Blockchain-based federated learning knowledge ticket verification and global model distillation method

Abstract

The application discloses a kind of federal learning knowledge ticket verification and global model distillation method based on blockchain, including the following steps: model publisher creates training task, constructs reference sample set and challenge sample set, and writes configuration information into blockchain;Participating node trains local teacher model based on local data, generates knowledge ticket and verification information;Knowledge ticket is stored in off-chain, and commitment information is submitted to blockchain;Aggregation party is based on verification information and preset parameter to determine the effectiveness of knowledge ticket;Knowledge fusion and distillation training are carried out based on effective knowledge ticket, and global model is updated.The application realizes the verifiability of federal learning process and the ability of resisting malicious attack while protecting data privacy by the separation architecture of storing summary on-chain and storing entity off-chain, in combination with multi-dimensional effectiveness verification mechanism based on challenge sample set, significantly improves the security and robustness of federal learning.

Description

Technical Field

[0001] This invention relates to the field of blockchain and federated learning integration technology, and in particular to a blockchain-based federated learning knowledge ticket verification and global model distillation method. Background Technology

[0002] With the development of artificial intelligence, distributed computing, and data element applications, an increasing number of scenarios require multiple data holders to conduct collaborative modeling without directly exchanging raw data. Federated learning avoids the direct privacy risks associated with the centralized aggregation of raw data by having each participant train the model locally using private data and then sharing the training updates externally. Meanwhile, blockchain, with its immutability and traceability, is also being used to enhance the trusted recording and process monitoring capabilities in multi-party collaboration.

[0003] However, while existing federated learning schemes typically do not directly upload raw data, they generally rely on uploading gradients, model parameters, or local update results to complete global model training. This type of parameter-level or gradient-level interaction still has several shortcomings: First, model update information itself may carry sensitive features related to local data distribution, and malicious parties or attackers may use this update information to infer training data, thus introducing new privacy risks. Second, in multi-party heterogeneous collaboration scenarios, the authenticity, stability, and validity of the content submitted by each participant lack a unified and reliable verification mechanism, making it easy for abnormal updates, forged updates, maliciously perturbed updates, or low-quality updates to be mixed into the global training process, thereby affecting model convergence and global model performance.

[0004] Furthermore, existing solutions for introducing blockchain into federated learning mostly focus on registering, verifying, or protecting model parameters or gradient updates, with their core technology still built upon parameter or gradient uploads. While these solutions improve process reliability or privacy protection to some extent, they have limited ability to express heterogeneous knowledge at a fine-grained level and struggle to balance the validity of updated content with the stable updating of the global model.

[0005] Therefore, there is an urgent need to provide a new federated collaborative modeling method that can achieve reliable verification of content submitted by multiple parties and effective updating of the global model without uploading local teacher model parameters or gradients. Summary of the Invention

[0006] This invention addresses the problems of existing federated learning schemes, which generally rely on uploading model parameters or gradients, resulting in privacy risks, difficulty in effectively verifying updated content, and insufficient stability of global model updates in heterogeneous data scenarios. It proposes a blockchain-based federated learning knowledge ticket verification and global model distillation method to achieve trusted knowledge interaction and global model updates without uploading local teacher model parameters or gradients.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a blockchain-based federated learning knowledge ticket verification and global model distillation method, comprising the following steps:

[0009] S1. Training task configuration and on-chain registration: The model publisher creates federated learning training tasks, constructs reference sample sets and challenge sample sets, and writes the training task configuration information and validity judgment parameters into the blockchain.

[0010] S2. Knowledge Ticket Generation and Submission: Participating nodes train a local teacher model on local private data based on the global student model. They then use the local teacher model to infer knowledge tickets from the reference sample set and to infer verification information from the challenge sample set. Participating nodes store the knowledge tickets and verification information off-chain and submit the commitment information of the knowledge tickets to the blockchain.

[0011] S3. Knowledge Ticket Verification and Validity Determination: The aggregator obtains off-chain knowledge tickets and verification information based on on-chain commitment information, and determines the validity of the knowledge tickets based on the verification information and preset validity determination parameters to obtain valid knowledge tickets and their fusion weights.

[0012] S4. Knowledge Fusion and Global Model Distillation Update: The aggregator performs knowledge fusion based on valid knowledge tickets and their fusion weights, and distills the global student model to obtain the updated global model. The updated global model information is then written into the blockchain.

[0013] Further, in step S1, the model publisher writes the reference sample set summary value, the challenge sample set summary value, the initial global model summary value, and the validity determination threshold into the blockchain, and stores the reference sample set, the challenge sample set, and the initial global model entity in the off-chain storage unit; the validity determination threshold includes at least the stability threshold, the uncertainty threshold, and the prototype distance threshold.

[0014] Furthermore, in step S2, the knowledge ticket includes at least the soft label matrix and uncertainty vector of the local teacher model on the reference sample set; the verification information is the response matrix of the local teacher model on the challenge sample set.

[0015] Furthermore, the knowledge ticket also includes a low-dimensional semantic prototype matrix, which is obtained by extracting, aggregating, and dimensionality-reducing the intermediate layer features of the local teacher model.

[0016] Further, in step S3, the aggregator calculates a stability score based on the verification information, a coverage score and an incremental score based on the knowledge document, calculates a comprehensive contribution score based on the stability score, coverage score and incremental score, and determines the validity of the knowledge document according to a preset threshold.

[0017] Furthermore, in step S4, the knowledge fusion includes weighted fusion of the soft label matrix, uncertainty vector, and / or low-dimensional semantic prototype matrix in the valid knowledge tickets to generate a global distillation target; the distillation training takes the reference sample set as input and optimizes the global distillation target as supervision.

[0018] Furthermore, in step S1, the blockchain only stores the sample set summary value, the model summary value, and their off-chain access index and version mapping relationship, and does not store their entity content; in step S2, the participating nodes encrypt the knowledge ticket and then store it in the off-chain storage unit.

[0019] Secondly, this invention proposes a blockchain-based federated learning knowledge ticket verification and global model distillation system, comprising:

[0020] Blockchain network: used to store training task configuration information, sample set summary values, model summary values, knowledge ticket commitment information, and version mapping relationships;

[0021] Off-chain storage units: used to store reference sample sets, challenge sample sets, global model entities, and knowledge ticket entities;

[0022] Model publisher node: Used to create training tasks, build reference sample sets and challenge sample sets, and write configuration information to the blockchain;

[0023] Participating nodes: are used to train local teacher models on local private data based on global student models, generate knowledge tickets and verification information, store knowledge tickets off-chain, and submit commitment information to the blockchain;

[0024] Aggregator Node: Used to obtain knowledge tickets and verification information from off-chain, perform validity determination and fusion weight calculation, perform knowledge fusion and distillation training based on valid knowledge tickets to obtain the updated global model, and write the updated model information into the blockchain.

[0025] Furthermore, the participating nodes include:

[0026] Local training module: used to train the local teacher model on local private data based on the global student model;

[0027] The document generation module is used to generate knowledge documents by reasoning on the reference sample set using the local teacher model.

[0028] Verification generation module: Used to perform reasoning on the challenge sample set using the local teacher model to generate verification information;

[0029] Storage and Submission Module: Used to store knowledge tickets and verification information off-chain, and submit the commitment information of knowledge tickets to the blockchain.

[0030] Furthermore, the aggregator node includes:

[0031] Acquisition module: Used to retrieve knowledge tickets and verification information from off-chain based on on-chain commitment information;

[0032] Verification module: Used to determine the validity of knowledge tickets based on verification information and preset validity judgment parameters, and to identify valid knowledge tickets and their fusion weights;

[0033] Fusion module: Used to perform weighted fusion of valid knowledge vouchers to generate a global distillation target;

[0034] Update module: Used to perform distillation training on the global student model based on the global distillation objective, to obtain the updated global model;

[0035] On-chain module: Used to write the updated global model information into the blockchain.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] First, this invention uses knowledge tickets as the interactive carrier for federated collaborative updates. Participating nodes do not need to upload local teacher model parameters or gradients. While ensuring that the original data does not leave the local machine, it reduces the risk of privacy leakage during the model update process and improves data security.

[0038] Secondly, this invention registers the commitment value, access index, and version identifier of knowledge tickets through blockchain, and combines challenge sample response and validity judgment mechanisms to achieve credible verification of knowledge tickets, which helps to improve the authenticity, stability, and validity of submitted content.

[0039] Third, by fusing effective knowledge tickets and distilling the global student model for updates, this invention enhances the system's ability to collaboratively utilize multi-source heterogeneous knowledge, improves the stability of global model updates, and helps improve the model's generalization ability.

[0040] Fourth, by leveraging the immutability and traceability of blockchain, this invention enables the trusted registration of training tasks, knowledge tickets, and model versions, which helps to improve the transparency, traceability, and overall credibility of the federated learning process. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0042] Figure 1 This is a schematic diagram of the overall architecture of the blockchain-based federated learning knowledge ticket verification and global model distillation method provided in this embodiment of the invention.

[0043] Figure 2 This is a flowchart illustrating the blockchain-based federated learning knowledge ticket verification and global model distillation method provided in this embodiment of the invention.

[0044] Figure 3 A schematic diagram illustrating the core mechanism of the blockchain-based federated learning knowledge ticket verification and global model distillation method provided in this embodiment of the invention. Detailed Implementation

[0045] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Example 1

[0047] This invention proposes a blockchain-based federated learning knowledge ticket verification and global model distillation system, the overall architecture of which is as follows: Figure 1 As shown, it includes:

[0048] Blockchain network: used to store training task configuration information, sample set summary values, model summary values, knowledge ticket commitment information, and version mapping relationships;

[0049] Off-chain storage units: used to store reference sample sets, challenge sample sets, global model entities, and knowledge ticket entities;

[0050] Model publisher node: Used to create training tasks, build reference sample sets and challenge sample sets, and write configuration information to the blockchain;

[0051] Participating nodes: are used to train local teacher models on local private data based on global student models, generate knowledge tickets and verification information, store knowledge tickets off-chain, and submit commitment information to the blockchain;

[0052] Aggregator Node: Used to obtain knowledge tickets and verification information from off-chain, perform validity determination and fusion weight calculation, perform knowledge fusion and distillation training based on valid knowledge tickets to obtain the updated global model, and write the updated model information into the blockchain.

[0053] The participating nodes include:

[0054] Local training module: used to train the local teacher model on local private data based on the global student model;

[0055] The document generation module is used to generate knowledge documents by reasoning on the reference sample set using the local teacher model.

[0056] Verification generation module: Used to perform reasoning on the challenge sample set using the local teacher model to generate verification information;

[0057] Storage and Submission Module: Used to store knowledge tickets and verification information off-chain, and submit the commitment information of knowledge tickets to the blockchain.

[0058] The aggregator node includes:

[0059] Acquisition module: Used to retrieve knowledge tickets and verification information from off-chain based on on-chain commitment information;

[0060] Verification module: Used to determine the validity of knowledge tickets based on verification information and preset validity judgment parameters, and to identify valid knowledge tickets and their fusion weights;

[0061] Fusion module: Used to perform weighted fusion of valid knowledge vouchers to generate a global distillation target;

[0062] Update module: Used to perform distillation training on the global student model based on the global distillation objective, to obtain the updated global model;

[0063] On-chain module: Used to write the updated global model information into the blockchain.

[0064] The model publisher registers training task information, reference sample set summaries, challenge sample set summaries, and global student model summaries on the blockchain. Participating nodes train teacher models based on their local private data and generate knowledge tickets containing soft-label matrices, uncertainty vectors, and low-dimensional semantic prototype matrices using a unified reference sample set. These tickets are encrypted and stored off-chain, and the commitment value, access index, and version identifier are submitted to the blockchain. The aggregator verifies the knowledge tickets based on the on-chain registration information and challenge response results, selects valid knowledge tickets, constructs distillation targets, and distills and updates the global student model. This method eliminates the need to upload local teacher model parameters or gradients, achieving reliable updates to the federated learning model while ensuring that the original data remains locally. This improves the security, verifiability, and traceability of the update process.

[0065] Example 2

[0066] This invention provides a blockchain-based federated learning knowledge ticket verification and global model distillation method, the process of which is as follows: Figure 2 As shown, it includes the following steps:

[0067] Step 1: Training Task Configuration and On-Chain Registration

[0068] The model publisher first creates a federated learning training task and generates a task identifier, round identifier, and initial model version number for this task. Simultaneously, the publisher constructs a unified reference sample set and a challenge sample set corresponding to this training task. The reference sample set carries the knowledge representations output by each participating node, while the challenge sample set verifies the consistency and stability of the knowledge documents of each participating node. The reference sample set consists of publicly available data, historical anonymized data, and / or synthetic samples constructed by the generative model. The challenge sample set is independent of the reference sample set and is used only for verifying the validity of knowledge documents; it does not participate in the local teacher model training of the nodes.

[0069] The model publisher stores the reference sample set, challenge sample set, and initial global student model in off-chain storage units, and writes the summary values ​​of the reference sample set, challenge sample set, and initial global student model, the current model version number, knowledge ticket submission deadline, distillation temperature parameters, knowledge compression dimension parameters, and validity judgment thresholds to the blockchain. The validity judgment thresholds are pre-determined based on the statistical results of the preheating rounds and / or the evaluation results of the independent validation sample sets. These thresholds include at least a stability threshold, an uncertainty threshold, and a prototype distance threshold, and are uniformly written to the blockchain before the start of the current training round. The blockchain does not store the entity content of the reference sample set, challenge sample set, and global student model; it only stores their summary values, off-chain access indexes, and version mapping relationships. After receiving the training task, each participating node downloads the current round's global student model, reference sample set, and challenge sample set based on the on-chain version number and index information, and uses the version registered on-chain as the unique and legitimate update baseline for this round.

[0070] In this step, the objects circulating on-chain are only the task description information, sample summaries, model summaries, and access indexes; the objects circulating off-chain are the reference samples, challenge samples, and the global student model ontology. This ensures version verifiability while avoiding the on-chain bloat problem caused by directly writing large-scale model files or reference data to the blockchain.

[0071] Step Two: Knowledge Ticket Generation and Submission

[0072] Each participating node performs local adaptive training on its local private dataset, using the current round's global student model as the initialization base, to obtain a teacher model that is retained only locally. This teacher model is used to infer from a unified reference sample set to generate corresponding knowledge tickets. Each knowledge ticket includes at least a soft-label matrix, an uncertainty vector, and a low-dimensional semantic prototype matrix of the teacher model on the reference sample set; these components constitute the main body of the knowledge ticket. The teacher model's response on the challenge sample set constitutes ticket-attached verification information, which corresponds one-to-one with the knowledge ticket main body and is submitted bound to it. This verification information does not participate in the global model distillation training. The knowledge ticket generation mechanism is as follows: Figure 3 As shown.

[0073] Specifically, let the first Participating nodes in round training The local teacher model is The unified reference sample set is ,in This represents the total number of samples in the reference sample set. Participating nodes construct a soft-label matrix using the teacher model's output on the reference sample set.

[0074] ,

[0075] in, For distillation temperature parameters, Represents a node In the The soft probability distribution matrix is ​​given by the reference sample set of the round. To ensure that knowledge tickets include not only category preference information but also uncertainty information, participating nodes further calculate the prediction entropy corresponding to each reference sample, forming an uncertainty vector:

[0076] ,

[0077] in, For the number of categories, Represents a node For reference sample The uncertainty characterization.

[0078] To further reflect the structural information of local data distribution, participating nodes extract feature representations of reference samples from the pre-defined intermediate feature layer of the teacher model, and aggregate feature vectors belonging to the same category to form prototype vectors for each category. Then, using the projection matrix seeds registered on the chain, deterministic orthogonal random projection matrices are generated to perform dimensionality reduction mapping on the prototype vectors of each category, resulting in low-dimensional semantic prototype matrices. In this way, the knowledge ticket not only contains output layer probability information but also compressed distribution structure information. However, since the prototype vector comes from the combined mapping of reference samples and intermediate features, and has undergone dimensionality reduction and quantization processing, the original private samples of the participating nodes cannot be directly deduced.

[0079] Subsequently, participating nodes combine the soft-label matrix, uncertainty vector, and low-dimensional semantic prototype matrix to form a knowledge ticket. To reduce storage and transmission pressure, participating nodes first quantize and encode the knowledge ticket according to a preset quantization bit width and then encapsulate it in fragments. Each fragment is then encrypted and stored in off-chain storage. Afterward, participating nodes use a one-way hash function to calculate the commitment value of the entire knowledge ticket and submit this commitment value, the off-chain address of the knowledge ticket, the current round identifier, and signature information to the blockchain.

[0080] In addition, participating nodes need to use the teacher model to reason about the challenge sample set to obtain a challenge response matrix. This matrix is ​​then compressed, encrypted, and stored in an off-chain storage unit. Its digest value and off-chain address are submitted along with the ticket as verification material. This challenge response matrix is ​​not used for global model distillation training; it is only used for subsequent stability checks and anomaly detection on the knowledge ticket.

[0081] Step 3: Verification and Validity Determination of Knowledge-Based Invoices

[0082] After the current round of knowledge ticket submission window closes, the aggregator pulls the encrypted knowledge ticket shards and challenge response encrypted data from each participating node based on the on-chain index. The aggregator first performs a consistency check on each knowledge ticket based on the commitment value, timestamp, signature, and version number. Only when the version number is consistent, the commitment value matches, the signature verification passes, and the submission time does not exceed the preset submission deadline, is the corresponding knowledge ticket included in the candidate set. For knowledge tickets included in the candidate set, the aggregator restores the knowledge ticket content and challenge response matrix, and determines their validity based on whether the knowledge ticket is stable, sufficiently comprehensive, and possesses effective incremental information.

[0083] For knowledge tickets entering the candidate set, the aggregator calculates a stability score based on the challenge response matrix. The stability score reflects whether the output of the participating node's teacher model on independent challenge samples in this round is basically consistent with the current global student model and maintains reasonable continuity with its previous submission. If a node's response on a challenge sample deviates significantly from the current global student model's response, or exhibits abnormal abrupt changes compared to its previous challenge response, then the node's knowledge ticket is considered to have a risk of forgery or training instability. Simultaneously, the aggregator calculates a coverage score based on the average probability distribution of the soft label matrix along the category dimension and its information entropy. This score characterizes the degree of distribution balance of the knowledge ticket in the category space; a more balanced distribution of the soft label matrix along the category dimension results in a higher coverage score. Furthermore, the aggregator calculates an incremental score based on the distance between the low-dimensional semantic prototype matrix and the current global prototype template. This score measures the extent to which the knowledge provided by the node supplements the current global knowledge distribution. A higher incremental score is achieved when the distance is within a preset effective range; a lower preset limit indicates limited new information; and a higher preset upper limit indicates a mismatch risk between the knowledge document and the current task space. The document verification and screening mechanism is as follows: Figure 3 As shown.

[0084] Based on the above three dimensions, the aggregator is a node. Calculate the overall contribution score:

[0085] ,

[0086] in, For stability score, To score coverage, For incremental scoring, The preset weight coefficients are used. When the stability score is lower than the preset stability threshold, or the mean and / or median of the uncertainty vector is higher than the preset uncertainty threshold, or the distance between the low-dimensional semantic prototype matrix and the current global prototype template exceeds the preset valid range, the aggregator will determine the knowledge ticket as an invalid ticket and remove it, and it will not participate in this round of global model update.

[0087] For all valid knowledge tickets, the aggregator normalizes their overall contribution scores to obtain the corresponding knowledge fusion weights:

[0088] ,

[0089] in, Indicates the first A set of valid knowledge tickets is generated. The aggregator normalizes the overall contribution score of all valid knowledge tickets to obtain the knowledge fusion weight corresponding to each valid knowledge ticket.

[0090] Step 4: Knowledge Fusion and Global Model Distillation Update

[0091] After obtaining the set of valid knowledge documents and their fusion weights, the aggregator performs weighted fusion on the soft label matrix, uncertainty vector, and low-dimensional semantic prototype matrix of the valid knowledge documents to generate a global distillation target. Specifically, for the reference sample... The aggregator constructs the fused soft target based on the soft tag matrix submitted by each valid node:

[0092] ,

[0093] And construct fusion uncertainty weights based on the uncertainty vectors submitted by each node:

[0094] ,

[0095] Simultaneously, the low-dimensional semantic prototype matrix is ​​weighted and fused to obtain the global semantic prototype template for the current round.

[0096] Subsequently, the aggregator uses the reference sample set as input and employs fusion soft objective, fusion uncertainty weights, and fusion semantic prototype templates as supervision to distill and train the current round's global student model, thereby obtaining the next round's global student model. Its optimization objective can be expressed as:

[0097] ,

[0098] in, For the next round of global student model, For Kullback-Leibler divergence, To predict the entropy function, For semantic prototype alignment loss, Here are the loss weight coefficients. The semantic prototype alignment loss is defined as:

[0099] ,

[0100] in For the fusion of the first Class global semantic prototype For model feature extraction layer, For the first Class reference sample, This represents the total number of categories.

[0101] In the optimization objective formula, the first term is used to make the global student model approximate the soft knowledge distribution after multi-node fusion; the second term uses the fusion uncertainty weight to weight and adjust the prediction entropy term to reduce the contribution of high-uncertainty samples to distillation training; and the third term is used to maintain the semantic structure consistency of the global model. The fusion distillation update mechanism is as follows: Figure 3 As shown.

[0102] After distillation training is complete, the aggregator stores the new global student model in off-chain storage and writes the new model's summary value, access index, current round's set of valid knowledge ticket commitment values, fusion weights of each valid node, and the new model version number to the blockchain. Subsequently, each participating node downloads the new global student model based on the new version index registered on-chain, entering the next round of local teacher model training and knowledge ticket generation until a preset stopping condition is met. This stopping condition can be set as follows: the average output change of the global student model on the reference sample set in two adjacent rounds is lower than a threshold; or the performance improvement of the global student model on the independent validation sample set is lower than a threshold for several consecutive rounds.

[0103] This invention, through a separate architecture of on-chain storage of digests and off-chain storage of entities, combined with a multi-dimensional validity verification mechanism based on challenge sample sets, achieves verifiability and resistance to malicious attacks in the federated learning process while protecting data privacy, significantly improving the security and robustness of federated learning.

[0104] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A blockchain-based federated learning knowledge ticket verification and global model distillation method, characterized in that, Includes the following steps: S1. Training task configuration and on-chain registration: The model publisher creates federated learning training tasks, constructs reference sample sets and challenge sample sets, and writes the training task configuration information and validity judgment parameters into the blockchain. S2. Knowledge Ticket Generation and Submission: Participating nodes train a local teacher model on local private data based on the global student model. The local teacher model is used to infer knowledge tickets from the reference sample set and to infer verification information from the challenge sample set. Participating nodes store knowledge tickets and verification information off-chain, and submit the commitment information of the knowledge tickets to the blockchain; S3. Knowledge Ticket Verification and Validity Determination: The aggregator obtains off-chain knowledge tickets and verification information based on on-chain commitment information, and determines the validity of the knowledge tickets based on the verification information and preset validity determination parameters to obtain valid knowledge tickets and their fusion weights. S4. Knowledge Fusion and Global Model Distillation Update: The aggregator performs knowledge fusion based on valid knowledge tickets and their fusion weights, and distills the global student model to obtain the updated global model. The updated global model information is then written into the blockchain.

2. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 1, characterized in that, In step S1, the model publisher writes the reference sample set summary value, the challenge sample set summary value, the initial global model summary value, and the validity judgment threshold into the blockchain, and stores the reference sample set, the challenge sample set, and the initial global model entity in the off-chain storage unit; The validity determination thresholds include at least the stability threshold, the uncertainty threshold, and the prototype distance threshold.

3. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 1, characterized in that, In step S2, the knowledge ticket includes at least the soft label matrix and uncertainty vector of the local teacher model on the reference sample set; the verification information is the response matrix of the local teacher model on the challenge sample set.

4. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 3, characterized in that, The knowledge ticket also includes a low-dimensional semantic prototype matrix, which is obtained by extracting, aggregating, and dimensionality-reducing the intermediate layer features of the local teacher model.

5. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 1, characterized in that, In step S3, the aggregator calculates a stability score based on the verification information, a coverage score and an incremental score based on the knowledge document, calculates a comprehensive contribution score based on the stability score, coverage score and incremental score, and determines the validity of the knowledge document according to a preset threshold.

6. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 1, characterized in that, In step S4, the knowledge fusion includes weighted fusion of the soft label matrix, uncertainty vector and / or low-dimensional semantic prototype matrix in the valid knowledge tickets to generate a global distillation target; the distillation training takes the reference sample set as input and the global distillation target as supervision for optimization.

7. The blockchain-based federated learning knowledge ticket verification and global model distillation method according to claim 1, characterized in that, In step S1, the blockchain only stores the sample set summary value, the model summary value, and their off-chain access index and version mapping relationship, and does not store their entity content; in step S2, the participating nodes encrypt the knowledge ticket and store it in the off-chain storage unit, and encrypt the verification information and store it in the off-chain storage unit.

8. A blockchain-based federated learning knowledge ticket verification and global model distillation system, characterized in that, include: Blockchain network: used to store training task configuration information, sample set summary values, model summary values, knowledge ticket commitment information, and version mapping relationships; Off-chain storage units: used to store reference sample sets, challenge sample sets, global model entities, and knowledge ticket entities; Model publisher node: Used to create training tasks, build reference sample sets and challenge sample sets, and write configuration information to the blockchain; Participating nodes: are used to train local teacher models on local private data based on the current global student model, generate knowledge tickets and verification information, store the knowledge tickets and verification information off-chain, and submit commitment information to the blockchain; Aggregator Node: Used to obtain knowledge tickets and verification information from off-chain, perform validity determination and fusion weight calculation, perform knowledge fusion and distillation training based on valid knowledge tickets to obtain the updated global model, and write the updated model information into the blockchain.

9. The blockchain-based federated learning knowledge ticket verification and global model distillation system according to claim 8, characterized in that, The participating nodes include: Local training module: used to train the local teacher model on local private data based on the global student model; The document generation module is used to generate knowledge documents by reasoning on the reference sample set using the local teacher model. Verification generation module: Used to perform reasoning on the challenge sample set using the local teacher model to generate verification information; Storage and Submission Module: Used to store knowledge tickets and verification information off-chain, and submit the commitment information of knowledge tickets to the blockchain.

10. The blockchain-based federated learning knowledge ticket verification and global model distillation system according to claim 8, characterized in that, The aggregator node includes: Acquisition module: Used to retrieve knowledge tickets and verification information from off-chain based on on-chain commitment information; Verification module: Used to determine the validity of knowledge tickets based on verification information and preset validity judgment parameters, and to identify valid knowledge tickets and their fusion weights; Fusion module: Used to perform weighted fusion of valid knowledge vouchers to generate a global distillation target; Update module: Used to perform distillation training on the global student model based on the global distillation objective, to obtain the updated global model; On-chain module: Used to write the updated global model information into the blockchain.

Images