Federal learning privacy protection method based on multi-stream network feature separation

Abstract

The invention relates to a federated learning privacy protection method based on multi-stream network feature separation. The method comprises the following steps: S1, transmitting a feature separator and a head of a server side to a client based on the server side; s2, performing training iteration on the feature separator and the head of the server side by using the client side; s3, uploading the feature separator after iteration and the head of the server side to the server side; s4, aggregating each part based on the server side to form a new feature separator and a new head; and S5, continuously iterating S1 to S4 until the training of the feature separator and the head is completed. The invention aims to train lightweight global and fine-grained flow features on the premise of not uploading original data of a client, and after feature separation, the last layer of a full-connection layer is used as the head of a network to perform an image classification task so as to achieve a better image recognition effect.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence privacy protection, and particularly to a privacy protection method for federated learning based on multi-stream network feature separation. Background Art

[0002] With the rapid development of artificial intelligence and big data technologies, federated learning (FL), as an emerging distributed learning framework, has gradually become a research hotspot in the field of privacy-preserving machine learning. Federated learning allows multiple participants (clients) to collaboratively train a global model while protecting local data privacy, without the need to upload local data to a central server. This decentralized learning mode is widely applied in fields such as healthcare, finance, and the Internet of Things. However, due to significant differences in data distribution and quality among clients, federated learning faces great challenges in data heterogeneity.

[0003] Data heterogeneity usually manifests in the following aspects: 1. inconsistent feature distributions, 2. unbalanced label distributions, 3. imbalanced sample quantities, 4. data quality differences; specifically, for example, in Internet of Things applications, the sensor data collected by different devices may have different feature ranges and distributions, and this inconsistency in feature distributions can lead to insufficient generalization ability of the global model on some clients; in the healthcare field, the case data of different hospitals usually focuses on specific diseases, which can result in the global model lacking robustness to rare classes. Large enterprises have a large amount of high-quality labeled data, while small enterprises have limited data volume; this imbalance in data scale can lead to the training process of the model being dominated by clients with a larger data volume, ignoring the contributions of clients with a smaller data volume but important information.

[0004] To address the above problems, in recent years, personalized federated learning (PFL), as an extended technology of federated learning, has received extensive attention. Different from traditional federated learning that constructs a single global model, personalized federated learning aims to generate a personalized model adapted to the local data distribution for each client while satisfying global collaboration. By combining global information with local characteristics, personalized federated learning has shown significant advantages in data heterogeneity scenarios.

[0005] Currently, the personalized federated learning technologies used to address data heterogeneity mainly include the following:

[0006] 1. Method based on model decomposition: The model is divided into a shared layer and a personalized layer. Cross-client collaboration is achieved through the shared layer, and adaptation to the local data distribution is achieved through the personalized layer.

[0007] 2. Meta - learning - based methods: Utilize meta - learning techniques to quickly adapt to the local distribution of clients, for example, generate personalized models through a small number of gradient updates.

[0008] 3. Multi - task learning - based methods: Model the federated learning problem as a multi - task learning problem to optimize the personalized performance of each client's model and the global collaboration effect.

[0009] Although these methods alleviate the impact of data heterogeneity on federated learning to a certain extent, there are still some problems, such as the trade - off between model personalization and global sharing, and the computational and communication overhead in multi - client collaboration. Summary of the Invention

[0010] The purpose of the present invention is to provide a privacy - protection method for federated learning based on multi - stream network feature separation to address the deficiencies in the prior art. The aim is to train lightweight global and fine - grained flow features without uploading the original client data, and use the last layer of the fully - connected layer as the head of the network after feature separation for image classification tasks to achieve better image recognition effects.

[0011] To achieve the above - mentioned purpose, the present invention provides the following solutions:

[0012] A privacy - protection method for federated learning based on multi - stream network feature separation, comprising:

[0013] S1. Transmit a feature separator and a head from the server - side to the client; wherein, the feature separator includes: a global feature separator and a fine - grained feature separator, and the head includes: a global feature head and a fine - grained feature head;

[0014] S2. Use the client to perform training iterations on the feature separator and the head;

[0015] S3. Upload the iterated feature separator and head to the server - side;

[0016] S4. Aggregate the iterated feature separator and head based on the server - side to form a new feature separator and head;

[0017] S5. Continuously iterate S1 - S4 until the training of the feature separator and the head is completed; perform federated learning privacy protection based on the trained feature separator and head.

[0018] Optionally, transmitting a feature separator and a head from the server - side to the client includes:

[0019] Divide the backbone of the client into a feature separator and a head;

[0020] Initialize the model parameters on the server side and distribute them to each client participating in the training;

[0021] In each iteration, the server randomly selects a preset part of the clients to participate in the training.

[0022] Optionally, the backbone of the client is divided into a feature separator including:

[0023] For the global flow feature separator, map the input sample to the feature space, and the head maps the extracted features to the label space;

[0024] For the fine-grained flow feature separator, map the input sample to the feature space, and the head maps the features to the label space.

[0025] Optionally, use the client to perform training iterations on the feature separator and the head on the server side, including:

[0026] Select a dataset and divide the dataset into a training set and a test set;

[0027] Based on a preset loss function, use the training set to train the feature separator and the head, and use the test set for testing; among them, the global feature separator is used for global flow feature separation, and the fine-grained feature separator is used for fine-grained flow feature separation.

[0028] Optionally, the global feature separator includes: encryptor M enc and decoder M dec ;

[0029] The global feature separator performs global flow feature separation including:

[0030] Given an input feature map L = H×W is the number of elements, D is the dimension of the input feature, through the global flow feature model parameters Initialize as the intermediate layer of the global flow feature, D′ is the initialized feature dimension, that is Through linear layer W enc and linear layer W dec Randomly initialize to obtain the encryptor That is: The decryptor K is the feature dimension of the decryptor, that is: M enc Encode the input feature information to generate a unique representation for each input, that is, generate the encrypted information, so that the input data cannot be recognized before being matched;

[0031] The encrypted information is as follows:

[0032]

[0033] Among them, Norm(·) is standard normalization;

[0034] Using the encrypted information and M dec Perform matrix multiplication to generate global flow features after the decryptor

[0035] Optionally, the fine-grained feature separator is obtained by improving VGG;

[0036] Improving VGG includes: simplifying the middle network layer of VGG into three branches, where two branches are respectively equipped with a 1×1 convolutional layer and a 3×3 convolutional layer, and only batch normalization operation is performed after each convolution. The other branch performs average pooling operations on the input in the H and W directions respectively to extract pixel-level features. At the same time, a selective attention mechanism is introduced in this branch to enable the network to suppress irrelevant features and strengthen key information;

[0037] The selective attention mechanism is as follows:

[0038]

[0039]

[0040]

[0041] Among them, f n*n (·) represents an n*n convolution, Avgpooling(·) represents an average pooling operation, respectively represent Flow One, Flow Two, and Flow Three; the total feature of the fine-grained flow is:

[0042]

[0043] Optionally, the head is as follows:

[0044]

[0045]

[0046] Among them, respectively represent the global head and the fine-grained head parameterized by , is the logical value of the global flow feature, is the logical value of the fine-grained flow feature.

[0047] Optionally, the preset loss function is:

[0048]

[0049] Among them, η is a hyperparameter, respectively representing the losses of the global flow feature separator and the fine-grained flow feature separator, W i represents the model parameters, x i represents the input of the i-th client, y i represents the corresponding label.

[0050] Optionally, the loss of the global flow feature separator is:

[0051]

[0052] Among them, α is the weight factor of the category, γ is the focusing parameter, represents the logical value of the c-th category of the m-th sample in the global flow head from the i-th client, N j represents the number of the j-th category, N m,c represents the number of the current sample m belonging to the category c;

[0053] The loss of the fine-grained flow feature separator is set as the cross-entropy loss function

[0054] Optionally, aggregating each part based on the server side includes:

[0055] The server side aggregates the two separators and their respective heads received and performs weighted averaging. The formula is as follows:

[0056]

[0057] Among them, is the loss of the i-th client, is the weight of the i-th client.

[0058] The beneficial effects of the present invention are:

[0059] The present invention first transmits a feature separator and the head of the server side to the client based on the server side; secondly, uses the client to perform training iteration on the feature separator and the head of the server side; then uploads the iterated feature separator and the head of the server side to the server side; again, aggregates each part based on the server side to form a new feature separator and head; finally, continuously iterates the above steps until the training of the feature separator and the head is completed. The purpose of the present invention is to train lightweight global and fine-grained flows under the premise of not uploading the original data of the client, and use the last layer of the fully connected layer as the head of the network after feature separation to perform image classification tasks, achieving a better image recognition effect. Description of the Drawings

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0061] Figure 1 Schematic diagram of a framework for a federated learning privacy protection method based on multi-stream network feature separation according to an embodiment of the present invention;

[0062] Figure 2 Schematic diagram of the process of a federated learning privacy protection method based on multi-stream network feature separation according to an embodiment of the present invention;

[0063] Figure 3 Comparison chart of the accuracy of the method according to the embodiment of the present invention and other advanced federated learning algorithms. Detailed implementation manners

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0065] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the drawings and specific implementation manners.

[0066] As Figure 1 - Figure 2 shown, this embodiment proposes a federated learning privacy protection method based on multi-stream network feature separation, including:

[0067] Step 1, the server broadcasts the global and fine-grained feature separators, and their respective heads (the last fully connected layer of the network) to the clients;

[0068] Step 2, each client updates the global and fine-grained feature separators and the global and fine-grained feature heads through local training iterations;

[0069] Step 3, the client uploads the locally updated multi-stream network feature separators and heads to the server;

[0070] Step 4, the server aggregates each part to form new global feature separators, fine-grained feature separators, global heads, and fine-grained heads, and continuously iterates the process of Steps 1-4 to complete the training.

[0071] Perform federated learning privacy protection based on the trained feature separator and head. Feature separator: responsible for extracting features from the original data, which are usually low-dimensional representations with sensitive information removed; Head: used for classifying tasks; With this separation, the client only needs to upload features instead of the original data, reducing the risk of sensitive information leakage.

[0072] Furthermore, based on the server transmitting the feature separator to the client, and the head on the server side includes:

[0073] Divide the backbone of the client into a feature separator and a head;

[0074] Initialize the model parameters on the server side and distribute them to each client participating in the training;

[0075] In each round of iteration, the server randomly selects a preset part of the clients to participate in the training.

[0076] Specifically, in this embodiment, step 1 further includes the following content:

[0077] Step 1-1, divide the backbone of the client into a feature separator g(·;·), p(·;·) and a head h g (·;·), h p (·;·), different from this, this embodiment designs an MFPN module, and at the same time separates the global flow features of the client and the personalized features including the fine-grained flow features of the client respectively; Specifically, for the global flow feature separator parameterized by W g In this embodiment, the input sample is mapped to the feature space, that is by W gh The head parameterized by maps the extracted features to the label space, that is For the fine-grained flow feature separator parameterized by W p Similarly, the input sample is mapped to the feature space, that is by W ph The head parameterized by maps the features to the label space, that is D, K, and C represent the dimensions of the input, feature space, and label respectively.

[0078] Step 1-2, initialization: Initialize the model parameters on the server side and distribute them to each client participating in the training.

[0079] Step 1-3, client selection: In each round of iteration, the server randomly selects a part of the clients S t to participate in the training.

[0080] Further, the training iteration of the feature separator and the server - side header using the client includes:

[0081] Select a data set and divide the data set into a training set and a test set;

[0082] Based on a preset loss function, use the training set to train the feature separator and the header, and use the test set for testing; among them, the global feature separator is used for global flow feature separation, and the fine - grained feature separator is used for fine - grained flow feature separation.

[0083] Specifically, in this embodiment, step 2 includes the following steps:

[0084] Step 2 - 1, select a data set.

[0085] Step 2 - 2, divide the data set into a training set and a test set.

[0086] To better simulate the degree of heterogeneity of data among different clients, this embodiment uses the Dirichlet distribution function with parameter α to divide the training data set. Specifically, parameter α controls the distribution diversity of data for each client: when α is small, the data distribution is more uneven, and the difference in data categories on each client is greater; when α is large, the data distribution is more uniform, and the difference in data categories on each client is smaller. For each client k, the ten - category data distribution vector p k =[p k0 ,p k1 ,...,p k9 is generated by the Dirichlet distribution Dir(α) with parameter α. The specific formula is as follows:

[0087] p k ~Dir(α)

[0088] where α is the distribution parameter for each category, and p k is the generated data distribution vector.

[0089] Step 2 - 3, in the MFPN module designed in this embodiment, design different feature separators for feature separation. Specifically:

[0090] (1) Global flow feature separation

[0091] Considering the computing performance of edge devices and the fact that global features dominate the features of a single client, this embodiment proposes a model independent of the input in the global feature separator: the encryptor M enc and the decoder M dec .

[0092] Given the input feature map L = H × W is the number of elements, and D is the dimension of the feature; in this embodiment, the linear layer W i g Initialization That is Through the linear layer W enc 、W dec The encryptor is randomly initialized That is: The decryptor That is: M enc Encode the input feature information to generate a unique representation for each input, making the input data unrecognizable before being matched by W i g At the same time, M enc Can change according to the context, flexibly select different information, ensure that only correctly matched information can be encrypted and used, reducing the risk of information leakage. The encrypted information is defined as:

[0093]

[0094] Norm(·) is standard normalization; M dec As the decryptor, it is dynamically changing. Each time F i Changes, it will trigger different decryption forms, enabling M dec To provide different information, and only when F i 、M enc Successfully match, M dec Will be activated to ensure that the final output is consistent with the current Reducing information interference and distortion.

[0095] Finally, use the encrypted information to perform matrix multiplication with M dec To generate after the decryptor

[0096]

[0097] (2) Fine-grained flow feature separation

[0098] Fine-grained flow features are crucial for accurately identifying images and obtaining the information required by the images. Considering the requirements of edge devices for low latency and low overhead, and the fact that the 5×5 convolutional layer greatly increases the communication overhead, this embodiment uses a 1×1 convolutional layer and a 3×3 convolutional layer multi-stream network to process the features at different scales, avoiding more sequential processing and greater depth, and at the same time fusing the features from different streams in the later stage of the network.

[0099] In this embodiment, three branches are mainly used for extraction, corresponding to Flow 1, Flow 2, and Flow 3 respectively. Among them, Flow 1 and Flow 2 use a 1*1 convolutional layer and a 3*3 convolutional layer for operations. Another branch, Flow 3, performs average pooling operations on the input in the H and W directions respectively to extract pixel-level features, and at the same time introduces a selective attention mechanism.

[0100] In this embodiment, VGG is improved and redesigned to be more suitable for edge devices. Specifically, in this embodiment, two of the middle network layers of VGG are respectively equipped with a 1×1 convolutional layer and a 3×3 convolutional layer. After each convolution, only batch normalization operation is performed to extract effective fine-grained features while reducing the number of parameters. For another branch, in this embodiment, average pooling operations are performed on the input in the H and W directions respectively to extract pixel-level features. At the same time, in this branch, a selective attention mechanism (sigmoid) is introduced to enable the network to suppress irrelevant features and strengthen key information. The proposed expression is as follows:

[0101]

[0102]

[0103]

[0104] f n*n (·) represents an n*n convolution, and Avgpooling(·) represents an average pooling operation. They respectively represent Flow 1, Flow 2, and Flow 3. Therefore, the total features of the fine-grained flow are

[0105]

[0106] In step 2 - 4, for two different separators, in this embodiment, the last fully connected layer of the network layer is connected to the features separated by the separator. The last layer is called the head of the network in this embodiment:

[0107]

[0108]

[0109] h g (·; ·), h p (·; ·) respectively represent the global head and the fine-grained head parameterized by W i gh 、W i ph parameters, Y i g is the logit (logical value) of the global flow feature, Y ip The logit (logical value) of the fine-grained flow feature.

[0110] Step 2-5. Considering that the cross-entropy loss may be more inclined to the categories with a majority quantity and ignores the categories with a minority quantity, in this embodiment, the Focal loss is used for the branch of the global flow feature and is improved. The number of samples of different categories is balanced in the loss calculation, so that the model pays more attention to rare categories and improves the performance ability for unbalanced data. The formula proposed in this embodiment is as follows:

[0111]

[0112] Among them, α is the weight factor of the category, and γ is the focusing parameter, which is used to adjust the attention to the samples with a smaller category quantity; represents the logits (logical values) of the m-th sample of the c-th category in the global flow head from the client i; in this embodiment, N is added to the parameters j 、N m,c parameter, N j represents the quantity of the j-th category, and the category samples are normalized to improve the sensitivity to rare category samples, N m,c represents the quantity of the current sample m belonging to the category c;

[0113] Since the fine-grained flow feature is less affected by category imbalance, the loss of it is set as the cross-entropy loss function in this embodiment

[0114] To sum up, this embodiment proposes:

[0115]

[0116] Among them, η is a hyperparameter, respectively represent the losses of the global flow feature separator and the fine-grained flow feature separator, is expressed as the model parameter.

[0117] Furthermore, perform the operation of Step 3:

[0118] Step 3. The client uploads the locally updated multi-flow network feature separator and the head to the server side.

[0119] Furthermore, Step 4 includes:

[0120] Step 4-1. The server side aggregates the two received separators and their respective heads and performs weighted averaging. The formula proposed in this embodiment is as follows:

[0121]

[0122] Among them, is the loss of the i-th client, is the weight of the i-th client.

[0123] Step 4-2: Broadcast the result obtained in Step 4-1 to each client, and continue to loop through Steps 1 to 4 until the algorithm converges.

[0124] To comprehensively evaluate the superior performance of the method innovatively proposed in the present invention in practical applications, the present invention selects a dedicated experimental dataset CIFAR-100, and according to the evaluation criteria detailed in this section, compares and analyzes the proposed method with traditional federated learning algorithms. The experimental environment was set up on the Ubuntu 18.04.3 LTS platform, using an Intel COREI7 13700KF CPU and an NVIDIA GeForce RTX3060TI GPU (32GB RAM). In the selection of programming languages and frameworks, the deep learning models in this article were built using the Python language and the Pytorch framework.

[0125] Specifically, in the specific implementation of the present invention, the selected dataset is the CIFAR-10 dataset, which is widely used to evaluate the performance of image classification algorithms and models due to its moderate difficulty and wide application. Its diversity and complexity make it a classic benchmark for image classification tasks.

[0126] In the specific implementation, the CIFAR-10 dataset is divided into a training set and a test set, where the training set has 50,000 images and the test set has 10,000 images.

[0127] The specific implementation parameters in the present invention are set as follows:

[0128] (1) Dirichlet parameter: The Dirichlet parameter α is set to 0.1 to simulate a relatively high degree of client data heterogeneity.

[0129] (2) Number of clients: The total number of clients is set to 20, and the number of online clients participating in training in each round is 10. The online clients are selected by random sampling.

[0130] (3) Training of client model parameters: The batch size is set to 64, the learning rate is 0.001, the number of local training rounds for the client is set to 1, and the total number of training rounds is 200.

[0131] To verify the effectiveness of the method proposed in the present invention, the most conventional accuracy metric is used to measure the effectiveness of the present invention. In the specific implementation method, the method of the present invention is experimentally compared with advanced models such as the most classic federated learning algorithm FedAvg. As Figure 3As shown, compared with the existing federated learning methods, the performance indicators of the method proposed by the present invention have significant advantages, with an accuracy improvement of 5.6% compared to FedAvg, and the image recognition performance in the case of non-independent and identically distributed data is optimized.

[0132] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A privacy protection method for federated learning based on multi-stream network feature separation, characterized in that: include: S1, broadcasting a feature separator and a header to a client based on a server; wherein the feature separator includes: a global feature separator and a fine-grained feature separator, and the header includes: a global feature header and a fine-grained feature header; S2. Using the client, perform local training iterations on the feature separator and the head; S3, uploading the iterated feature separator and header to the server; S4, aggregating the iterated feature separator and header based on the server to form a new feature separator and header; S5. Continue to iterate S1-S4 until the training of the feature separator and the head is completed; perform federated learning privacy protection based on the trained feature separator and the head.

2. According to claim 1, the method for privacy protection of federated learning based on multi-stream network feature separation is characterized in that: Transmitting feature separators and headers from the server to the client includes: Split the server-side backbone into a feature separator and a header; Initialize model parameters on the server side and distribute them to each client participating in training; In each round of iteration, the server randomly selects a preset portion of clients to participate in training.

3. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 2 is characterized in that: Separating the client's backbone into feature separators consists of: For the global stream feature separator, the input samples are mapped to the feature space, and the head maps the extracted features to the label space; For the fine-grained stream feature separator, the input samples are mapped to the feature space, and the head maps the features to the label space.

4. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 1, characterized in that: Using the client to iterate the feature separator and the server-side header includes: Select a data set and divide the data set into a training set and a test set; Based on a preset loss function, the feature separator and the head are trained using a training set, and tested using a test set; wherein the global feature separator is used for global flow feature separation, and the fine-grained feature separator is used for fine-grained flow feature separation.

5. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 4 is characterized in that: The global feature separator includes: an encryptor M enc With decoder M dec ; The global feature separator performs global flow feature separation including: Given an input feature map L = H × W is the number of elements, D is the dimension of the input feature, and the global flow feature model parameters initialization is the middle layer of the global flow feature, and D' is the feature dimension after initialization, that is, Through the linear layer W enc , linear layer W dec Randomly initialize the encryptor Right now: Decryptor K is the characteristic dimension of the decryptor, that is: M enc Encode the input feature information to generate a unique representation for each input, that is, generate encrypted information so that the input data can be Unable to identify before matching; The encrypted information is: Among them, Norm(·) is standard normalization; Using the encrypted information and M dec Perform matrix multiplication to generate global flow features after the decryptor 6. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 4 is characterized in that: The fine-grained feature separator is obtained based on improving VGG; The improvements to VGG include: simplifying the intermediate network layer of VGG into three branches, two of which are equipped with 1×1 convolutional layers and 3×3 convolutional layers respectively. After each convolution, only batch normalization is performed. The other branch performs average pooling operations in the H and W directions on the input to extract pixel-level features. At the same time, a selective attention mechanism is introduced in this branch to enable the network to suppress irrelevant features and enhance key information. The selective attention mechanism is: Among them, f n*n (·) represents n*n convolution, Avgpooling(·) represents average pooling operation, They represent flow one, flow two, and flow three respectively; the characteristics of the fine-grained flow are:

7. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 4 is characterized in that: The header is: Among them, h g (·;·),h p (·;·) respectively represent parameterized global and fine-grained headers, is the logical value of the global flow feature, A logical value for fine-grained flow features.

8. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 4 is characterized in that: The default loss function is: Among them, η is a hyperparameter, They represent the loss of the global flow feature separator and the fine-grained flow feature separator, respectively. i Expressed as model parameters, x i represents the input of the i-th client, y i Indicates the corresponding label.

9. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 8, characterized in that: The loss of the global flow feature separator is: Among them, α is the weight factor of the category, γ is the focusing parameter, represents the logical value of the cth class of the mth sample in the global stream header from client i, N j It represents the number of the jth class, N m,c It indicates the number of current samples m belonging to category c; The loss of the fine-grained flow feature separator is set to the cross entropy loss function 10. The method for privacy protection of federated learning based on multi-stream network feature separation according to claim 1, characterized in that: Aggregating each part based on the server side includes: The server aggregates the two separators and their respective headers received and performs weighted average, and the formula is as follows: in, is the loss of the ith client, is the weight of the i-th client.

Images