A semi-centralized edge federated partition learning method of a federated learning system for internet of things terminal equipment under a wireless network

Abstract

The application discloses a semi-centralized edge federal segmentation learning method of a federal learning system for Internet of Things terminal equipment under a wireless network. The method comprises the following steps: establishing a to-be-trained model; splitting the to-be-trained model into a server-side first model and a client-side first model; distributing the client-side first model to each client in a to-be-trained client group; training each client-side first model to obtain a client-side second model corresponding to each client and a first gradient; updating the server-side first model according to the multiple first gradients to obtain a server-side second model and a server-side second model feature; updating each client-side second model according to the server-side second model feature to obtain a client-side third model corresponding to each client in the to-be-trained equipment group; repeating the execution; selecting models meeting a first condition from the multiple client-side third models to aggregate the models to obtain an aggregated client-side third model; and combining the aggregated client-side third model and the server-side second model to obtain a trained model. The application has the characteristics of small computing power consumption and small training delay.

Description

Technical Field

[0001] This invention relates to the field of federated segmentation learning technology, and more specifically, to a semi-centralized edge federated segmentation learning method under wireless networks. Background Technology

[0002] Driven by recent advancements in the Internet of Things (IoT), the interconnection of various devices in the physical world via the internet allows for the generation of vast amounts of data from these devices, which can then be used to train large-scale deep learning models. However, transmitting all local data to the cloud or a central base station for model training poses privacy risks.

[0003] Traditionally, distributed training algorithms are often used to address privacy concerns. However, in the Internet of Things (IoT) ecosystem, the network structures of terminal devices are typically not identical, and each device has limited resources such as computing power and battery energy. Therefore, it is often necessary to design multiple dynamic segmentation algorithms for different network structures on multiple terminal devices. This results in complex algorithms and high computational consumption.

[0004] Existing technology discloses a semi-centralized adversarial training method for federated learning, belonging to the interdisciplinary field of security and artificial intelligence. Specifically, it involves: first, determining the features to be prioritized for training within the federated learning framework and obtaining a sample dataset with feature labels; then, pre-training an adversarial generative model based on sample attributes, enabling the model to generate adversarial examples targeting those attributes; next, splitting the adversarial generative model into an encoder and a decoder, sending the encoder to the client, and initiating the federated process; the client uses the received encoder to encode a small number of random samples (5%-10%) locally and sends them to the server; after receiving samples in each round, the server fine-tunes the parameters of the federated learning framework using the adversarial examples and performs the final release. This method consumes significant computational resources on each terminal. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies, such as high computational consumption and complex training methods, by providing a semi-centralized edge federated segmentation learning method for federated learning systems of IoT terminal devices under wireless networks. This method features low computational consumption and low training latency.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0007] A semi-centralized edge federated segmentation learning method for a federated learning system for IoT terminal devices under a wireless network, wherein the federated learning system for IoT terminal devices includes a server and a group of clients to be trained; the group of clients to be trained includes multiple clients, and the semi-centralized edge federated segmentation learning method includes:

[0008] S1: Establish the model to be trained;

[0009] S2: Split the model to be trained into a server-side first model and a client-side first model;

[0010] S3: Assign the first client model to each client in the group of clients to be trained;

[0011] S4: Train the first model for each client to obtain the second model and first gradient for each client;

[0012] S5: Update the server-side first model based on multiple first gradients to obtain the server-side second model and its features;

[0013] S6: Update the second model of each client according to the features of the second model on the server side, and obtain the third model of each client in the client group to be trained;

[0014] S7: Determine whether the first termination condition has been met; if yes, proceed to step S8; if no, take the third client model corresponding to each client in the client group to be trained as the new first client model for each client in the client group to be trained, take the second server model as the new first server model, and proceed to step S4.

[0015] S8: Select multiple client third-party models that meet the first condition and aggregate them to obtain the aggregated client third-party model;

[0016] S9: Combine the aggregated client-side third model and server-side second model to obtain the trained model.

[0017] Further, in step S5, the server-side first model is updated based on multiple client-side second models to obtain the server-side second model, including:

[0018]

[0019] Indicates the learning rate. This indicates the first model on the server side. Denotes the first gradient of client n. This represents the second model on the server side, where m represents the total number of clients and n represents the client sequence number.

[0020] Further, in step S6, based on the server-side second model, the second model of each client is updated to obtain the third model of each client in the client group to be trained, including:

[0021]

[0022] Describes the third client model for client n. Describes the second client model for client n. This represents the local gradient obtained by client n through backpropagation based on the features of the second model on the server.

[0023] Furthermore, in the backpropagation process, data pairs in the client... The formula for calculating the loss function is as follows:

[0024]

[0025]

[0026]

[0027] Represents classification loss, Indicates the reconstruction loss. Represents the reconstruction weights; C represents the number of classification samples. Indicates the true label, This represents the probability of being predicted as class c. This represents the second model feature on the server side. Indicates input data, This represents the structural similarity index.

[0028] Furthermore, in step S2, the model to be trained includes: a feature extraction modules connected end-to-end and a fully connected module;

[0029] The client-side first model includes the first b feature extraction modules in the model to be trained, and the server-side first model includes the last ab feature extraction modules in the model to be trained and a fully connected module.

[0030] A federated segmentation learning method for a clustered federated learning system for IoT terminal devices, wherein the clustered federated learning system for IoT terminal devices includes multiple clusters; each cluster includes a server and a group of clients to be trained; the federated segmentation learning method includes:

[0031] S01: Establish the model to be trained;

[0032] S02: Apply the semi-centralized edge federated segmentation learning method of the federated learning system for IoT terminal devices under a wireless network to learn the model to be trained and multiple client groups to be trained respectively, and obtain multiple trained models.

[0033] S03: Integrate multiple trained models to obtain the final model.

[0034] A federated segmentation learning optimization method includes:

[0035] S11: Obtain the first learning parameters

[0036] S12: Based on the first learning parameters, execute the semi-centralized edge federated segmentation learning method of the federated learning system for IoT terminal devices under a wireless network to obtain the observation parameters;

[0037] S13: Based on the observed parameters, optimize the first learning parameters using the optimization objective to obtain the second learning parameters;

[0038] S14: Use the second learning parameter as the new first learning parameter, and repeat steps S12~S13 until the preset condition is met; use the second learning parameter as the optimal learning parameter.

[0039] Furthermore, the optimization objective includes:

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] Indicates the optimization objective. Indicates the privacy risk weighting factor. This represents the energy consumption weighting factor. Indicates the energy consumption optimization target. Indicates privacy optimization goals, This represents the bandwidth of client n. Indicates the total bandwidth. Indicates the segmentation parameters. This represents the transmit power of client n. This indicates the maximum transmit power. This represents the segmentation ratio parameter. This represents the local training time of client n. This indicates the maximum local training time.

[0047] A semi-centralized edge federated segmentation learning system under a wireless network includes:

[0048] First model acquisition module: Establish the model to be trained;

[0049] Model splitting module: splits the model to be trained into a server-side first model and a client-side first model;

[0050] Model allocation module: Assigns the first client model to each client in the group of clients to be trained;

[0051] First training module: Train the first model for each client to obtain the second model and first gradient for each client;

[0052] First model update module: Based on multiple first gradients, update the server-side first model to obtain the server-side second model and the features of the server-side second model;

[0053] Second model update module: Based on the features of the server-side second model, update the second model of each client to obtain the third model of each client in the client group to be trained;

[0054] Loop module: Determine whether the first termination condition has been met; if yes, execute the model aggregation module; if no, take the third client model corresponding to each client in the client group to be trained as the new first client model for each client in the client group to be trained, take the second server model as the new first server model, and then execute the first training module.

[0055] Model aggregation module: Selects models that meet the first condition from multiple client-side third-party models and aggregates them to obtain the aggregated client-side third-party models;

[0056] Model Combination Module: Combines the aggregated client-side third model and server-side second model to obtain the trained model.

[0057] A cluster-based federated segmentation learning system includes:

[0058] Second model acquisition module: Establish the model to be trained;

[0059] The second training module applies a semi-centralized edge federated segmentation learning method of a federated learning system for IoT terminal devices under a wireless network to learn the model to be trained and multiple client groups to be trained, thereby obtaining multiple trained models.

[0060] Model integration module: Integrates multiple trained models to obtain the final model.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] This invention reduces data transmission and local computation by splitting the model to be trained into a server-side first model and a client-side first model, and updating the server-side model and the client-side model respectively. This results in low training latency and low energy consumption on the terminal device, while protecting the data on the client side.

[0063] This invention can also flexibly adjust the parameters in the learning method according to the learning process, thereby further reducing energy consumption, reducing training latency, and minimizing the risk of privacy leakage. Attached Figure Description

[0064] Figure 1 The flowchart is provided for a semi-centralized edge federation segmentation learning method under a wireless network, as shown in Example 1.

[0065] Figure 2 This is a schematic diagram of network segmentation provided in Example 1.

[0066] Figure 3 The flowchart is provided for a clustered asynchronous federated segmentation learning method in Example 1.

[0067] Figure 4 This is a flowchart of a federated segmentation learning optimization method provided in Example 1.

[0068] Figure 5 This is a schematic diagram of the structure of a clustered asynchronous federated segmentation learning system provided in Example 1. Detailed Implementation

[0069] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0070] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;

[0071] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0072] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] Example 1

[0074] like Figure 1 As shown, a semi-centralized edge federated segmentation learning method for a federated learning system for IoT terminal devices under a wireless network is disclosed. The federated learning system for IoT terminal devices includes a server and a group of clients to be trained. The group of clients to be trained includes multiple clients. The semi-centralized edge federated segmentation learning method includes:

[0075] S1: Establish the model to be trained;

[0076] S2: Split the model to be trained into a server-side first model and a client-side first model;

[0077] S3: Assign the first client model to each client in the group of clients to be trained;

[0078] S4: Train the first model for each client to obtain the second model and first gradient for each client;

[0079] S5: Update the server-side first model based on multiple first gradients to obtain the server-side second model and its features;

[0080] S6: Update the second model of each client according to the features of the second model on the server side, and obtain the third model of each client in the client group to be trained;

[0081] S7: Determine whether the first termination condition has been met; if yes, proceed to step S8; if no, take the third client model corresponding to each client in the client group to be trained as the new first client model for each client in the client group to be trained, take the second server model as the new first server model, and proceed to step S4.

[0082] S8: Select multiple client third-party models that meet the first condition and aggregate them to obtain the aggregated client third-party model;

[0083] S9: Combine the aggregated client-side third model and server-side second model to obtain the trained model.

[0084] In one specific embodiment, the first condition is that it is less than the deadline. The deadline is designed by the edge server based on various considerations within the cluster. Once the deadline is set... Arrival, base station It will be from its associated client node Normalized model updates are collected, and models exceeding a certain age threshold are discarded from the model aggregation phase. Due to differences in available computing resources across clients, the time required for each client to complete local model training varies, resulting in different frequencies of participation in intra-cluster model aggregation. Therefore, setting a deadline can prevent these clients from wasting resources.

[0085] In one specific embodiment, such as Figure 2 As shown, this patent designs a segmentation method for network structure privacy based on a segmentation learning network: the segmentation points are set in the output layer of the convolutional block, and the number of feature channels in the output layer is [number missing]. And satisfy This corresponds to different stages of a convolutional network.

[0086] set up This is a dynamic segmentation layer index, with values ​​that are positive integers. ,in This represents the total number of separable layers in the model. Model scaling parameter. Indexed by segmentation layer Determined through the following piecewise linear mapping relationship:

[0087]

[0088] Edge server According to user node Privacy requirements weight Generate proportionality coefficient Dynamically transform the complete local model Segmented into user node parts server-side part User node The obtained customer model is a complete user node model. The former layer.

[0089] Therefore, the model size can be expressed as , For segmentation layers The value of . We assume that each client (device) has its own local dataset. Size is Equipment integration This indicates that collaborative training aims to train a global model that achieves high inference performance across all devices. Segmentation points are placed at the output layer of convolutional blocks, corresponding to different stages of the convolutional network, to address the issue of device heterogeneity. Different model ratios acquire corresponding device-side and server-side models; the choice of ratio affects the ease with which the server can reconstruct intermediate feature data and the energy consumption for processing model transmission. Simultaneously, this distributed design can significantly alleviate the computational burden on the client side. This segmentation learning design, applied in federated learning, can effectively address the issue of device heterogeneity and respond to the privacy needs of terminal devices and the energy consumption considerations of the client itself.

[0090] Further, in step S5, the server-side first model is updated based on multiple client-side second models to obtain the server-side second model, including:

[0091]

[0092] Indicates the learning rate. This indicates the first model on the server side. Denotes the first gradient of client n. This represents the second model on the server side, where m represents the total number of clients and n represents the client sequence number.

[0093] Further, in step S6, based on the server-side second model, the second model of each client is updated to obtain the third model of each client in the client group to be trained, including:

[0094]

[0095] Describes the third client model for client n. Describes the second client model for client n. This represents the local gradient obtained by client n through backpropagation based on the features of the second model on the server.

[0096] Furthermore, in the backpropagation process, data pairs in the client... The formula for calculating the loss function is as follows:

[0097]

[0098]

[0099]

[0100] Represents classification loss, Indicates the reconstruction loss. Represents the reconstruction weights; C represents the number of classification samples. Indicates the true label, This represents the probability of being predicted as class c. This represents the second model feature on the server side. Indicates input data, This represents the structural similarity index.

[0101] Furthermore, in step S8, the formula for the aggregated client-side third model is obtained as follows:

[0102]

[0103] This represents the aggregated client-side third model. This indicates the number of models in the client's third model that meet the first condition. This represents the nth client third model that meets the first condition.

[0104] Furthermore, such as Figure 2As shown, in step S2, the model to be trained includes: a feature extraction modules connected end to end, and a fully connected module;

[0105] The client-side first model includes the first b feature extraction modules in the model to be trained, and the server-side first model includes the last ab feature extraction modules in the model to be trained and a fully connected module.

[0106] The pseudocode for model segmentation is as follows:

[0107]

[0108] like Figure 3 As shown, a federated segmentation learning method for a clustered federated learning system for IoT terminal devices is disclosed. The clustered federated learning system for IoT terminal devices comprises multiple clusters; each cluster includes a server and a group of clients to be trained. The federated segmentation learning method includes:

[0109] S01: Establish the model to be trained;

[0110] S02: Apply the semi-centralized edge federated segmentation learning method of the federated learning system for IoT terminal devices under a wireless network to learn the model to be trained and multiple client groups to be trained respectively, and obtain multiple trained models.

[0111] S03: Integrate multiple trained models to obtain the final model.

[0112] Furthermore, in step S03, the multiple trained models are integrated to obtain the formula for the final model, as follows:

[0113]

[0114]

[0115] This represents the nth trained model, where n is the index. This represents the nth weight. This represents the final model.

[0116] like Figure 4 As shown, a federated segmentation learning optimization method includes:

[0117] S11: Obtain the first learning parameters

[0118] S12: Based on the first learning parameters, execute the semi-centralized edge federated segmentation learning method of the federated learning system for IoT terminal devices under a wireless network to obtain the observation parameters;

[0119] S13: Based on the observed parameters, optimize the first learning parameters using the optimization objective to obtain the second learning parameters;

[0120] S14: Use the second learning parameter as the new first learning parameter, and repeat steps S12~S13 until the preset condition is met; use the second learning parameter as the optimal learning parameter.

[0121] Furthermore, the optimization objective includes:

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] Indicates the optimization objective. Indicates the privacy risk weighting factor. This represents the energy consumption weighting factor. Indicates the energy consumption optimization target. Indicates privacy optimization goals, This represents the bandwidth of client n. Indicates the total bandwidth. Indicates the segmentation parameters. This represents the transmit power of client n. This indicates the maximum transmit power. This represents the segmentation ratio parameter. This represents the local training time of client n. This indicates the maximum local training time.

[0129] The pseudocode for a federated segmentation learning optimization method is as follows:

[0130]

[0131] In one specific embodiment, the privacy index is evaluated as follows:

[0132] Assuming the server is honest but curious, it can still reconstruct the original input data from the client's shredded data; this process is called a reconstruction attack. In this case, the privacy risk associated with such a reconstruction attack can be assessed by measuring the similarity between the original image and the reconstructed image (called the Structural Similarity Index (SSIM)). As the segmentation layer index increases, the quality of the reconstructed image decreases, indicating increased model complexity, but with minimal impact on model accuracy. This degradation and enhanced privacy are interchangeable. This is because the complexity of the model deployed on the client introduces advanced nonlinearities into the output; this increased complexity makes it more difficult for adversaries to reverse engineer and retrieve private input data. Therefore, the selection of segmentation points needs to consider energy consumption and training quality. Simultaneously, regarding model performance requirements, to assess the degree of privacy exposure risk after segmentation, the similarity between the original input and the reconstructed image can be measured. of To assess the privacy risks associated with reconstruction attacks ( (This can be approximately mapped to the index of the cutting layer), the structural similarity index is expressed as...

[0133]

[0134] in and These are image patches representing the original image and the reconstructed image, respectively. (Luminance) represents the comparison of brightness. (Contrast) is used to represent contrast. The range of this indicator is: .

[0135]

[0136]

[0137]

[0138]

[0139] in and They are and The mean; and yes and The variance; yes and covariance; and It is a small constant added to stabilize the denominator. ;

[0140] However, the structural similarity index formula does not directly reflect the segmentation layer. The necessary connection between them It only records the image reconstruction quality corresponding to different segmentation layers, lacking details about the segmentation layers themselves. Theoretical analysis is needed. In this case, empirical measurements can be used to apply a data-driven approach to the model. Regression-based modeling is employed, one of the most widely used methods, including mobile CPU attribute models (e.g., modeling CPU power and temperature variations). Hypotheses are made based on observed data. and There exists a quadratic functional relationship (convex function) between them, in the form of:

[0141]

[0142] in, These are the parameters to be fitted.

[0143] The experimental data were fitted using the least squares method, with the optimization objective being to minimize the mean squared error (MSE).

[0144]

[0145] Solving for parameters using matrix operations or optimization libraries, a smaller root mean square error (RMSE) indicates higher model fitting accuracy. (The coefficients of the quadratic term are also relevant.) If it is positive, it means Follow The increase exhibits a convex function change, i.e., deep cutting ( When the value is large, the structural similarity index decreases, and privacy protection is enhanced. (First-order term coefficient) A negative value reflects right The linear suppression effect is observed. The quadratic coefficient is positive, the model is a convex function, which meets the conditions for the global optimum in the optimization problem. Therefore, the optimal cutting layer can be solved through convex optimization. This patent uses this method to determine the segmentation ratio. and The connection is to adapt to the network training of reinforcement learning. At the same time, the optimal model ratio obtained in reinforcement learning can also be used to appropriately adjust the selection of the segmentation position of the entire model.

[0146] In one specific embodiment, the communication index is evaluated as follows:

[0147] Assume that channels can be perfectly obtained on the device (client) through uplink or downlink channel estimation. These channels remain static during one round of training but vary in different training sessions. To obtain a hyperspectral signal while ensuring communication quality, Each device uses non-vertical multiple access to upload data and download the gradient of fragmented gradient data or the device model.

[0148] Assuming the channels are arranged in descending order , indicating equipment Having the first The strongest instantaneous channel. In this scenario, each device receives the optimal channel allocation equally. Uplink speed and downlink speed They are respectively

[0149]

[0150]

[0151] in For terminal devices System bandwidth, For equipment The transmission power, For base station as equipment Distributed power, and Representing base station and equipment respectively The power of the received additive white Gaussian noise (AWGN).

[0152] ① Client-side model download and upload latency: Let The size of the complete model (in bits) can be represented as follows: The device-side model can be represented as... .equipment Download and upload latency are calculated using the following formulas:

[0153]

[0154]

[0155] ② Client-side model forward propagation delay:

[0156] Assume each participating terminal device uses a fixed size Random small batch Conduct a round of training. (Set) The computational workload required to process each piece of data in the segmentation layer during the forward propagation at the device end (which can be expressed in FLOPs). This represents the computational workload required to train the complete model on the device. Therefore, it can be represented as... .equipment The forward propagation delay is calculated as follows:

[0157]

[0158] in It is equipment The frequency of the central processing unit (CPU) (in cycles per second). It represents the number of FLOPs per cycle.

[0159] ③ Data uplink transmission delay:

[0160] make This indicates the output size (in bits) of one data segmentation layer required in the device-side model. Device The gradient transmission delay of the crushing data is determined as follows:

[0161]

[0162] ④ Server-side model latency:

[0163] make This indicates the first step in the forward and backward propagation on the server side. The computational workload required for each data segmentation layer to process each piece of data. The forward propagation latency and backward propagation latency of the server-side model are given by the following formula.

[0164]

[0165] in It is the CPU frequency of the edge server, and It is a floating-point number for each cycle of the edge server.

[0166] ⑤ Downlink transmission delay of shredded data:

[0167] make This indicates the output size (in bits) of one data point required by the server-side model segmentation layer. Device The gradient transmission delay of the crushing data is determined as

[0168]

[0169] ⑥ Backpropagation delay of the device-side model:

[0170] set up This represents the computational workload (in FLOPs) required for the segmentation layer to process each piece of data during the backpropagation process at the device end. This represents the computational workload required for backpropagation of the complete device-side model. Therefore, it can be represented as... .equipment The backpropagation delay is calculated as follows:

[0171]

[0172] Considering the above latency components, the total latency for both local and global training epochs is:

[0173]

[0174] Therefore, when selecting latency within the cluster, the device with the highest latency during local training should be chosen.

[0175]

[0176] The designed training cutoff time should be related to the average training latency in the cluster, and is designed as follows:

[0177]

[0178]

[0179] in This is a constant, representing the training redundancy time.

[0180] In one specific embodiment, the energy consumption index is evaluated as follows:

[0181] Considering the ample resources of edge servers and base stations, we ignore the energy consumption during the gradient descent process, as well as the energy consumption for model aggregation and updates between edge server clusters and server-side computation. In our model, energy consumption mainly occurs during local computation and model upload. Due to the heterogeneity of computing resources and communication conditions among different smart devices, the time spent on local computation and model upload varies for each device, resulting in different energy consumption. We will model the energy consumption of these two processes separately below:

[0182] ① Communication power consumption:

[0183] Considering the downloading and uploading of the model on the device, the transmission of compressed data, and the gradient of the compressed data, the communication energy consumption per device per round of training is:

[0184]

[0185] ② Calculate energy consumption:

[0186] Since computational power consumption includes the selected equipment, the computational power consumption for one round of training is given by the following formula:

[0187]

[0188] in It is an effective switched capacitor, depending on the chip architecture. Considering the above two parts, as well as local and global training, the cumulative power consumption of a single device is...

[0189]

[0190] Therefore, the total energy consumption for training all devices is

[0191]

[0192] The average energy consumption of the terminal devices within the cluster is

[0193]

[0194] Since it is to be used as the final optimization target, it is normalized:

[0195]

[0196] in and This represents the extreme energy consumption within the sliding window. It is a very small constant.

[0197] In one specific embodiment, the resource scheduling problem of terminal devices under the federated split learning framework can be represented by the following objective equation: This objective equation focuses on reducing battery energy consumption of terminal devices and improving privacy protection throughout the entire scenario. Because the system involves the collaborative operation of multiple clusters, using energy consumption and privacy protection within each cluster as optimization objectives can optimize the system performance of multiple clusters and prevent resource depletion in isolated clusters.

[0198] Dynamically calculate weights:

[0199]

[0200] When the overall privacy requirements of the device are high Increase; when energy consumption is strictly limited, Increase.

[0201] A semi-centralized edge federated segmentation learning system under a wireless network includes:

[0202] First model acquisition module: Establish the model to be trained;

[0203] Model splitting module: splits the model to be trained into a server-side first model and a client-side first model;

[0204] Model allocation module: Assigns the first client model to each client in the group of clients to be trained;

[0205] First training module: Train the first model for each client to obtain the second model and first gradient for each client;

[0206] First model update module: Based on multiple first gradients, update the server-side first model to obtain the server-side second model and the features of the server-side second model;

[0207] Second model update module: Based on the features of the server-side second model, update the second model of each client to obtain the third model of each client in the client group to be trained;

[0208] Loop module: Determine whether the first termination condition has been met; if yes, execute the model aggregation module; if no, take the third client model corresponding to each client in the client group to be trained as the new first client model for each client in the client group to be trained, take the second server model as the new first server model, and then execute the first training module.

[0209] Model aggregation module: Selects models that meet the first condition from multiple client-side third-party models and aggregates them to obtain the aggregated client-side third-party models;

[0210] Model Combination Module: Combines the aggregated client-side third model and server-side second model to obtain the trained model.

[0211] A cluster-based federated segmentation learning system includes:

[0212] Second model acquisition module: Establish the model to be trained;

[0213] The second training module applies a semi-centralized edge federated segmentation learning method of a federated learning system for IoT terminal devices under a wireless network to learn the model to be trained and multiple client groups to be trained, thereby obtaining multiple trained models.

[0214] Model integration module: Integrates multiple trained models to obtain the final model.

[0215] like Figure 5 As shown, the system scenario is One edge cluster, including Each cluster consists of one client and one edge server (ES). ) and a set of associated client nodes (represented as The system consists of several components. Each edge server coordinates at least one client node, and each client node is associated with only one edge server. Devices are selected based on predefined criteria (e.g., physical proximity and network coverage). Furthermore, edge servers are connected to neighboring servers via high-speed cables. The nodes are connected to form a connected graph. The scene is represented in the form of a cluster, which is independent of the central server and adopts a semi-centralized and semi-asynchronous training method, effectively mitigating the "lagging effect".

[0216] This patent's network architecture is designed to balance privacy protection and task performance in distributed scenarios. It achieves both the core task objective and the privacy protection objective through a dual-output branch of "classification + reconstruction." Simultaneously, the training of the reconstructed image simulates an attacker's approach to assess the risk of feature privacy leakage under this network structure, allowing for the rational selection of segmentation layers. Through client-server collaborative computation, the classification task of the input data is achieved, improving classification accuracy. The original input image is reconstructed based on the intermediate feature maps output by the client. The core objective is to protect data privacy through "feature reversibility" (even if intermediate features are leaked, the risk of sensitive information exposure can be reduced through reconstruction loss constraints). The client is responsible for the first half of feature extraction, outputting intermediate feature maps; the server is responsible for the second half of feature processing and classification output, while the reconstruction branch reconstructs the input image based on the intermediate feature maps.

[0217] The same or similar labels correspond to the same or similar parts;

[0218] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0219] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A semi-centralized edge federated segmentation learning method for a federated learning system for IoT terminal devices under a wireless network, wherein, A federated learning system for IoT terminal devices includes a server and a group of client devices to be trained. The client group to be trained includes multiple clients, characterized in that the semi-centralized edge federated segmentation learning method includes: S1: Establish the model to be trained; S2: Split the model to be trained into a server-side first model and a client-side first model; S3: Assign the first client model to each client in the group of clients to be trained; S4: Train the first model for each client to obtain the second model and first gradient for each client; S5: Update the server-side first model based on multiple first gradients to obtain the server-side second model and its features; S6: Update the second model of each client according to the features of the second model on the server side, and obtain the third model of each client in the client group to be trained; S7: Determine whether the first termination condition has been met; if yes, proceed to step S8; if no, take the third client model corresponding to each client in the client group to be trained as the new first client model for each client in the client group to be trained, take the second server model as the new first server model, and proceed to step S4. S8: Select multiple client third-party models that meet the first condition and aggregate them to obtain the aggregated client third-party model; S9: Combine the aggregated client-side third model and server-side second model to obtain the trained model; In step S6, based on the server-side second model, the second model of each client is updated to obtain the third model of each client in the client group to be trained, including: Describes the third client model for client n. Describes the second client model for client n. This represents the local gradient obtained by client n through backpropagation based on the features of the second model on the server. Data pairs in the client during backpropagation The formula for calculating the loss function is as follows: Represents classification loss, Indicates the reconstruction loss. Represents the reconstruction weights; C represents the number of classification samples. Indicates the true label, This represents the probability of being predicted as class c. This represents the second model feature on the server side. Indicates input data, Represents the structural similarity index; This is obtained by measuring the similarity between the original image and the reconstructed image; In step S2, the model to be trained includes: a feature extraction modules connected end to end and a fully connected module; The client-side first model includes the first b feature extraction modules in the model to be trained, and the server-side first model includes the last ab feature extraction modules in the model to be trained and a fully connected module. The client is responsible for the first half of feature extraction and outputs the intermediate feature map; the server is responsible for the second half of feature processing and classification output, and at the same time, the reconstruction branch reconstructs the input image based on the intermediate feature map.

2. The semi-centralized edge federated segmentation learning method for a federated learning system for IoT terminal devices under a wireless network according to claim 1, characterized in that, In step S5, the server-side first model is updated based on multiple client-side second models to obtain the server-side second model, including: Indicates the learning rate. This indicates the first model on the server side. Denotes the first gradient of client n. This represents the second model on the server side, where m represents the total number of clients and n represents the client sequence number.

3. A federated segmentation learning method for a clustered federated learning system for Internet of Things (IoT) terminal devices, wherein, A clustered federated learning system for IoT terminal devices comprises multiple clusters; each cluster includes a server and a group of clients to be trained. Its key feature is a federated segmentation learning method, including: S01: Establish the model to be trained; S02: Apply the semi-centralized edge federated segmentation learning method of the federated learning system for IoT terminal devices under wireless network as described in any one of claims 1 to 2, and learn the model to be trained and multiple client groups to be trained respectively to obtain multiple trained models; S03: Integrate multiple trained models to obtain the final model.

4. A federated segmentation learning optimization method, characterized in that, include: S11: Obtain the first learning parameters S12: Based on the first learning parameters, execute the semi-centralized edge federation segmentation learning method of the federated learning system for IoT terminal devices under a wireless network as described in any one of claims 1 to 2 to obtain the observation parameters; S13: Based on the observed parameters, optimize the first learning parameters using the optimization objective to obtain the second learning parameters; S14: Use the second learning parameter as the new first learning parameter, and repeat steps S12~S13 until the preset condition is met; use the second learning parameter as the optimal learning parameter.

5. The federated segmentation learning optimization method according to claim 4, characterized in that, The optimization objectives include: Indicates the optimization objective. Indicates the privacy risk weighting factor. This represents the energy consumption weighting factor. Indicates the energy consumption optimization target. Indicates privacy optimization goals, This represents the bandwidth of client n. Indicates the total bandwidth. Indicates the segmentation parameters. This represents the transmit power of client n. This indicates the maximum transmit power. This represents the segmentation ratio parameter. This represents the local training time of client n. This indicates the maximum local training time.

6. A semi-centralized edge federated segmentation learning system under a wireless network, applied to the learning method described in any one of claims 1 to 2, characterized in that, include: First model acquisition module: Establish the model to be trained; Model splitting module: splits the model to be trained into a server-side first model and a client-side first model; Model allocation module: Assigns the first client model to each client in the group of clients to be trained; First training module: Train the first model for each client to obtain the second model and first gradient for each client; First model update module: Based on multiple first gradients, update the server-side first model to obtain the server-side second model and the features of the server-side second model; Second model update module: Based on the features of the server-side second model, update the second model of each client to obtain the third model of each client in the client group to be trained; Loop module: Determines whether the first termination condition has been met; If so, then execute the model aggregation module; If not, the third client model corresponding to each client in the client group to be trained is used as the new first client model for each client in the client group to be trained, and the second server model is used as the new first server model, then the first training module is executed. Model aggregation module: Selects models that meet the first condition from multiple client-side third-party models and aggregates them to obtain the aggregated client-side third-party models; Model Combination Module: Combines the aggregated client-side third model and server-side second model to obtain the trained model.

7. A cluster-based federated segmentation learning system, applied to the learning method of claim 3, characterized in that, include: Second model acquisition module: Establish the model to be trained; The second training module applies a semi-centralized edge federated segmentation learning method of a federated learning system for IoT terminal devices under a wireless network to learn the model to be trained and multiple client groups to be trained, thereby obtaining multiple trained models. Model integration module: Integrates multiple trained models to obtain the final model.

Images