Decentrainment joint training system capable of ensuring credibility and high efficiency of data

Abstract

The invention discloses a decentration joint training system capable of ensuring data credibility and high efficiency. In the system, a sensor in a sensor Internet of Things layer collects original data of a client training node; each client training node of the node calculation layer is responsible for preprocessing corresponding original data and pre-training a local model, generating gradient statistics, encrypting and packaging the gradient statistics into transaction records, and submitting the transaction records to the block chain network; the block chain network adopts a consensus algorithm, manages a dynamic committee election process through an intelligent contract, screens out nodes based on a designed multi-dimensional incentive mechanism to form a temporary committee, and forms a federated learning aggregation layer; in the federated learning aggregation layer, each temporary committee member node completes weighted fusion of global statistical characteristics in a decentration mode, and writes the global statistical characteristics into a distributed account book, and then the global statistical characteristics are automatically distributed to each edge node by using an intelligent contract to trigger local model updating. According to the invention, the communication and storage overhead is effectively reduced, and the data security is improved.

Description

Technical Field

[0001] This invention relates to the field of blockchain technology, and more specifically, to a decentralized joint training system that ensures data trustworthiness and efficiency. Background Technology

[0002] Traditional centralized learning models struggle to balance the diversity and privacy of multi-source data across enterprises, and cannot effectively address the challenges of balancing data diversity with complex communication environments. While different enterprises have accumulated massive amounts of data, their decentralized and closed management methods have created "data silos." A single-enterprise model may perform well on a single type of data, but it cannot fully reflect the diverse characteristics of different data types under complex operating conditions, failing to meet the accurate diagnostic needs across enterprises and scenarios. Federated learning offers a new solution for joint training of multi-source data, enabling joint modeling of dispersed data without disclosing the original data from each party, balancing data security and model performance. However, existing federated learning still faces challenges in practical applications, including single-point-of-failure risks, insufficient incentive mechanisms, and the need to improve model security and traceability.

[0003] Blockchain, as a decentralized data storage and processing technology based on cryptography, possesses characteristics such as immutability, traceability, and decentralization, providing secure and reliable distributed ledger and incentive mechanisms for federated learning. However, with the introduction of blockchain, the load on network communication and node collaboration also increases. For practical needs such as multi-source heterogeneous data and high-speed dynamic environments, how to achieve an efficient, low-communication-overhead federated learning blockchain system while ensuring data privacy and trustworthiness remains a bottleneck in current research.

[0004] In summary, federated learning, as a key technology for solving the "data silo" problem, achieves collaborative optimization of data privacy protection and model training through distributed collaboration. While the introduction of blockchain technology can enhance system robustness to address the single point of failure risk of the centralized architecture of traditional federated learning, it also introduces challenges such as increased communication load and data storage pressure. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a decentralized federated training system that ensures data reliability and efficiency. This system includes a sensor IoT layer, a node computing layer, and a federated learning aggregation layer built using a blockchain network, wherein:

[0006] The sensor IoT layer is equipped with multiple sensors, each responsible for collecting raw data from the corresponding client training node;

[0007] The node computing layer comprises multiple client training nodes. Each client training node acts as a computing node, responsible for the preprocessing of the corresponding raw data and local model pre-training, generating gradient statistics, which are then encrypted and encapsulated into transaction records and submitted to the blockchain network. The blockchain employs a consensus algorithm, managing a dynamic committee election process through smart contracts. Based on a designed multi-dimensional incentive mechanism, nodes are selected to form a temporary committee, creating the federated learning aggregation layer to replace the federated learning center server. The gradient statistics include mean and variance.

[0008] In the federated learning aggregation layer, each temporary committee member node completes the weighted fusion of global statistical features in a decentralized manner. The global statistical features are written into the distributed ledger through asynchronous consensus, and then automatically distributed to each edge node using smart contracts, triggering local model updates.

[0009] Compared with existing technologies, the advantages of this invention are that it designs an efficient and reliable joint training system, constructs a Gaussian conditional probability framework, guides global representation learning by constructing a generative classifier, and designs a dynamic interpolation mechanism to balance the bias-variance tradeoff, which significantly improves the model performance of heterogeneous clients in scenarios with scarce data and distribution offset. Furthermore, it replaces the on-chain storage of complete model parameters with the on-chain storage of key statistics, reducing the storage and communication volume of the blockchain and improving the operating efficiency.

[0010] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0012] Figure 1 This is an architecture diagram of a decentralized joint training system that ensures data reliability and efficiency according to an embodiment of the present invention. Detailed Implementation

[0013] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0014] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0015] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0016] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0017] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0018] To address the significant performance bottlenecks of traditional federated learning when dealing with non-independent and identically distributed (non-iID) client data, and the high communication overhead associated with simply combining federated learning with blockchain, this invention provides a personalized federated learning method based on adaptive feature distribution combined with blockchain. This method guides global representation learning by constructing a generative classifier, designs a dynamic interpolation mechanism to balance bias-variance tradeoffs, significantly improving model performance under scenarios of data scarcity and distributional bias in heterogeneous clients, and replaces the on-chain storage of complete model parameters with on-chain storage of key statistics, reducing blockchain storage requirements and communication overhead, and improving operational efficiency.

[0019] In summary, this invention first constructs a trusted decentralized joint training system (or framework). This system mainly comprises three layers: a sensor IoT layer, a node computing layer, and a federated learning aggregation layer. In the sensor IoT layer, sensors are responsible for collecting raw data from each client training node. In the node computing layer, each client training node acts as a computing node, forming the node computing layer, responsible for preprocessing the raw data and pre-training the local model, generating gradient statistics containing mean and variance parameters, which are then encrypted and encapsulated as transaction records and submitted to the blockchain network. The blockchain employs an improved consensus algorithm, managing a dynamic committee election process through smart contracts. A multi-dimensional incentive mechanism selects nodes to form a temporary committee, creating the federated learning aggregation layer to replace the central federated learning server. Within the federated learning aggregation layer, each committee member node performs weighted fusion of global statistical features in a decentralized manner. Through asynchronous consensus, the global statistical features are written into a distributed ledger and automatically distributed to edge nodes using smart contracts, triggering local model updates and forming a closed-loop optimization mechanism of "distributed feature learning - trusted gradient aggregation - targeted model update".

[0020] Based on the designed trusted decentralized joint training system, this invention addresses the heterogeneity of client data in federated learning (non-i.i.d.) and the high overhead of blockchain communication. It guides global representation learning by constructing a generative classifier and combines it with a dynamic distribution interpolation mechanism to optimize the bias-variance tradeoff. Specifically, representation learning is reconstructed as a generative modeling task of global feature distribution. Under the reasonable assumption that the feature space follows a class-conditional Gaussian distribution, the parameters of the class-conditional distribution are estimated by aggregating the mean and covariance matrices of data features from each active node, forming a generative classifier. Bayesian linear interpolation is then performed between the estimated local feature distribution parameters and the aggregated global feature distribution parameters to personalize the model and balance the performance of local and global models. Blockchain storage of mean, variance statistics, and task performance scalars reduces transmission costs and improves the model's generalization and accuracy.

[0021] To alleviate the significant disparities in data quality and model training performance among enterprises due to differences in geographical environment, equipment conditions, and technical capabilities, one embodiment constructs a reputation evaluation system that includes multi-dimensional indicators such as model performance score and data quality score, providing a basis for adjusting the contribution weight of each enterprise and for committee elections.

[0022] In summary, this invention designs a reliable and efficient decentralized joint training system. This system is a decentralized joint training architecture that supports multi-party collaboration and is secure and reliable. Furthermore, it designs an efficient shared data transmission mechanism for the joint training framework, a multi-dimensional reputation-oriented incentive mechanism that integrates data quality and model training performance, and so on.

[0023] The following section will provide specific examples of the decentralized joint training architecture, the corresponding efficient shared data transmission mechanism, and the multi-dimensional reputation-oriented incentive mechanism.

[0024] 1. A decentralized joint training architecture that supports multi-party collaboration and is secure and reliable.

[0025] Figure 1This is an architecture diagram of a secure and reliable decentralized federated training system that supports multi-party collaboration. During system operation, sensors are distributed across various client training nodes, responsible for collecting raw data. For example, the sensor IoT layer can be configured with multiple sensors to collect different types of data, such as sound and images. Each client training node acts as a computing node, generating gradient statistics including mean and variance parameters, and submitting them to the blockchain network. The blockchain manages the dynamic committee election process through smart contracts, selecting nodes based on a multi-dimensional incentive mechanism to form a temporary committee, thereby constructing a federated learning aggregation layer, replacing the traditional federated learning center server. In the federated learning aggregation layer, committee member nodes perform weighted aggregation of global statistical features in a decentralized manner, and write the results to a distributed ledger through a consensus algorithm. Subsequently, with the help of smart contracts, updated data is automatically distributed to each edge node, triggering local model updates. This process is continuously iterated until training converges.

[0026] 2. Efficient shared data transmission mechanism of the joint training framework

[0027] The specific process of the efficient shared data transmission mechanism of the joint training framework is as follows:

[0028] ① Representation learning from the perspective of generative modeling

[0029] Assume the feature space follows a class-conditional Gaussian distribution (the probability distribution of data in each class is an independent Gaussian distribution). For class c, the global feature distribution is defined as:

[0030]

[0031] Train a shared feature extractor φ such that the features z = φ(x) of all nodes are aligned with the global distribution p(z|y).

[0032] Where, μ c ∈R d Let ∑ be the feature mean of the global class c. c ∈R d×d Let y be the feature covariance matrix of the global category c, and y represent the label. Let x represent a Gaussian distribution, x represent a sample, and d represent the feature dimension of the sample.

[0033] ② Construct a global generative classifier

[0034] The committee uses an incentive mechanism to calculate the contribution of each active node and the proportion of each category in the Gaussian distribution, and then weights and aggregates the global feature distribution parameters.

[0035]

[0036] Where M is the total number of active nodes. Let ω be the number of samples of category c in node i. i The contribution of node i to the federated aggregation is calculated through a multi-dimensional incentive mechanism. This represents the mean of category c in node i.

[0037] Estimate the parameters of the class-conditional distribution to form a generative classifier. For example, for input feature z, the classification decision is based on Bayes' theorem:

[0038]

[0039] Where c′ represents a possible category, used to enumerate all categories. Each candidate category in the list. μ represents the set of categories. c′ Let p(y=c′) represent the mean vector of category c′, p(y=c′) represent the prior probability of category c′, and p(y=c) be the global category prior, estimated by weighted average of the data statistics transmitted from each node.

[0040]

[0041] Under this Gaussian assumption, the classifier is generated as follows:

[0042]

[0043] This form is equivalent to the linear classifier Wz+b, where:

[0044]

[0045] This demonstrates that generative classifiers can seamlessly replace traditional fully connected layers.

[0046] ③ Estimate local characteristic distribution parameters

[0047] The local distribution p of node i i (z|y) may be offset from the global distribution p(z|y), therefore node i can utilize local data. Estimate the local class condition distribution:

[0048]

[0049] in, Let n represent the feature mean vector of category c. c Let j represent the number of samples belonging to category c, j represent the sample index, and n represent the total number of samples. This represents the eigenvalue matrix after centering. express The transpose and dimension transformation are used for covariance calculation. This represents the estimation of the overall characteristic covariance matrix.

[0050] when At that time, a regularization term is added to the covariance to achieve small sample correction:

[0051]

[0052] Where d represents the feature dimension. Let represent the covariance matrix of class c on client i after regularization. Let represent the covariance matrix of the local estimate, γ represent the regularization coefficient (e.g., γ = 0.01), Tr represent the trace operation (i.e., the sum of the diagonal elements of the covariance matrix), and I represent the identity matrix.

[0053] ④ Bayesian interpolation fusion

[0054] By interpolation coefficient (β) i Dynamically fuse global and local distributed parameters ∈[0,1]):

[0055]

[0056] in, This represents the final interpolated mean of category c on client i. This represents the interpolation covariance matrix of class c on client i. This represents the mean vector of class c estimated locally by the client. β represents the mean vector of class c in the global model. i Determined by minimizing the loss on the local validation set:

[0057]

[0058] in This is the interpolated classifier. Let (x, y) represent the local validation set used by client i for validation, and let (x, y) represent a sample pair in the validation set.

[0059] 3. A multi-dimensional reputation-oriented incentive mechanism that integrates data quality and model training performance.

[0060] The specific process of the multi-dimensional reputation-oriented incentive mechanism that integrates data quality and model training performance is as follows:

[0061] ① Calculate the task performance score (PS) i The formula is as follows:

[0062] PS i =ω1×Accuaracy i +ω2×Stability i +ω3×Efficiency i (14)

[0063] Where the subscript i represents the client number, ω1, ω2, and ω3 are the corresponding weights, and ω1 + ω2 + ω3 = 1. Accuracy is denoted as Accuracy. i Stability is measured by the reciprocal of the accuracy fluctuation over a certain period, and is denoted as Stability. i Task completion efficiency is expressed as the reciprocal of the ratio of actual completion time to expected completion time, denoted as Efficiency. i .

[0064] ② Calculate the Data Quality Score (DQS) i The formula is as follows:

[0065] DQS i = (1-PollutionFlag) i Integrity i (15)

[0066] If the mean and variance of the uploaded parameters deviate from the historical range or show abnormal peaks, PollutionFlag i It is recorded as 1 if it is not, and 0 otherwise. Additionally, the data integrity metric Integrity is introduced. i The value range is [0,1], where 1 indicates that the data is complete. For example, the data integrity index is set as follows:

[0067]

[0068] Where m represents the number of covered types, and M represents the total number of types.

[0069] ③ Calculate the Comprehensive Reputation Score (CRS) i The formula is as follows:

[0070] CRS i =α×PS i +β×DQS i (17)

[0071] The task performance score is weighted by α, the data quality score is weighted by β, and α+β=1.

[0072] ④ Select nodes with higher overall reputation scores as committee members to aggregate the global model, and then use a consensus method such as Dumbo-NG to efficiently and with low latency determine a reasonable global model.

[0073] ⑤ The overall reputation score is converted into model aggregation weights, giving higher aggregation weights to client training nodes with higher reputations. This ensures a fair and comprehensive evaluation of each node's contribution, incentivizing each node to actively participate and improve its model training performance and data quality. The formula is as follows:

[0074]

[0075] Where i represents the i-th client node, and j represents the index of all client nodes, used for summation in the normalized denominator.

[0076] The designed multi-dimensional reputation-oriented incentive mechanism generates reputation scores by quantifying model training performance and training data quality, guides committee member elections and model aggregation, and achieves positive feedback of contribution and reward, which can improve the enthusiasm of multi-node collaboration and global model performance.

[0077] The decentralized joint training system combining blockchain and federated learning provided by this invention can be applied to fields of cross-institutional data collaboration, such as financial services, healthcare, smart terminals, the Internet of Things (IoT), smart cities, and industrial manufacturing. It achieves multi-party joint modeling without sharing raw data, solving problems such as data silos and privacy leaks. For example, in predictive maintenance scenarios for industrial equipment, the system first collects raw data (such as vibration waveforms, temperature curves, and RMS current values) of various industrial equipment (e.g., CNC machine tools, compressors, assembly line motors) through a sensor IoT layer using vibration sensors, temperature sensors, and current sensors. Each factory client training node in the node computing layer preprocesses the raw data, performing noise reduction and normalization, and pre-trains a local model based on equipment fault characteristics (e.g., bearing wear, gearbox noise), generating gradient statistics containing the mean and variance of fault characteristics. This data is then encrypted, encapsulated as a transaction record, and submitted to the blockchain network. The blockchain executes a multi-dimensional incentive mechanism through smart contracts, comprehensively evaluating the task performance and data quality scores of each node. Nodes with high overall reputation scores are selected to form a temporary committee, creating a federated learning aggregation layer. Each committee member node is based on a Gaussian distribution assumption. The system weights and fuses global fault feature distribution parameters according to contribution weights. The aggregation result is written to the distributed ledger through the Dumbo-NG consensus algorithm. The smart contract automatically distributes the updated fault diagnosis model parameters to the edge nodes of each factory, triggering local model updates. Each factory uses the updated model to analyze real-time equipment data. When a potential fault is detected, gradient statistics are generated and uploaded to the blockchain again, forming a closed-loop iteration of "equipment data collection - local feature extraction - blockchain trusted aggregation - global model optimization". At the same time, the blockchain dynamically adjusts the model aggregation weights according to the comprehensive reputation score of each node, incentivizing nodes with high contributions to have more say in the fault prediction model iteration. Ultimately, this improves the accuracy of industrial equipment fault warnings and reduces maintenance costs. Moreover, the data is uploaded to the blockchain only in the form of statistics, which reduces communication overhead compared to traditional full model transmission, ensuring industrial data privacy and efficient system operation.

[0078] In summary, compared with the prior art, the advantages of the present invention are mainly reflected in the following aspects:

[0079] 1) This invention achieves decentralized joint training that is collaborative, secure, and reliable, enabling cross-domain collaborative training while ensuring data security and effectively reducing system communication and storage overhead. The designed decentralized multi-party joint training framework based on federated learning and blockchain allows each committee member node in the federated learning aggregation layer to perform weighted fusion of global statistical features in a decentralized manner, solving the single point of failure problem of traditional federated learning.

[0080] 2) This invention replaces traditional feature learning techniques with generative modeling and replaces the technique of uploading complete model parameters to the blockchain with the generated statistical features. While ensuring data credibility, it significantly improves communication efficiency, thereby achieving efficient shared data transmission.

[0081] 3) This invention guides the adjustment of contribution weights for each node and the election of the committee through a multi-dimensional reputation-oriented incentive mechanism. The characteristic of this multi-dimensional reputation-oriented incentive mechanism lies in establishing a comprehensive multi-dimensional reputation-oriented incentive mechanism that considers both model training performance and data quality, guiding the adjustment of contribution weights for each node and the election of the committee.

[0082] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0083] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0084] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0085] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0086] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0087] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0088] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0089] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0090] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A decentralized federated training system that ensures data trustworthiness and efficiency, comprising a sensor IoT layer, a node computing layer, and a federated learning aggregation layer constructed using a blockchain network, wherein: The sensor IoT layer is equipped with multiple sensors, each responsible for collecting raw data from the corresponding client training node; The node computing layer comprises multiple client training nodes. Each client training node acts as a computing node, responsible for the preprocessing of the corresponding raw data and local model pre-training, generating gradient statistics, which are then encrypted and encapsulated into transaction records and submitted to the blockchain network. The blockchain employs a consensus algorithm, managing a dynamic committee election process through smart contracts. Based on a designed multi-dimensional incentive mechanism, nodes are selected to form a temporary committee, creating the federated learning aggregation layer to replace the federated learning center server. The gradient statistics include mean and variance. In the federated learning aggregation layer, each temporary committee member node completes the weighted fusion of global statistical features in a decentralized manner. The global statistical features are written into the distributed ledger through asynchronous consensus, and then automatically distributed to each edge node using smart contracts, triggering local model updates.

2. The system according to claim 1, characterized in that, The system employs a shared data transmission mechanism, including: Assuming the feature space follows a class-conditional Gaussian distribution, for class c, the global feature distribution is defined as: Here, a shared feature extractor φ is trained such that the features z = φ(x) of all client nodes are aligned with the global distribution p(z|y), μ c ∈R d Let ∑ be the feature mean of the global class c. c ∈R d×d Let y be the feature covariance matrix of the global category c, and y represent the label. Let x represent a Gaussian distribution, d represent the feature dimension of the sample, and d represent the feature dimension of the sample. The contribution of each active client node and the proportion of each category in the Gaussian distribution are calculated and then weighted to aggregate the global feature distribution parameters, which are expressed as follows: Where M is the total number of active client nodes, Let ω be the number of samples of category c in client node i. i The contribution of client node i to the federated aggregation is calculated through a multi-dimensional incentive mechanism. This represents the mean of category c in client node i; Estimate the parameters of the class-conditional distribution to form a generative classifier. For feature z, the classification decision is based on Bayes' theorem: Where c′ represents a possible category, μ represents the set of categories. c′ Let p(y=c′) represent the mean vector of category c′, p(y=c′) represent the prior probability of category c′, and p(y=c) be the global category prior, estimated by weighting the statistical data transmitted from each client node. The classifier is generated based on the following formula: Client node i uses local data The estimated local conditional distribution is expressed as: in, Let n represent the feature mean vector of category c. c Let j represent the number of samples belonging to category c, j represent the sample index, and n represent the total number of samples. This represents the eigenvalue matrix after centering. This represents the estimation of the overall characteristic covariance matrix; when When a regularization term is added to the covariance, it is represented as: Where d represents the feature dimension. Let represent the covariance matrix of class c on client node i after regularization. Let represent the covariance matrix of the local estimate, γ represent the regularization coefficient, Tr represent the trace operation, and I represent the identity matrix; Through the interpolation coefficient β i Dynamically fusing global and local distributed parameters is represented as follows: in, This represents the final interpolated mean of category c at client node i. This represents the interpolation covariance matrix of class c on client node i. This represents the mean vector of class c estimated locally by the client. β represents the mean vector of class c in the global model. i Determined by minimizing the loss on the local validation set, expressed as: in, This is the interpolated classifier. Let (x, y) represent the local validation set used by client node i for validation, and let (x, y) represent a sample pair in the validation set.

3. The system according to claim 1, characterized in that, The multi-dimensional incentive mechanism includes: Calculate the task performance score PS using the following formula. i : PS i =ω1×Accuracy i +ω2×Stability i +ω3×Efficiency i Where ω1, ω2, and ω3 are the corresponding weights, and ω1 + ω2 + ω3 = 1, the accuracy is denoted as Accuracy. i It refers to the accuracy of the task, stability. i It refers to the stability and efficiency of the task. i This refers to the task completion efficiency, where i represents the client number; Calculate the Data Quality Score (DQS) using the following formula. i : DQS i =(1-PollutionFlag i )×Integrity i Among them, PollutionFlag i It is an exception indicator; in an abnormal state, the PollutionFlag indicates an error. i Integrity is 1 if it is not 0 otherwise i It is a data integrity indicator, with a value range of [0, 1], where 1 indicates that the data is complete; The Comprehensive Reputation Score (CRS) is calculated using the following formula. i : CRS i =α×PS i +β×QDS i Where α is the task performance score weight, β is the data quality score weight, and α+β=1; Nodes with higher overall reputation scores are selected as committee members to aggregate the global model, and then the global model is determined through a consensus method. The overall reputation score is converted into model aggregation weights, so that client training nodes with higher reputations have higher aggregation weights, as shown below: Where j is the index of the client and n is the number of clients.

4. The system according to claim 3, characterized in that, Stability i Measured by the inverse of the accuracy fluctuation over a set period.

5. The system according to claim 3, characterized in that, Efficiency i It is expressed as the reciprocal of the ratio of actual completion time to expected completion time.

6. The system according to claim 3, characterized in that, m represents the number of covered types, and M represents the total number of types.

7. The system according to claim 1, characterized in that, The global model is determined using the Dumbo-NG consensus method.

8. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images