A federated learning method and system for power system edge device heterogeneity

Abstract

This invention belongs to the interdisciplinary field of federated learning and privacy protection technology, and discloses a federated learning method and system for the heterogeneity of edge devices in power systems. By constructing dynamic dual-cooperative subsets—a demand-cooperative subset and a support-cooperative subset—a transmission efficiency model jointly optimized by communication and computation is designed to achieve efficient complementary collaboration of gradients among edge devices and adaptive resource allocation. Through innovative design of a time-varying learning rate compensation function and a two-stage aggregation strategy, fair participation of heterogeneous devices and efficient fusion of Non-IID data are achieved. This invention breaks down barriers of device capability differences through a dynamic gradient complementarity mechanism and eliminates the impact of transmission latency through the construction of communication-aware dynamic cooperative subsets, ultimately realizing a highly robust, low-latency, and privacy-protected distributed model training paradigm, providing key technical support for the intelligent deployment of power system edges.

Description

Technical Field

[0001] This application relates to the intersection of federated learning and privacy protection technologies, specifically to a federated learning method and system for addressing the heterogeneity of edge devices in power systems. Background Technology

[0002] Federated learning, as a privacy-preserving distributed training paradigm, allows multiple devices to collaboratively train a global model without sharing local data. However, federated learning faces significant challenges in power systems: 1) Device heterogeneity: significant differences in computing power among edge devices (e.g., GPU servers and embedded devices) lead to strong devices waiting for weaker devices during synchronous aggregation, resulting in low resource utilization; 2) Communication constraints: power line carrier communication (PLC) bandwidth fluctuations >40%, and unstable device communication bandwidth (packet loss rate >15%); Non-IID data: differences in load curve distribution between industrial and commercial areas and residential areas cause global model bias, increasing the measured load prediction error by 40%.

[0003] Currently, common solutions to address federation issues stemming from device heterogeneity include the following: 1. Asynchronous aggregation mechanism: Allowing devices to independently complete training and immediately upload gradients, with the server updating the model through staleness compensation (such as latency decay factor). 2. Device grouping strategy: Grouping by computing power / network status, with synchronous aggregation within groups and asynchronous collaboration between groups. 3. Adaptive local training: Dynamically adjusting the number of local iterations for weaker devices (such as early stopping mechanism) or increasing the number of local rounds for stronger devices.

[0004] However, the methods described above for addressing the issues of heterogeneous devices, limited resources, and non-independent and identically distributed (non-IID) data among participating devices in existing federated learning still have some significant technical problems: First, the asynchronous aggregation mechanism described above is difficult to design compensation functions for, and fixed decay factors (such as γ=0.99) are difficult to adapt to dynamic latency, exacerbating the Non-IID bias, with high-frequency, powerful devices dominating model updates and reducing accuracy; Second, the device grouping strategy described above suffers from group mismatch due to dynamic changes in device status, resulting in a doubling of communication overhead, with intra-group broadcasting and inter-group synchronization increasing traffic by nearly 35%; Third, the adaptive local training described above suffers from underfitting on weak devices, leading to increased gradient noise, unstable model convergence, a three-fold increase in computational load on powerful devices, a surge in energy consumption, and an imbalance in resource utilization.

[0005] Against this backdrop, this invention provides a federated learning method and system for addressing the heterogeneity of edge devices in power systems. Through a pioneering dual-cooperative subset dynamic architecture (combining demand subsets with support subsets) coupled with adaptive learning rate compensation, it completely breaks the "stronger gets stronger" dilemma. It increases the participation rate and gradient contribution weight of weaker devices, suppressing the monopoly effect of high-frequency, high-performance devices. A flexible cooperation mechanism based on real-time transmission efficiency perception is proposed; this dynamic load balancing reduces the total communication volume and energy consumption of high-performance devices. Through a negative correlation gradient fusion algorithm, the gradients of the least similar devices are intelligently weighted, effectively neutralizing data bias. It can adapt to extreme Non-IID scenarios, improving convergence stability. Summary of the Invention

[0006] The purpose of this application is to provide a federated learning method for addressing the heterogeneity of edge devices in power systems. Through an innovative dual-cooperative subset dynamic architecture (combining demand subsets with support subsets) coupled with adaptive learning rate compensation, the participation rate of weaker devices and their gradient contribution weights are improved. An elastic cooperation mechanism based on real-time transmission efficiency awareness reduces the total communication volume and energy consumption of stronger devices. Through a negative correlation gradient fusion algorithm, the gradients of the least similar devices are intelligently weighted, effectively neutralizing data bias.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] In a first aspect, this invention discloses a federated learning method for addressing the heterogeneity of edge devices in a power system, comprising the following steps: S1, the federated central server initializes the global model and distributes it to all participating devices; S2, the server collects the available bandwidth of each device. Time required for a round of local training S3. The server constructs the required collaboration subset and the support collaboration subset for each device; S4. After completing one round of local training, the device encrypts and uploads the gradients to the server, and simultaneously constructs the support collaboration subset... Gradients are distributed to the devices, either directly or via a server; S5. After receiving the gradients uploaded by the devices, the server sends the global model parameters to the devices and then updates the global model using a dynamic staleness compensation strategy; S6. After receiving the global model parameters, the devices combine the gradient information of the required collaborative subset received during this training round and perform two-stage aggregation; S7. When the number of global aggregations on the server reaches Tg, the federated learning training ends.

[0009] More specifically, the federated learning method for heterogeneous edge devices in a power system, as described above, is characterized in that, in step S2, the available bandwidth B of each device and the time required for one round of local training are... After completing one round of local training, the device collects the currently available bandwidth and the time required to complete this round of training, and then encrypts and uploads the data to the server. Preferably, Take the moving average of the most recent 3 rounds.

[0010] More specifically, the federated learning method for addressing the heterogeneity of edge devices in a power system, as described above, is characterized in that, in step S3, the server constructs a demand collaboration subset and a support collaboration subset for each device, specifically including the following steps: S31, calculating the gradient similarity matrix between devices based on the latest gradient information uploaded by each device collected by the server. The gradient similarity between device i and device j , Let i be the model gradient of device i. Let be the model gradient of device j;

[0011] S32. Calculate the transmission efficiency of each device, such as the transmission efficiency of device i. Among them, molecules Indicates equipment The theoretical amount of data that can be transmitted during training (bandwidth × time) reflects the actual communication capability. Denominator This represents the theoretical requirement for gradient transfer. Let i be the model gradient memory. All device models are the same, and the model gradient memory is the same size. 2 is because the gradient of the receiving demand collaboration subset and the gradient of the transmission to the helping collaboration subset are both used.

[0012] S33, Based on transmission efficiency and number of participating devices Determine the number of required collaborative subsets for device i. ;

[0013] S34. Selection and Equipment Least similar One device ( The smallest (i.e., the least similar) subset of the required collaboration for device i ;

[0014] S35. Derive the assistance collaboration subset for each device from the demand collaboration subset of all devices. That is, which equipment's demand collaboration subset includes that equipment;

[0015] S36, If the equipment The subset of assistance and cooperation Size greater than Then retain the least similar Each device is notified to update its required collaboration subset.

[0016] The required collaboration subsets and assistance collaboration subsets for each device are dynamically constructed, and the federal center server executes them each time. Global model update involves performing a construction operation to build a complementary requirement collaboration subset for each participating device, preferably... .

[0017] More specifically, the federated learning method for heterogeneous edge devices in a power system, as described above, is characterized in that, in step S4, the gradient-encrypted transmission of the cooperative subset helps the devices... gradient The encryption / decryption module performs the following operations: 1) Direct transmission: Sends the data using end-to-end encryption. 1) Devices in the middle; 2) Server relay: After decryption by the central server's encryption / decryption module, it is re-encrypted and forwarded to the device. The equipment in the middle.

[0018] More specifically, the federated learning method for addressing the heterogeneity of power system edge devices described above is characterized in that, in step S5, updating the global model using a dynamic staleness compensation strategy specifically includes the following steps:

[0019] S51, the Federation Center server comes from the device. gradient global model parameters Transmitted to device ;

[0020] S52, Update Global Model Its learning rate The calculation method is as follows:

[0021] ,

[0022] in , The total number of times the global model is updated. This represents the number of times the current global model has been updated. For equipment The number of times gradients were uploaded to the federated central server. The global model learning rate. For weight hyperparameters, preferably .

[0023] The learning rate Simultaneously, dynamic adjustments are made based on the participation frequency: when the device participates in the frequency... At that time, the amplification compensation coefficient is increased to ;when When the attenuation compensation coefficient reaches .

[0024] The reason why the server first transmits the global model to the device that uploaded the model and then performs gradient aggregation is that the device will perform local aggregation based on the global model it sends out, thus avoiding duplicate aggregation.

[0025] More specifically, the federated learning method for addressing the heterogeneity of power system edge devices described above is characterized in that, in step S6, the device locally performs two-stage aggregation, specifically including the following steps:

[0026] S61. Perform the first-stage interpolation aggregation update with the global model: ;

[0027] in These are the interpolation weight hyperparameters of the global model, which gradually decrease as training progresses.

[0028] S62, and the gradient set of the required collaboration subset received during this round of training. Perform the second phase of weighted aggregate update. , , .

[0029] in It is the local learning rate of device i. It is the similarity between device i from the federated center server and device k in its demand collaboration subset. The smaller the value, the larger the gradient weight corresponding to device k, and the stronger the complementarity.

[0030] Secondly, the present invention discloses a system for implementing a federated learning method for the heterogeneity of edge devices in a power system, comprising: a central server module and a client module.

[0031] The central server module includes: a resource monitor, a dynamic collaborative subset manager, a global aggregator, and an encryption / decryption module;

[0032] The resource monitor is used to collect and store the network bandwidth of each device and the time required to perform one round of local training;

[0033] The dynamic collaboration manager is used to calculate the gradient similarity matrix between devices and to build the required collaboration subset and the assistance collaboration subset.

[0034] The global aggregator is used to execute dynamic obsolescence compensation strategies to update the global model;

[0035] The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, and supports protocols such as AES-256 / SM4.

[0036] The client module includes: a resource detector, a local training engine, a collaborative gradient aggregator, and an encryption / decryption module;

[0037] The resource monitor is used to locally measure network bandwidth and the time required for a round of local training, and reports it to the server after signing it using the encryption and decryption module;

[0038] The local training engine performs local model training, calculates gradients, and uploads the gradients to the server and the support subset.

[0039] The collaborative gradient aggregator is used to implement two-stage aggregation. The first stage uses the global model from the server to perform interpolation aggregation, and the second stage uses gradients from the required collaborative subset to perform weighted aggregation.

[0040] The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, and supports protocols such as AES-256 / SM4.

[0041] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention improves the participation rate and gradient contribution weight of weak devices by employing an innovative dual-cooperative subset dynamic architecture (combining demand subsets with support subsets) with adaptive learning rate compensation; a flexible cooperative mechanism based on real-time transmission efficiency awareness reduces the total communication volume and energy consumption of strong devices; and a negative correlation gradient fusion algorithm intelligently weights the gradients of the least similar devices, effectively neutralizing data bias. This provides a reliable solution for heterogeneous federated learning of edge devices in power systems. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart for a federated learning method for addressing the heterogeneity of edge devices in a power system, provided for the application.

[0044] Figure 2 This is an architecture diagram of a federated learning system for addressing the heterogeneity of edge devices in a power system, provided for the application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0046] The terms “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0047] The terms “first,” “second,” etc., are used only to distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance, nor as requiring or implying any such actual relationship or order between these entities or operations.

[0048] The basic idea of ​​this invention is that the federated central server initializes the global model and distributes it to all participating devices; the server collects the available bandwidth B and the time required for one round of local training from each device. The server constructs the required collaboration subset and the support collaboration subset for each device; after completing one round of local training, the device encrypts and uploads the gradients to the server, and simultaneously, based on the support collaboration subset... Gradients are distributed to devices, either directly or via a server. Upon receiving gradients from a device, the server sends global model parameters to that device and updates the global model using a dynamic staleness compensation strategy. After receiving the global model parameters, the device performs a two-stage aggregation, combining the gradient information from the required collaborative subset received during the current training round. Federated learning training ends when the number of global aggregations on the server reaches Tg. This innovative dual-collaboration subset dynamic architecture (combining required and support subsets) with adaptive learning rate compensation enhances the participation rate and gradient contribution weight of weaker devices. A flexible collaboration mechanism based on real-time transmission efficiency awareness reduces total communication volume and energy consumption of stronger devices. A negative correlation gradient fusion algorithm intelligently weights the gradients of the least similar devices, effectively neutralizing data bias. This provides a reliable solution for heterogeneous federated learning of edge devices in power systems.

[0049] like Figure 1 As shown, this invention provides a federated learning method for addressing the heterogeneity of edge devices in power systems, comprising the following steps:

[0050] S1. The Federation Center server initializes the global model and distributes it to all participating devices;

[0051] S2, The server collects the available bandwidth of each device. Time required for a round of local training ;

[0052] After the device completes one round of local training, it collects the device's currently available bandwidth and the time required to complete this round of training, and then encrypts and uploads the data to the server. Preferably, Take the moving average of the most recent 3 rounds.

[0053] S3, the server constructs the required collaboration subset and the support collaboration subset for each device;

[0054] Specifically, the steps include: S31, calculating the gradient similarity matrix between devices based on the latest gradient information uploaded by each device collected by the server. The gradient similarity between device i and device j , Let i be the model gradient of device i. Let be the model gradient of device j;

[0055] S32. Calculate the transmission efficiency of each device, such as the transmission efficiency of device i. Among them, molecules Indicates equipment The theoretical amount of data that can be transmitted during training (bandwidth × time) reflects the actual communication capability. Denominator This represents the theoretical requirement for gradient transfer. Let i be the model gradient memory. All device models are the same, and the model gradient memory is the same size. 2 is because the gradient of the receiving demand collaboration subset and the gradient of the transmission to the helping collaboration subset are both used.

[0056] S33, Based on transmission efficiency and number of participating devices Determine the number of required collaborative subsets for device i. ;

[0057] S34. Selection and Equipment Least similar One device ( The smallest (i.e., the least similar) subset of the required collaboration for device i ;

[0058] S35. Derive the assistance collaboration subset for each device from the demand collaboration subset of all devices. That is, which equipment's demand collaboration subset includes that equipment;

[0059] S36, If the equipment The subset of assistance and cooperation Size greater than Then retain the least similar Each device is notified to update its required collaboration subset.

[0060] The required collaboration subsets and support collaboration subsets for each device are dynamically constructed, and the federal center server executes them each time. Global model update involves performing a construction operation to build a complementary requirement collaboration subset for each participating device, preferably... .

[0061] S4. After completing one round of local training, the device encrypts and uploads the gradients to the server, and simultaneously performs training based on the collaborative subset. The gradient is distributed to the devices within it, either directly or via a server.

[0062] equipment gradient The encryption / decryption module performs the following operations: 1) Direct transmission: Sends the data using end-to-end encryption. 1) Devices in the middle; 2) Server relay: After decryption by the central server's encryption / decryption module, it is re-encrypted and forwarded to the device. The equipment in the middle.

[0063] S5. After receiving the gradient uploaded by the device, the server sends the global model parameters to the device and then updates the global model using a dynamic staleness compensation strategy.

[0064] Specifically, the following steps are included:

[0065] S51, the Federation Center server comes from the device. gradient global model parameters Transmitted to device ;

[0066] S52, Update Global Model Its learning rate The calculation method is as follows:

[0067] ,

[0068] in , The total number of times the global model is updated. This represents the number of times the current global model has been updated. For equipment The number of times gradients were uploaded to the federated central server. The global model learning rate. For weight hyperparameters, preferably .

[0069] Learning rate Simultaneously, dynamic adjustments are made based on the participation frequency: when the device participates in the frequency... At that time, the amplification compensation coefficient is increased to ;when When the attenuation compensation coefficient reaches .

[0070] The reason why the server first transmits the global model to the device that uploaded the model and then performs gradient aggregation is that the device will perform local aggregation based on the global model it sends out, thus avoiding duplicate aggregation.

[0071] S6. After receiving the global model parameters, the device combines the gradient information of the required collaborative subset received during this round of training and performs two-stage aggregation.

[0072] Specifically, the following steps are included:

[0073] S61. Perform the first-stage interpolation aggregation update with the global model: ;

[0074] in These are the interpolation weight hyperparameters of the global model, which gradually decrease as training progresses.

[0075] S62, and the gradient set of the required collaboration subset received during this round of training. Perform the second phase of weighted aggregate update. , , .

[0076] in It is the local learning rate of device i. It is the similarity between device i from the federated center server and device k in its demand collaboration subset. The smaller the value, the larger the gradient weight corresponding to device k, and the stronger the complementarity.

[0077] S7. When the number of global aggregations on the server reaches Tg, the federated learning training ends.

[0078] like Figure 2 As shown, the present invention provides an architecture diagram of a federated learning system for the heterogeneity of edge devices in a power system, comprising: a central server module and a client module.

[0079] The central server module includes: a resource monitor, a dynamic collaborative subset manager, a global aggregator, and an encryption / decryption module;

[0080] The resource monitor is used to collect and store the network bandwidth of each device and the time required to perform one round of local training;

[0081] The dynamic collaboration manager is used to calculate the gradient similarity matrix between devices and to build the required collaboration subset and the assistance collaboration subset.

[0082] The global aggregator is used to execute dynamic obsolescence compensation strategies to update the global model;

[0083] The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, and supports protocols such as AES-256 / SM4.

[0084] The client module includes: a resource detector, a local training engine, a collaborative gradient aggregator, and an encryption / decryption module;

[0085] The resource monitor is used to locally measure network bandwidth and the time required for a round of local training, and reports it to the server after signing it using the encryption and decryption module;

[0086] The local training engine performs local model training, calculates gradients, and uploads the gradients to the server and the support subset.

[0087] The collaborative gradient aggregator is used to implement two-stage aggregation. The first stage uses the global model from the server to perform interpolation aggregation, and the second stage uses gradients from the required collaborative subset to perform weighted aggregation.

[0088] The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, and supports protocols such as AES-256 / SM4.

[0089] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A federated learning method for addressing the heterogeneity of edge devices in power systems, characterized in that, Includes the following steps: S1. The Federation Center server initializes the global model and distributes it to all participating devices; S2, The server collects the available bandwidth of each device. Time required for a round of local training ; S3, the server constructs the required collaboration subset and the support collaboration subset for each device; S4. After completing one round of local training, the device encrypts and uploads the gradients to the server, and simultaneously performs training based on the collaborative subset. The gradient is distributed to the devices within it, either directly or via a server. S5. After receiving the gradient uploaded by the device, the server sends the global model parameters to the device and then updates the global model using a dynamic staleness compensation strategy. S6. After receiving the global model parameters, the device combines the gradient information of the required collaborative subset received during this round of training and performs two-stage aggregation. S7. When the global aggregation count on the server reaches... At that time, the federal learning and training will end; In step S3, the server constructs the required collaboration subset and the support collaboration subset for each device, specifically including the following steps: S31. Based on the latest gradient information uploaded by each device collected by the server, calculate the gradient similarity matrix between devices. The gradient similarity between device i and device j , Let i be the model gradient of device i. Let be the model gradient of device j; S32. Calculate the transmission efficiency of each device, and the transmission efficiency of device i. Among them, molecules Indicates equipment The theoretical amount of data that can be transmitted during training is bandwidth × time, reflecting the actual communication capability, where the denominator is... This represents the theoretical requirement for gradient transfer. Let i be the model gradient memory for device i. Since all device models are the same, the model gradient memory is of the same size. S33, Based on transmission efficiency and number of participating devices Determine the number of required collaborative subsets for device i. ; S34. Selection and Equipment Least similar Each device constitutes a subset of the collaborative needs of device i. ; S35. Derive the assistance collaboration subset for each device from the demand collaboration subset of all devices. That is, which equipment's demand collaboration subset includes that equipment; S36, If the equipment The subset of assistance and cooperation Size greater than Then retain the least similar Each device, and notify the relevant devices to update their required collaboration subset; The required collaboration subsets and assistance collaboration subsets for each device are dynamically constructed, and the federal center server executes them each time. Each global model update involves constructing a subset of complementary needs collaboration among participating devices. ; In step S5, a dynamic staleness compensation strategy is used to update the global model, which specifically includes the following steps: S51, the Federation Center server comes from the device. gradient , global model parameters Transmitted to device ; S52, Update Global Model Its learning rate The calculation method is as follows: ,in , The total number of times the global model is updated. This represents the number of times the current global model has been updated. For equipment The number of times gradients were uploaded to the federated central server. The global model learning rate. For weight hyperparameters, The learning rate Simultaneously, dynamic adjustments are made based on the participation frequency: when the device participates in the frequency... At that time, the amplification compensation coefficient is increased to ;when When the attenuation compensation coefficient reaches ; The reason why the server first transmits the global model to the device that uploaded the model and then performs gradient aggregation is that the device will perform local aggregation based on the global model it sends out, thus avoiding duplicate aggregation.

2. The federated learning method for addressing the heterogeneity of power system edge devices according to claim 1, characterized in that, In step S2, the available bandwidth of each device Time required for a round of local training After the device completes one round of local training, it collects the device's current available bandwidth and the time required to complete this round of training, and then encrypts and sends it to the server. Take the moving average of the most recent 3 rounds.

3. The federated learning method for addressing the heterogeneity of power system edge devices according to claim 1, characterized in that, In step S4, gradient encryption transmission assists the collaborative subset, and the device... gradient Perform the following operations using the encryption / decryption module: 1) Direct transmission: Send to [the specified location] using end-to-end encryption. 1) Devices in the middle; 2) Server relay: After decryption by the central server's encryption / decryption module, it is re-encrypted and forwarded to the device. The equipment in the middle.

4. The federated learning method for addressing the heterogeneity of power system edge devices according to claim 1, characterized in that, In step S6, the device performs two-stage aggregation locally, specifically including the following steps: S61. Perform the first-stage interpolation aggregation update with the global model: ;in These are the interpolation weight hyperparameters of the global model, which gradually decrease as training progresses. S62, and the gradient set of the required collaboration subset received during this round of training. Perform the second phase of weighted aggregate update. , , ;in It is the local learning rate of device i. It is the similarity between device i from the federated center server and device k in its demand collaboration subset. The smaller the value, the larger the gradient weight corresponding to device k, and the stronger the complementarity.

5. An asynchronous federated learning system based on dynamic gradient complementary cooperation, implemented using the federated learning method for heterogeneity of power system edge devices as described in any one of claims 1-4, characterized in that, Includes: a central server module and a client module; The central server module includes: a resource monitor, a dynamic collaborative subset manager, a global aggregator, and an encryption / decryption module; The resource monitor is used to collect and store the network bandwidth of each device and the time required to perform one round of local training; The dynamic collaboration manager is used to calculate the gradient similarity matrix between devices and to build the required collaboration subset and the assistance collaboration subset. The global aggregator is used to execute dynamic obsolescence compensation strategies to update the global model; The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, supporting protocols such as AES-256 / SM4; The client module includes: a resource detector, a local training engine, a collaborative gradient aggregator, and an encryption / decryption module; The resource monitor is used to locally measure network bandwidth and the time required for a round of local training, and reports it to the server after signing it using the encryption and decryption module; The local training engine performs local model training, calculates gradients, and uploads the gradients to the server and the support subset. The collaborative gradient aggregator is used to implement two-stage aggregation. The first stage uses the global model from the server to perform interpolation aggregation, and the second stage uses gradients from the required collaborative subset to perform weighted aggregation. The encryption / decryption module provides encryption and decryption services for gradient or parameter transmission, and supports protocols such as AES-256 / SM4.

Images