A Blockchain-Based Group Collaborative Learning Method

Abstract

The present invention discloses a group collaborative learning method based on blockchain, including: S1, constructing a collaborative learning model based on blockchain, dividing a training set and a test set; S2, presetting a smart contract module, setting initial model parameters, global aggregation rounds and initial iterations, and dividing local training groups; S3, selecting the node with the highest credit in the local training group as the leader node, and the others as training nodes, and all the leader nodes form a global aggregation group; S4, the training node locally trains to obtain a local training model, and the leader node aggregates the local model using federated learning; S5, the global aggregation group aggregates the local model to obtain a global model; S6, the leader node verifies the local model, and updates the credit of the edge node according to the verification accuracy and historical credit; S7, updating the initial model parameters, judging i<I, and returning to S3 if i=i+1 is executed. The present invention not only effectively reduces the wireless communication delay, but also enhances the efficiency of collaborative learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial edge data processing, and specifically relates to a group collaborative learning method based on blockchain. Background Art

[0002] In the era of Industry 4.0, complex computing tasks and the accompanying massive amounts of data provide huge development potential for machine learning and decentralized industrial edge. Federated Learning (FL), as a distributed machine learning, solves data privacy issues while reducing the risk of privacy leakage, providing a potential solution for the industrial edge. However, due to high-latency wireless communication, strong heterogeneity, low security and other reasons, traditional federated learning is difficult to handle large-scale tasks in the Industry 4.0 scenario.

[0003] In addition, in general, in each iteration of the federated learning process, all clients need to upload their local training model parameters to the central server. After aggregation, the federated learning central server needs to broadcast the global model to each client, which causes huge wireless communication pressure and delay on the wireless network. In the Industry 4.0 scenario, a large number of distributed devices are distributed in different locations. In addition, these devices have heterogeneous computing power and data attributes, so aggregating the models trained by these heterogeneous devices into a single model may greatly reduce convergence. Traditional federated learning is prone to single point failures and some security issues. If a malicious client uploads the wrong model to the central server, the performance of the aggregated global model will be greatly degraded, which will in turn affect the local training of the next iteration, resulting in reduced collaborative learning effects. Summary of the invention

[0004] In order to solve the technical problem that the collaborative learning effect is reduced when using existing federated learning for aggregation, the present invention proposes a group collaborative learning method based on blockchain, which improves the efficiency of collaborative learning by using the decentralized, open and transparent data storage technology of blockchain. To solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0005] A group collaborative learning method based on blockchain, comprising the following steps:

[0006] S1, builds a collaborative learning model corresponding to the industrial edge based on blockchain technology, and divides all edge nodes in the industrial edge into training sets and test sets;

[0007] S2, preset the smart contract module in the blockchain, set the initial model parameters, the number of global aggregation rounds I and the initial iteration number i=1, and use the partition clustering method to divide the edge nodes in the training set into local training groups;

[0008] S3, select the edge node with the highest credit in each local training group as the leader node, and the other edge nodes are training nodes, and the leader nodes of all local training groups form a global aggregation group;

[0009] S4, each training node performs local training using the neural network and the initial model parameters to obtain a local training model, and the leader node aggregates the received local training models using the federated learning method to obtain a local model;

[0010] S5, the smart contract module triggers global aggregation, and the global aggregation group aggregates the received local models to obtain a global model;

[0011] S6, the leader node of each local training group verifies the local model using the test set and updates the credit of all edge nodes based on the verification accuracy and historical credit;

[0012] S7, update the initial model parameters according to the global model, judge i<I, if it is, execute i=i+1 and return to step S3, otherwise end.

[0013] In step S2, the smart contract module includes:

[0014] Node registration module: divides the roles of edge nodes according to credit, including training nodes and leader nodes;

[0015] Local model generation module: The training node generates a local training model based on the training set and the global model. The leader node normalizes the credit calculated in the previous round and uses it as the weight of this round to aggregate the local training model to generate a local model.

[0016] Local model upload module: After the local model is generated, the leader node uploads the model parameters of the local model to the blockchain;

[0017] Global model aggregation module: aggregates the received local models to obtain a global model;

[0018] Global model parameter acquisition module: responsible for obtaining the model parameters of the global model from the blockchain;

[0019] Credit upload module: upload credit to blockchain storage;

[0020] Status query module: obtains training status information, which includes the current number of local model training rounds, model update status and credit calculation quantity.

[0021] In step S2, the partitioning clustering method is used to divide the edge nodes in the training set into local training groups, including the following steps:

[0022] ①The task publisher publishes the collaborative learning task, and the edge nodes that wish to join the collaborative learning task send a request with identity and contribution to the task publisher;

[0023] ②The task publisher verifies the received request and selects edge nodes to join the collaborative learning task based on their contribution;

[0024] ③ Use the partition clustering method to divide the edge nodes that join the collaborative learning task into local training groups.

[0025] The calculation formula of the contribution is:

[0026]

[0027] In the formula, Contribution k represents the contribution of edge node k, D k represents the local dataset size of edge node k, N represents the total number of edge nodes in the training set, and f k Represents the local CPU frequency of edge node k.

[0028] In step S4, each leader node aggregates the received local training models using the federated learning method to obtain a local model, including the following steps:

[0029] ⑴Use the normalization method to calculate the aggregation weight of each training node;

[0030] (2) The leader node aggregates the received local training models using the federated learning method according to the aggregation weights to obtain the local model.

[0031] When i≥2, the calculation formula of the aggregation weight is:

[0032]

[0033] In the formula, G represents the number of local training groups, represents the credit of training node k in round i-1, represents the aggregate weight of training node k in the i-th round, and N represents the total number of edge nodes in the training set.

[0034] In step S6, the calculation formula of the edge node's credit is:

[0035]

[0036] In the formula, represents the credit of edge node k in the i-th round of iterative training, u represents the weight factor, represents the local model parameters of edge node k during the i-th round of iterative training, represents the global model parameters obtained by global aggregation, and represents the test set based on edge node k, δ represents the learning rate, represents the credit of edge node k in the mth round, and f(·) represents the model function.

[0037] Beneficial effects of the present invention:

[0038] A Blockchain-Enabled Group Federated Learning (BGFL) framework was built for wireless industrial edge scenarios. By dividing industrial equipment into multiple groups, the accuracy of collaborative learning is improved while meeting diverse needs. At the same time, the wireless communication delay required to achieve convergence is reduced, and data isolation and high confidentiality on the same chain based on the group dimension are achieved. BGFL uses hierarchical aggregation, that is, a combination of off-chain local aggregation and on-chain global aggregation to reduce aggregation complexity and improve the learning efficiency of heterogeneous industrial equipment. At the same time, smart contracts are used for on-chain model aggregation to ensure the trustworthiness and security of the framework and the learning process. It not only effectively reduces wireless communication delays and enhances the robustness of the collaborative learning framework, but also provides greater privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 It is a schematic diagram of the architecture model of the present invention.

[0041] Figure 2 Schematic diagram of the relationship between the local training group and the global aggregation group.

[0042] Figure 3 Schematic diagram of the influence of learning rate on the prediction accuracy of the present invention.

[0043] Figure 4 It is a performance schematic diagram of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] A group collaborative learning method based on blockchain, comprising the following steps:

[0046] S1, builds a collaborative learning model corresponding to the industrial edge based on blockchain technology, and divides all edge nodes in the industrial edge into training sets and test sets;

[0047] The established collaborative learning model is an initialized global model, which is put into the genesis block, and each edge node corresponds to an edge device in the industrial edge.

[0048] S2, preset the smart contract module in the blockchain, set the initial model parameters, the number of global aggregation rounds I and the initial iteration number i=1, and use the partition clustering method to divide the edge nodes in the training set into local training groups;

[0049] The smart contract module includes a node registration module, a local model generation module, a local model upload module, a global model aggregation module, a global model parameter acquisition module, a credit upload module and a status query module, wherein the calling functions corresponding to each module are as follows:

[0050] Register(): Node registration, register and divide the roles of each edge node in the blockchain, including training nodes and leader nodes. Each local training group contains a leader node. The selection of leader nodes and training nodes is determined according to the credit of the edge node. The node with the highest credit in each local training group is the leader node.

[0051] LocalAggregation(): After registration, the training node updates the local parameters of the neural network based on the training set and the global model for local training. The local parameters refer to the weight parameters and bias parameters of each layer in the neural network. The global model is obtained from the blockchain through the Get_GlobalModel() function. The leader node is responsible for off-chain aggregation of the local training model, and the credit calculated in the previous round is normalized and used as the weight of this round for local aggregation to obtain the local model.

[0052] Put_Local_Model(): Local model upload. After several rounds of iterative training and local model generation, the leader node uploads the model parameters of the local model to the blockchain for later global aggregation;

[0053] GlobalAggregation(): Global model aggregation, where the edge nodes in the global aggregation group of the initial round are selected according to the contribution of the edge nodes. The edge nodes in the global aggregation group, i.e., the leader nodes, do not participate in local training;

[0054] Get_GlobalModel(): Gets the global model parameters, which is responsible for obtaining the model parameters of the global model from the blockchain for a new round of local training;

[0055] Put_Credit(): upload credit to the blockchain to store credit records for subsequent query and verification;

[0056] Get_status(): Status query, obtain the current local model training rounds, model update status, credit calculation number and other training status information off-chain, so as to judge the next round of training iteration and iteration termination conditions. When the global model is updated, credit calculation is performed on each edge node. If the update information is not obtained, credit calculation cannot be performed. Among them, the model update status refers to whether the global model has been updated, and the credit calculation number refers to whether the credit of all edge nodes has been updated.

[0057] The method of using the partition clustering method to divide the edge nodes in the industrial edge into local training groups includes the following steps:

[0058] ①The task publisher publishes the collaborative learning task, and the edge nodes that wish to join the collaborative learning task send a request with identity and contribution to the task publisher;

[0059] The task issuer may be a factory management enterprise in the industrial edge.

[0060] The calculation formula for the contribution of each edge node is:

[0061]

[0062] In the formula, Contribution k represents the contribution of edge node k, D k represents the local dataset size of edge node k, N represents the total number of edge nodes in the training set, and f k Represents the local CPU frequency of edge node k.

[0063] ②The task publisher verifies the received request and selects edge nodes to join the collaborative learning task based on their contribution;

[0064] First, determine the contribution threshold, and then add the edge nodes whose contribution is greater than the contribution threshold to the collaborative learning task. The initial credit of all edge nodes is contribution, which ensures that each local training group has edge nodes participating in global aggregation, greatly increasing the security and stability of collaborative learning.

[0065] ③ Use the partition clustering method to divide the edge nodes that join the collaborative learning task into local training groups;

[0066] The partition clustering method includes the DBSCAN algorithm (Density-Based Spatial clustering of Applications with Noise) or the K-means algorithm (K-means clustering algorithm, K-means clustering algorithm), the local training group is divided by a group of edge nodes, and each local training group is independent of each other, each edge node is an industrial device, which can perform local training on artificial intelligence (AI) tasks on industrial edge devices. Each group has its own ledger to maintain transactions and data, and executes the consensus process independently of the consensus mechanism of other groups. Therefore, group division ensures that groups can be separated from each other to operate independently, further providing better privacy protection. Blockchain nodes can participate in different groups at the same time and perform different multi-party collaborative services.

[0067] S3, select the edge node with the highest credit in each local training group as the leader node, and the other edge nodes are training nodes, and the leader nodes of all local training groups form a global aggregation group:

[0068] In the Industry 4.0 scenario, due to differences in hardware and other physical factors, the computing power and wireless communication capabilities of edge devices in different groups vary greatly. In addition, the devices in the group are usually located in different geographical locations, resulting in heterogeneous data sets and features. Therefore, if these heterogeneous devices are aggregated into one model, it may significantly affect the convergence of collaborative learning. After selecting the nodes participating in collaborative learning, the present application uses a partitioned clustering method to divide different edge nodes into different local training groups. The global aggregation group is composed of the leader nodes of the local training groups to aggregate the local models received from the local training groups.

[0069] S4, each training node of each local training group performs local training using the neural network and the initial model parameters to obtain a local training model, and each leader node aggregates the received local training models using a federated learning method to obtain a local model;

[0070] The neural network is a two-layer fully connected layer. The training node of each local training group uses the status query module to obtain the training status information from the blockchain, uses the neural network and the initial model parameters obtained by the global model parameter acquisition module to perform local training to obtain the local training model, and the leader node starts the local model generation module to aggregate the received local training model using the federated learning method to obtain the local model, and uses the local model upload module to upload the parameters of the local model to the blockchain. The local training group composed of edge nodes with similar contribution is responsible for coordinating the edge devices in the same group to complete the training task. Since the performance of each edge node in the group is similar, the completion time of the local training model is not much different. Therefore, the local training in the same local training group is basically synchronized. Aggregation will only be triggered if there are enough training nodes uploading their local training models to the leader node. In order to prevent malicious nodes from extending the training time and reducing the training performance, a time threshold is set in the smart contract module, and the training parameters of the training nodes that exceed the time threshold of this round of training will be abandoned.

[0071] Each leader node uses the federated learning method to aggregate the received local training models to obtain a local model, including the following steps:

[0072] ⑴Use the normalization method to calculate the aggregation weight of each training node;

[0073] When the current training round number i ≥ 2, the aggregate weight of training node k in the i-th round is obtained by normalizing the credit of training node k in the i-1th round. The calculation formula of the aggregate weight is:

[0074]

[0075] In the formula, G represents the number of local training groups, represents the credit of training node k in round i-1, Represents the aggregate weight of training node k in the i-th round.

[0076] (2) The leader node aggregates the received local training models using the federated learning method according to the aggregation weights to obtain the local model.

[0077] S5, starting the global model aggregation module, the global aggregation group aggregates the received local models to obtain a global model;

[0078] The global aggregation group is composed of the leader nodes of all local training groups. This ensures that each local training group has edge nodes participating in the global aggregation, greatly increasing the security and stability of collaborative learning. For each edge node in each local training group, the global model parameter acquisition module is used to obtain the initial model parameters from the main chain for local training. After several rounds of local training, the leader node aggregates the local training model using the local model generation module. Once the local training iterations in the local training group are completed, the aggregated local model parameters will be uploaded to the InterPlanetary File System (IPFS) to reduce the storage load of the blockchain. At the same time, the global model aggregation module is used to aggregate the global model according to the average weight, and the model parameters of the global model are stored in IPFS, and the returned hash result is stored on the blockchain, which is convenient for lightweight storage of model parameters. The global aggregation group is composed of multiple leader nodes with strong computing power and high credit, which are responsible for coordinating the training group devices to aggregate the local model parameters. Due to the heterogeneity of different local training groups, synchronous global updates are used to improve the generalization performance of the model.

[0079] IPFS is a distributed storage system for web 3.0 that stores unique hash files accessible to all nodes while eliminating redundancy in wireless networks. In addition, IPFS ensures the security of storage transactions because it provides a high-throughput content-addressed block storage model. Therefore, uploading data to IPFS and storing the hash returned by IPFS in the blockchain can significantly reduce storage load without compromising security.

[0080] S6, the leader node of each local training group verifies the local model using the test set and updates the credit of all edge nodes in the training set based on the verification accuracy and historical credit;

[0081] The calculation formula of the edge node's credit is:

[0082]

[0083] In the formula, represents the credit of edge node k in the i-th round of iterative training, u represents the weight factor, represents the local model parameters of edge node k during the i-th round of iterative training, represents the global model parameters obtained by global aggregation, and represents the test set based on edge node k, δ represents the learning rate, represents the credit of edge node k in the mth round, and f(·) represents the model function. Represents the model validation accuracy.

[0084] To prevent malicious nodes from providing false quality assessment results, each edge node is rated for credit at the end of each iteration. In this way, even if an edge node performed well in the previous training iteration, it will not be discarded due to poor performance in one iteration due to edge device or network factors. When the training nodes in each local training group send their model parameters to the leader node, the leader node will verify the local model with the test set. Then, the new credit of the node is obtained using the verification accuracy and the node's historical credit, and uploaded to the blockchain through the credit upload module for credit record storage.

[0085] S7, update the initial model parameters according to the global model, judge i<I, if it is, execute i=i+1 and return to step S3, otherwise end.

[0086] Factory occupancy detection is a popular problem in Industry 4.0. The following uses four edge computing devices Jetson Nano, a laptop and a wireless router to establish a four-group alliance chain architecture across physical hardware based on the blockchain underlying platform FISCO BCOS to simulate privacy computing in multiple factories and provide authentic and reliable blockchain services. Specifically, based on the blockchain underlying platform FISCO BCOS, a multi-group blockchain is built to expand the alliance chain from the original one-chain-one-ledger storage / execution mechanism to a one-chain-multiple-ledger storage / execution mechanism, and to achieve data isolation and confidentiality on the same chain based on the group dimension. In the blockchain-enabled group federated learning (BGFL) of this application, different groups share the blockchain network, and network message isolation between each group organization is achieved through network access and account whitelist, and it is assumed that the same Byzantine Fault Tolerance (PBFT) algorithm is used as the consensus algorithm in each group.

[0087] In order to facilitate the occupancy prediction of each factory within a region and realize the function of distributed smart contract deployment, a visual node front (WeBASE-Front) is planned to be built for each Jetson Nano. Since each factory manages its own private key to avoid the risk of leakage or forgetting when performing digital signatures, which leads to the loss of assets in its account on the chain, it is planned to build a WeBASE-Sign component to encrypt and store the private keys of each user and realize unified management. In addition, the whitelist is used to control access rights to further improve the security of the WeBASE-Sign component. Finally, it is planned to build a WeBASE management platform (WeBASE-Web) on the virtual machine of another laptop. On the one hand, it will unify the management of blockchain nodes, multiple smart contracts, platform account private keys, and certificate permissions on four Jetson Nanos, and on the other hand, it will monitor the node status and transaction behavior on the entire blockchain and display them in a visual way.

[0088] The following convolutional neural network is used to train the task, and the number of training times is set to 9 and the batch size is set to 50. Next, the robustness of some parameters in the learning model is tested. Figure 3 The effect of learning rate on the prediction accuracy of the present invention is shown. At the beginning of training, the model with the highest learning rate converges the fastest. Figure 3 As shown in (1) in . As the number of training times increases, the prediction accuracy of the model with a higher learning rate is lower than that of the model with a lower learning rate. This is because the smaller the learning rate, the slower the loss gradient decreases, and the longer the convergence time. At the same time, too large a gradient descent step may exceed the optimal value, causing the gradient to oscillate back and forth around the minimum value, thus affecting the performance. Therefore, the learning rate is set to 0.001, and the loss function is as follows Figure 3 As shown in (2) in .

[0089] Figure 4 (a) in the figure evaluates the effect of the number of nodes on the wireless communication delay. Assuming that the nodes in each group are evenly distributed, the experimental results show that the present application is more effective than the traditional federated learning algorithm. This trend becomes more obvious as the number of nodes increases, because local groups are usually concentrated at the industrial edge, where the data transmission rate is higher than the rate from the client to the centralized server. Similarly, the higher the number of groups, the lower the wireless communication delay. In subsequent experiments, 4 groups were simulated, each with 20 industrial nodes, and each group had 4 training nodes and 1 leader node. Assuming that the geographical locations are evenly distributed among each client and that the parameter data attenuates by 1% every 100 meters during transmission, the wireless communication delay of the present application (2 groups and 4 groups) is compared with that of the traditional federated learning algorithm method, and the results are shown in Figure 2. Figure 4 As shown in (b) in the figure. Figure 4It can be intuitively seen in (b) that in all cases, the performance of the present application is better than the traditional federated learning algorithm, and the more blockchain groups there are, the smaller the wireless communication delay of the algorithm. This is because, in BGFL, the client will first perform local training and aggregation within each group, which greatly reduces the increase in data transmission size due to data attenuation. In addition, as the distance between industrial nodes increases, the wireless communication delay will be higher due to the increase in the size of the required transmission.

[0090] Figure 4 (c) in Figure 1 shows the model prediction accuracy under different collaborative learning frameworks. The random collaborative learning framework does not use average weighting, but uses a random weighting method for global aggregation. Figure 4 (c) It can be seen that the prediction accuracy of this application is as high as 0.9327, and the convergence speed is also within an acceptable range compared with the other two frameworks. This is because in each training iteration, this application gives higher aggregation weights to training nodes with higher credit values, thereby greatly improving the training performance. Compared with traditional federated learning and random frameworks, the training accuracy of this framework is improved by 3.73% and 33.03% respectively. In the traditional federated learning architecture, each client trains a local model separately based on local data, providing an attacker with a vulnerable attack surface. Figure 4 As shown in (d) in the figure, a malicious node attack is simulated to demonstrate the learning efficiency of this application, traditional federated learning, and random framework under different numbers of malicious nodes. Figure 4 (d) It can be seen that the present application always outperforms the other two frameworks in training performance. Even when there are malicious nodes in the environment, the average training accuracy of the present application is still 0.9694. This is because the existing federated learning framework cannot judge the correctness of the local learning model, and the client can freely submit malicious models. Once the node uploads the malicious model parameters, the performance of collaborative learning will drop rapidly. When there are 8 malicious nodes in the edge scenario, the accuracy of the present application is 42.53% and 46.82% higher than the other two collaborative learning frameworks, respectively.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A group collaborative learning method based on blockchain, characterized in that: It includes the following steps: S1. Build a collaborative learning model corresponding to the industrial edge based on blockchain technology, and divide all edge nodes in the industrial edge into a training set and a test set; S2. Preset an intelligent contract module in the blockchain, set the initial model parameters, the global aggregation round number I, and the initial iteration number i = 1, and use the partition clustering method to divide the edge nodes in the training set into local training groups; S3. Select the edge node with the highest credit in each local training group as the leader node, and the other edge nodes are all training nodes, and the leader nodes of all local training groups form a global aggregation group; S4. Each training node uses a neural network and the initial model parameters to perform local training to obtain a local training model, and the leader node uses the federated learning method to aggregate the received local training models to obtain a local model; S5. The intelligent contract module triggers global aggregation, and the global aggregation group aggregates the received local models to obtain a global model; S6. The leader node of each local training group uses the test set to verify the local model, and updates the credit of all edge nodes according to the verification accuracy and historical credit; S7. Update the initial model parameters according to the global model, judge whether i < I, if so, execute i = i + 1 and return to step S3, otherwise end; In step S2, the method of using the partition clustering method to divide the edge nodes in the training set into local training groups includes the following steps: ① The task publisher publishes a collaborative learning task, and the edge nodes that hope to join the collaborative learning task send requests with identity identifiers and contribution degrees to the task publisher; ② The task publisher verifies the received requests, and selects edge nodes to join the collaborative learning task according to the contribution degrees of the edge nodes; ③ Use the partition clustering method to divide the edge nodes that join the collaborative learning task into local training groups.

2. The blockchain-based group collaborative learning method according to claim 1, characterized in that: In step S2, the intelligent contract module includes: Node registration module: Divide the roles of edge nodes according to credit, and the roles include training nodes and leader nodes; Local model generation module: The training node generates a local training model based on the training set and the global model, and the leader node aggregates the local training models by using the credit normalized in the previous round as the weight of this round to obtain a local model; Local model upload module: After the local model is generated, the leader node uploads the model parameters of the local model to the blockchain; Global model aggregation module: Aggregate the received local models to obtain a global model; Global model parameter acquisition module: Responsible for obtaining the model parameters of the global model from the blockchain; Credit upload module: Upload the credit to the blockchain for storage; Status query module: Obtain training status information, and the training status information includes the current local model training round number, model update situation, and credit calculation quantity.

3. The blockchain-based group collaborative learning method according to claim 2 is characterized in that: The calculation formula of the contribution degree is: In the formula, Contribution k represents the contribution of edge node k, D k represents the local dataset size of edge node k, N represents the total number of edge nodes in the training set, and f k Represents the local CPU frequency of edge node k.

4. The blockchain-based group collaborative learning method according to claim 1, characterized in that: In step S4, the method of each leader node respectively using the federated learning method to aggregate the received local training models to obtain a local model includes the following steps: ⑴ Use the normalization method to calculate the aggregation weight of each training node; (2) The leader node aggregates the received local training models using the federated learning method according to the aggregation weights to obtain the local model.

5. The blockchain-based group collaborative learning method according to claim 4 is characterized in that: When i≥2, the calculation formula of the aggregation weight is: In the formula, G represents the number of local training groups, represents the credit of training node k in round i-1, represents the aggregate weight of training node k in the i-th round, and N represents the total number of edge nodes in the training set.

6. The blockchain-based group collaborative learning method according to claim 1, characterized in that: In step S6, the calculation formula of the edge node's credit is: In the formula, represents the credit of edge node k in the i-th round of iterative training, u represents the weight factor, represents the local model parameters of edge node k during the i-th round of iterative training, represents the global model parameters obtained by global aggregation, and represents the test set based on edge node k, δ represents the learning rate, represents the credit of edge node k in the mth round, and f(·) represents the model function.

Images