training; and if the client-side personalized application is not received, skipping to the step (2.1) to perform the next round of federal learning training.

8. The comparative learning-based federal learning personalization method of claim 1, wherein step (3.2) comprises:

the goal of the updated local model is:

wherein, the mu control model compares the individualized hyperparameter of the loss weight, and is set by the client; w is a _g Weights, w, representing the global model _i The weights of the local model are represented by,

weights representing independent models, x representing D ⁱ Y represents the label to which x corresponds.

Images