Federated learning method, system and device based on contract theory incentive mechanism

Abstract

The present invention discloses a method, system and device for federated learning based on a contract theory incentive mechanism, the method comprising: obtaining initial model parameters; receiving a local model trained by a terminal device; aggregating the local model and sending the aggregated local model to other access points to update the global model parameters; sending the updated global model to the terminal device and performing iterative learning, judging that the iteration completion condition is met; and sending the final model to the edge cloud server. The system comprises: an access point, an edge server and a terminal device. The device comprises a memory and a processor for executing the above-mentioned federated learning method based on a contract theory incentive mechanism. By using the present invention, the global aggregation task of the edge cloud server is distributed to all access points connected thereto, which can alleviate the pressure of the edge cloud server in processing local models. The present invention can be widely applied to the field of federated learning.

Description

Technical Field

[0001] The present invention relates to the field of federated learning, and in particular to a federated learning method, system and device based on a contract theory incentive mechanism. Background Art

[0002] The Industrial Internet of Things is considered a revolutionary way to improve industrial production efficiency. In the Industrial Internet of Things, a large number of intelligent terminals are deployed to monitor and perceive the environment. The data collected by the intelligent terminals will be sent to the data center for analysis and processing to support intelligent decision-making and improve manufacturing accuracy. In this scenario, digital twin technology can effectively solve problems such as quality monitoring and fault detection by creating virtual twin models of physical elements (such as industrial equipment, products or services). Specifically, various mission-critical operations must be tested, analyzed, predicted and optimized in the twin model before they are implemented in the actual production process. For example, the twin model of a machine can accurately predict fault behavior in the production process before it affects equipment efficiency and output.

[0003] The construction of twin models is the basis for the application of digital twin technology in the field of industrial Internet of Things. In order to build twin models of physical elements, it is necessary to integrate data from various sources (such as sensor data, historical data, and simulation data), conduct comprehensive data analysis, and obtain an accurate description of the actual situation. In recent years, digital twin edge networks have become a new paradigm for deploying twin models at the edge of factory networks (such as access points AP, base stations, and factory gateways). In digital twin edge networks, edge servers create twin models locally based on incoming data. In addition, edge servers update the twin models at the local edge by constantly interacting with physical entities. Compared with cloud-based digital twins (where heterogeneous data is collected and processed by centralized servers for efficient data utilization), digital twin edge networks can bring flexible twin model construction for performance optimization in the production process. In edge networks, data is shared by edge cloud servers to computing devices managed by them. During the construction of edge twin models, malicious computing devices may exist to steal some sensitive privacy data, causing information and privacy security issues.

[0004] Traditional machine learning requires all terminal data to be collected to the main server, and then the main server is used to train the model. Since terminal users have sensitive data (such as personal privacy data), the processing of terminal data by a third party may lead to data leakage and cause security issues. Therefore, many terminals attach importance to the ownership and use rights of data, which may lead to data silos. Federated learning, as a new machine learning method, can use data for machine learning modeling while meeting the data privacy and security requirements of all parties. In the federated learning process, each participant does not need to upload their own data to the central server. They only need to use their own data for machine learning training to obtain relevant model parameters, intermediate data, etc., and upload these parameters to the central server. The central server aggregates the parameters or intermediate data from each federated learning (FL) terminal to obtain model data and feeds it back to each terminal. Each terminal updates its own model based on the feedback information. Federated learning allows each participant to jointly model without sharing data, which can effectively solve the data silo problem. Traditional federated learning alleviates the pressure on edge cloud servers to process data to a certain extent by distributing tasks to local processing. However, in the process of building edge digital twin models, the edge cloud servers cannot process them simultaneously due to the large number of local models trained by terminal devices, resulting in a long time in the end. Summary of the invention

[0005] In the process of building edge digital twins based on federated learning and blockchain, there is a lack of an incentive mechanism to motivate the FL terminal devices and DAG blockchain verification nodes that work diligently, and the central server of federated learning does not have enough understanding of the performance and modeling willingness of each terminal involved in the modeling process. It is necessary to design an incentive mechanism under the condition of information asymmetry to increase the enthusiasm of FL terminal devices and DAG blockchain verification nodes and improve the efficiency of edge digital twin construction.

[0006] In view of this, in order to solve the problem of low efficiency caused by the task of aggregating local models by edge cloud servers in existing federated learning methods, the present invention proposes a federated learning method based on a contract theory incentive mechanism, the method comprising, applied to an access point:

[0007] Obtain initial model parameters and send them to the terminal device;

[0008] Receiving the local model trained by the terminal device;

[0009] Aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model;

[0010] The updated global model is sent to the terminal device for iterative learning, and the final model is obtained when the iteration completion conditions are met;

[0011] The final model is sent to the edge cloud server.

[0012] Through this step, a semi-distributed federated learning framework is constructed, which distributes the global aggregation tasks of the edge cloud server to all access points connected to it, which can alleviate the pressure of the edge cloud server in processing local models.

[0013] In some embodiments, it also includes:

[0014] Verify the trained local model through the DAG blockchain network to obtain the verification result;

[0015] After verification, the trained local model is sent to the access point;

[0016] Save the verification result to the terminal device.

[0017] Through this optimization step, the DAG blockchain technology is introduced. Only the verified local model can be sent to the access point by the terminal device, eliminating the unreliable local model parameters provided by malicious terminals, ensuring the credibility of the local model parameters and the accuracy of the global model aggregated by the access point. The DAG blockchain provides a safe and reliable environment for the construction of edge twin models.

[0018] In some embodiments, the processing steps of the edge server are also included, which specifically include:

[0019] Release learning tasks and initialize model parameters;

[0020] After receiving the final model from the access point, the terminal device and the DAG blockchain network are incentivized according to the contract theory.

[0021] Through this optimization step, an incentive mechanism is introduced to improve the enthusiasm of terminal devices and verification nodes, thereby improving the work efficiency of federated learning.

[0022] In some embodiments, the incentives for the terminal device and the DAG blockchain network based on the contract theory specifically include:

[0023] Based on an iterative learning process, terminal devices are classified, contracts are provided for the classified terminal devices, and the trained local model is evaluated to obtain a first evaluation result;

[0024] Rewarding the classified terminal devices according to the first evaluation result;

[0025] Based on the iterative learning process, the verification nodes in the DAG blockchain network are classified, contracts are provided for the classified verification nodes, and the verification results are evaluated to obtain a second evaluation result;

[0026] The classified verification nodes are rewarded according to the second evaluation results.

[0027] In some embodiments, the classification indicators of the terminal device include the remaining power and the computing resources occupied for iterative learning, and the classification indicators of the verification node include the CPU cycles and energy consumption for verifying the local model.

[0028] In some embodiments, the processing step of the edge server further includes:

[0029] Based on the rewards and costs, define the utility function of the terminal device, the utility function of the verification node, and the utility function of the edge server.

[0030] The present invention also proposes a federated learning system based on a contract theory incentive mechanism, the system comprising an access point, an edge server and a terminal device:

[0031] The access point is used to:

[0032] Obtain initial model parameters and send them to the terminal device;

[0033] Receiving the local model trained by the terminal device;

[0034] Aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model;

[0035] The updated global model is sent to the terminal device for iterative learning, and the final model is obtained when the iteration completion conditions are met;

[0036] Send the final model to the edge cloud server;

[0037] The edge server is used for:

[0038] Release learning tasks and initialize model parameters;

[0039] Receive the final model from the access point and incentivize the terminal device and DAG blockchain network based on the contract theory;

[0040] The terminal device is used for local training of the model.

[0041] The present invention also proposes a federated learning device based on a contract theory incentive mechanism, comprising:

[0042] at least one processor;

[0043] at least one memory for storing at least one program;

[0044] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned federated learning method based on the contract theory incentive mechanism.

[0045] Based on the above scheme, the present invention provides a federated learning method, system and device based on the contract theory incentive mechanism. Through federated learning, each client device trains a local model according to its locally generated data, and uploads the local model parameters to the edge server. The edge server aggregates the collected local model parameters to build an edge twin model to ensure the data security of the client. At the same time, the DAG blockchain technology is introduced to provide a safe and reliable environment for the construction of the edge twin model in the FL process. Finally, the contract theory is used to incentivize the FL terminal device and the DAG blockchain verification node, improve work efficiency, and build the edge digital twin model more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the steps of a federated learning method based on the contract theory incentive mechanism of the present invention;

[0047] Figure 2 It is an application architecture diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0048] The Industrial Internet of Things is considered a revolutionary way to improve industrial production efficiency. In the Industrial Internet of Things, a large number of intelligent terminals are deployed to monitor and perceive the environment. The data collected by the intelligent terminals will be sent to the data center for analysis and processing to support intelligent decision-making and improve manufacturing accuracy. In this scenario, digital twin technology can effectively solve problems such as quality monitoring and fault detection by creating virtual twin models of physical elements (such as industrial equipment, products or services). Specifically, various mission-critical operations must be tested, analyzed, predicted and optimized in the twin model before they are implemented in the actual production process. For example, the twin model of a machine can accurately predict fault behavior in the production process before it affects equipment efficiency and output.

[0049] The construction of twin models is the basis for the application of digital twin technology in the field of industrial Internet of Things. In order to build twin models of physical elements, it is necessary to integrate data from various sources (such as sensor data, historical data, and simulation data), conduct comprehensive data analysis, and obtain an accurate description of the actual situation. In recent years, digital twin edge networks have become a new paradigm for deploying twin models at the edge of factory networks (such as access points AP, base stations, and factory gateways). In digital twin edge networks, edge servers create twin models locally based on incoming data. In addition, edge servers update the twin models at the local edge by constantly interacting with physical entities. Compared with cloud-based digital twins (where heterogeneous data is collected and processed by centralized servers for efficient data utilization), digital twin edge networks can bring flexible twin model construction for performance optimization in the production process. In edge networks, data is shared by edge cloud servers to computing devices managed by them. During the construction of edge twin models, malicious computing devices may exist to steal some sensitive privacy data, causing information and privacy security issues.

[0050] Federated learning (FL) is a distributed machine learning method. In the FL process, each client device trains a local model based on its locally generated data and uploads the local model parameters to the master node (such as the edge server) instead of sending the original data, which means that FL naturally has good privacy protection.

[0051] This solution can be applied to the construction of twin models. In the process of building edge digital twins based on federated learning and blockchain, there is a lack of an incentive mechanism to incentivize the FL terminal devices and DAG blockchain verification nodes that work diligently. In addition, the central server of federated learning does not have enough understanding of the performance and modeling willingness of each terminal involved in the modeling process. It is necessary to design an incentive mechanism under the condition of information asymmetry to increase the enthusiasm of FL terminal devices and DAG blockchain verification nodes and improve the efficiency of edge digital twin construction.

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0054] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0055] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0056] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0057] Reference Figure 1 and Figure 2 , which is a flow chart of an optional example of a federated learning method based on a contract theory incentive mechanism proposed in the present invention. The method can be applied to a computer device. The federated learning method proposed in this embodiment may include but is not limited to the following steps, which are applied to an access point:

[0058] Step S101, obtaining initial model parameters and sending them to the terminal device;

[0059] All APs connected to the edge cloud server first download the initial model parameters from the edge cloud server And select the terminal device it manages to send the model parameters, and the terminal device performs FL.

[0060] Step S102, receiving the local model trained by the terminal device;

[0061] After the local training of FL is completed, the FL terminal device sends the obtained local model to the AP.

[0062] Step S103: aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model;

[0063] The AP aggregates the local model and sends the aggregated model to other APs to update the global model parameters.

[0064] Step S104: Send the updated global model to the terminal device and perform iterative learning, and determine whether the iteration completion condition is met to obtain the final model;

[0065] Step S105: Send the final model to the edge cloud server.

[0066] The AP sends the updated global model to the FL terminal for a new round of FL. After the FL task is completed, the AP sends the latest global model to the edge cloud server to complete the construction of the twin model.

[0067] Specifically, a semi-distributed federated learning framework is introduced to distribute the global aggregation tasks of the edge cloud server to all APs connected to it, which alleviates the computing pressure of the edge cloud server. Large-scale FL terminal devices use locally stored physical space perception data to perform local training tasks related to twin model construction to obtain local models. At the same time, each AP collects the local model parameters trained by the FL terminal devices it manages and aggregates them into global model parameters. Through multiple iterative training of global model parameters and local model parameters of all parties, the construction task of the edge digital twin model is completed without interacting with the original data, and the final twin model is sent to the edge cloud server, which alleviates the computing pressure of the edge cloud server and realizes the protection of the privacy and security of physical space information.

[0068] In one possible implementation, it also includes:

[0069] S106. Verify the trained local model through the DAG blockchain network to obtain a verification result;

[0070] S107, after verification, the trained local model is sent to the access point;

[0071] S108. Save the verification result to the terminal device.

[0072] The local model trained by the FL terminal device must first be verified through the DAG blockchain network, and the verification results are saved locally in all FL terminals in the form of tangle.

[0073] Specifically, in the process of building an edge digital twin model based on federated learning, there may be malicious FL terminal devices that provide false local model parameters, resulting in a decrease in the accuracy of the global model aggregated by the AP. If there are a large number of malicious FL terminal devices in the edge digital twin network, the constructed edge digital twin model will lose its value. The DAG blockchain provides a safe and reliable environment for federated learning. The local model trained by the FL terminal device needs to be verified by the DAG blockchain node. The local model parameters that pass the verification will be saved locally in the form of Tangle, while the local model of the malicious FL terminal device will not pass the verification. The introduction of DAG blockchain technology ensures the reliability of local model parameters and improves the accuracy of global model aggregation.

[0074] The updated local model of FL is verified by DAG and cumulative weight consensus scheme. The blockchain network based on DAG directed acyclic graph can receive transactions through asynchronous communication, which significantly improves the throughput transaction rate of the blockchain. Among them, the unverified transactions in the DAG network are called tip transactions (Tip), and all transactions form a tangle. The more transactions involved in the network, the faster the new transactions are received. The updated local model enters the DAG network, and the Markov Chain Monte Carlo (MCMC) algorithm is used to determine the Tip to be verified. In the MCMC algorithm, two independent random walk particles are generated on the Tangle, starting from the genesis transaction, and moving to the end of the Tangle through the verification relationship between transactions. The choice of path is determined by the transition probability:

[0075]

[0076] In the above formula, x and y are two transactions in the DAG network, respectively. The transfer probability applies to all transactions z in the network. H represents the cumulative weight of the transaction, and α is the expansion coefficient of the difference between the cumulative weights of the two transactions. When the tip transaction Tip is reached, the walking process stops, and then the tip transaction is selected to be verified by the new transaction.

[0077] The cumulative weight of each transaction in the network is equal to its own weight plus the sum of the weights of all transactions that directly or indirectly verify the transaction. When the cumulative weight of transactions in the DAG network exceeds a certain threshold, the transaction is finally confirmed. The confirmed results will be saved in all terminal devices and sent to the base station via wireless transmission to form a Tangle for local model updates. Local models that have not been finally confirmed for a long time will be regarded as outdated models. Outdated models will affect the accuracy of global model aggregation and will be discarded by the DAG network.

[0078] In a possible implementation manner, the processing steps of the edge server are also included, which specifically include:

[0079] S201, publish learning tasks and initialize model parameters;

[0080] The FL task publisher (i.e., edge server) publishes tasks.

[0081] S202. After receiving the final model from the access point, the terminal device and the DAG blockchain network are incentivized according to the contract theory.

[0082] In one possible implementation, the incentives for the terminal device and the DAG blockchain network based on the contract theory specifically include:

[0083] S2021. Based on an iterative learning process, classify the terminal devices, provide contracts for the classified terminal devices, evaluate the trained local model, and obtain a first evaluation result;

[0084] The FL terminal devices are classified according to their preferences. A group of FL terminal devices can be represented by M = {M i , M2, ..., M m}. FL terminal devices have different preferences in terms of remaining power and occupied computing resources for participating in FL transactions. The M FL terminal devices with heterogeneous preferences are divided into S categories and sorted in ascending order: θ1<...<θ s ...<θ S , s∈{1,...,S}. θ s The larger it is, the more willing the FL terminal device is to contribute to FL and will receive higher rewards.

[0085] Evaluate the quality of the model trained on the FL terminal device. For the FL terminal device, in order to quantitatively evaluate the quality of the model trained on the FL terminal device, we define is the model training value (VoMT), where A MT represents the initial VoMT, B MT is the decay rate of VoMT, t m is the delay imposed by the mth FL end device on model training. VoMT decays exponentially with the model training delay, which means that the fastest trained model has the highest VoMT because the end user can receive the required data on time.

[0086] S2022. Reward the classified terminal devices according to the first evaluation result;

[0087] FL task publisher provides contracts for s-type FL terminal devices (R MT ,s,V MT, s), where RMT,s is the computational cost for s-type FL terminal devices. The higher the value of model training, the higher the reward for FL terminal devices.

[0088] S2023. Based on an iterative learning process, classify the verification nodes in the DAG blockchain network, provide contracts for the classified verification nodes, evaluate the verification results, and obtain a second evaluation result;

[0089] Nodes are classified according to their preferences. A set of DAG blockchain verification nodes can be N = {N1, N2, ..., N n}. The preferences of DAG blockchain verification nodes for local model verification vary in terms of CPU cycles, energy consumption, etc. N DAG blockchain verification nodes with heterogeneous preferences are divided into Q categories and sorted in ascending order: ψ1<...<ψ q ...<ψ Q ,q∈{1,...Q}.ψ q Large, indicating that the DAG blockchain verification nodes are more willing to contribute to the verification of the local model and will receive higher rewards.

[0090] Evaluate the verification quality of DAG blockchain verification nodes. For DAG block verification nodes, in order to quantify the effect of local model verification, define is the verification value (VoMV) of the local model. MV represents the initial VoMV, B MV is the decay rate of VoMV, t n The delay incurred by verifying the local model for the nth DAG validator. VoMV decays exponentially with the verification delay, meaning that the transactions that are verified the fastest have the highest VoMV.

[0091] S2024. Reward the classified verification nodes according to the second evaluation result.

[0092] FL task issuer provides contract for q-type verification equipment (R MV，q , V MV，q ), R MV，q is the verification fee for the q-type verification device. The faster the local model is verified, the higher the reward for the DAG blockchain verification node.

[0093] 1) In the process of motivating FL terminal devices, using contract theory, FL terminal devices are first divided into s categories according to the differences in the remaining power of participating in FL transactions, the computing resources occupied, etc., and then the quality of the model trained by the FL terminal devices is evaluated. The parameters of the fastest trained local model are the most timely, so the model has the best quality and the highest value. The FL task publisher provides a contract (R MT ,s,V MT ,s), the higher the value of model training, the higher the reward for FL terminal devices. 2) In the process of incentivizing DAG blockchain verification nodes, DAG blockchain verification nodes are first divided into q categories according to their differences in CPU cycles, energy consumption, etc. of the local model being verified, and then the results of local model verification are quantified, and the node with the fastest verification transaction has the highest verification value. The FL task publisher provides a contract (R MV，q , V MV，q ), the faster the DAG blockchain node verifies the model, the higher the reward. Therefore, contract theory can promote FL terminal devices and DAG blockchain verification nodes in the network to participate more actively in tasks, making the construction and verification of edge twin network models more efficient.

[0094] In a possible implementation, the processing step of the edge server further includes:

[0095] S203. Define the utility function of the terminal device, the utility function of the verification node, and the utility function of the edge server based on the reward and the cost.

[0096] Construct the utility function of the FL terminal device. The contract has been signed (R MT ,s,V MT , s) is defined as:

[0097] U FL (s) = θ s v1(R MT ,s)-ε1V MT ,s

[0098] The first term represents the reward for FL model training, v1(R MT , s) is the FL model training cost R MT , a monotonically increasing concave function of s; the second term is to provide model training V MT , the computing resource cost of the required value of s, and ε1 is the unit computing resource cost. FL (s)<0, the FL terminal device will not accept the contract and choose not to participate in FL transactions.

[0099] Construct the utility function of the DAG blockchain node. During the verification process of the local model, the signed contract (R MV，q , V MV，q ) is defined as:

[0100] U verifier (q)=ψ q v2(R MV，q )-ε2V MV，q

[0101] The first term represents the reward obtained by local model verification, v2(R MV，q ) is a monotonically increasing concave function. The second term provides the value V required for local model verification MV，q The backtracking cost (CPU and bandwidth cost) is ε2, and the unit backtracking cost is ε2. verifier (q)<0, the verifier will not accept the contract, that is, choose not to participate in the local model verification, defined as (R MV，q =0, V MV，q =0).

[0102] Benefits of FL task publishers. For FL task publishers who use s-type FL terminal devices for FL model training and q-type DAG blockchain verification nodes for verification, U publish (s, q) = V publish (s, q)-R publish (s, q) is the utility function of the FL task publisher. publish (s, q) is the service requirement of the FL task publisher (the value created by s, q), and is equal to V MT，s With V MV，q The sum of V MT，s is the model training value (VoMT) value, V MV，q is the model validation value (VoMV); R publish (s, q) is the remuneration that the FL task publisher needs to pay when the service requirements are met, and R publish (s,q)=l1u1R MT，s +l2u2R MV，q , where u1 is the unit cost of the FL task publisher for local model training, and u2 is the unit cost of the FL task publisher for local model parameter verification. l1 is the model training fee weight, l2 is the block verification fee weight, and l1+l2=1.

[0103] In summary, the utility function of the FL task publisher can be expressed as:

[0104]

[0105] where λ sis the distribution probability of S-type FL terminal equipment, λ q is the distribution probability of Q-type block validators. The main goal of FL task publishers is to maximize their expected utility while meeting the contract requirements of transaction FL terminal devices and DAG verification. In summary, the optimization goal can be constructed as:

[0106]

[0107] st

[0108] (a)θ s v1(R MTs )-ε1V MTs ≥0

[0109] (b)ψ q v2(R MV，q )-ε2V MV，q ≥0,

[0110] (c)θ s v1(R MTs )-ε1V MT，s ≥θ s v1(R MTs′ )-ε1V MTs′ ,

[0111] (d)ψ q v2(R MV，q )-ε2V BV，q ≥ψ q v2(R MV，q′ )-ε2V MV，q′ ,

[0112]

[0113]

[0114] Among them, (a) and (b) are the IR (Individual Rationality) individual rationality constraints for FL model training and DAG blockchain verification, respectively. The IR constraint means that FL terminal devices and DAG verification nodes are rational and will choose the contract that brings them the greatest benefit. (c) and (d) are the IC (Incentive Compatible) incentive compatibility constraints for FL model training and block verification, respectively. The IC constraint means that FL terminal devices and DAG verification nodes will choose contracts designed for their own types instead of other contracts. (e) Ensure that the FL model training fee and transaction verification fee will not exceed the maximum reward R provided by the FL task publisher max (f) is the weight constraint of model training fee and transaction verification fee.

[0115] In general, in the digital edge twin network, first, the FL task publisher, that is, the edge server, sends the task to all the APs connected to it through wireless transmission, and the AP sends the task to the FL terminal devices it manages for local model training to obtain the local model. Then, the FL terminal device sends the local model to the DAG blockchain verification node through wireless transmission for local model verification. The model parameters after verification will be saved in the form of Tangle locally on each DAG verification node. Finally, the verified local model parameters will be sent by the FL terminal device to the AP, and the AP will aggregate the global model. The AP sends the updated global model to other APs to update the global model parameters and sends it to the FL terminal devices for a new round of FL. After the FL task is completed, the AP transmits the constructed edge twin model to the edge cloud server through wireless communication. Finally, based on the contract theory, the FL task publisher provides contract incentives for the FL terminal devices and DAG blockchain verification nodes that contribute to the FL task.

[0116] A federated learning system based on contract theory incentive mechanism, including access points, edge servers and terminal devices:

[0117] The access point is used to:

[0118] Obtain initial model parameters and send them to the terminal device;

[0119] Receiving the local model trained by the terminal device;

[0120] Aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model;

[0121] The updated global model is sent to the terminal device for iterative learning, and the final model is obtained when the iteration completion conditions are met;

[0122] Send the final model to the edge cloud server;

[0123] The edge server is used for:

[0124] Release learning tasks and initialize model parameters;

[0125] Receive the final model from the access point and incentivize the terminal device and DAG blockchain network based on the contract theory;

[0126] The terminal device is used for local training of the model.

[0127] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0128] A federated learning device based on contract theory incentive mechanism:

[0129] at least one processor;

[0130] at least one memory for storing at least one program;

[0131] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned federated learning method based on the contract theory incentive mechanism.

[0132] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0133] A storage medium stores processor-executable instructions, which, when executed by a processor, are used to implement a federated learning method based on a contract theory incentive mechanism as described above.

[0134] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0135] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A federated learning method based on the contract theory incentive mechanism, characterized by: Applicable to access points, including: Obtain initial model parameters and send them to the terminal device; Receiving the local model trained by the terminal device; Aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model; The updated global model is sent to the terminal device for iterative learning, and the final model is obtained when the iteration completion conditions are met; Send the final model to the edge cloud server; The processing steps of the edge server are also included, which specifically include: Release learning tasks and initialize model parameters; Receive the final model from the access point and incentivize the terminal device and DAG blockchain network based on the contract theory; The incentives for terminal devices and DAG blockchain networks based on contract theory specifically include: Based on an iterative learning process, terminal devices are classified, contracts are provided for the classified terminal devices, and the trained local model is evaluated to obtain a first evaluation result; Rewarding the classified terminal devices according to the first evaluation result; Based on the iterative learning process, the verification nodes in the DAG blockchain network are classified, contracts are provided for the classified verification nodes, and the verification results are evaluated to obtain a second evaluation result; The classified verification nodes are rewarded according to the second evaluation results.

2. A federated learning method based on contract theory incentive mechanism according to claim 1, characterized in that: Also includes: Verify the trained local model through the DAG blockchain network to obtain the verification result; After verification, the trained local model is sent to the access point; Save the verification result to the terminal device.

3. A federated learning method based on contract theory incentive mechanism according to claim 2, characterized in that: The classification indicators of terminal devices include the remaining power and computing resources occupied for iterative learning, and the classification indicators of verification nodes include the CPU cycles and energy consumption for verifying local models.

4. A federated learning method based on contract theory incentive mechanism according to claim 3, characterized in that: The processing steps of the edge server also include: Based on the rewards and costs, define the utility function of the terminal device, the utility function of the verification node, and the utility function of the edge server.

5. A federated learning system based on contract theory incentive mechanism, characterized in that: Including access points, edge servers and terminal devices: The access point is used to: Obtain initial model parameters and send them to the terminal device; Receiving the local model trained by the terminal device; Aggregate the local models and send the aggregated local models to other access points to update the global model parameters to obtain an updated global model; The updated global model is sent to the terminal device for iterative learning, and the final model is obtained when the iteration completion conditions are met; Send the final model to the edge cloud server; The edge server is used for: Release learning tasks and initialize model parameters; Receive the final model from the access point and incentivize the terminal device and DAG blockchain network based on the contract theory; The terminal device is used for local training of the model; The incentives for terminal devices and DAG blockchain networks based on contract theory specifically include: Based on an iterative learning process, terminal devices are classified, contracts are provided for the classified terminal devices, and the trained local model is evaluated to obtain a first evaluation result; Rewarding the classified terminal devices according to the first evaluation result; Based on the iterative learning process, the verification nodes in the DAG blockchain network are classified, contracts are provided for the classified verification nodes, and the verification results are evaluated to obtain a second evaluation result; The classified verification nodes are rewarded according to the second evaluation results.

6. A federated learning device based on contract theory incentive mechanism, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a federated learning method based on a contract theory incentive mechanism as described in any one of claims 1 to 4.

Images