Federal learning service method and system oriented to data circulation

Abstract

The invention belongs to the technical field of communication, and discloses a federated learning service method and system oriented to data circulation, which effectively protects the data privacy and security of participants by introducing mechanisms such as task auditing, qualification verification, encrypted data interaction, model update parameter screening and contribution value-based reward distribution. Potential hostile attacks are prevented, and fair and reasonable benefit distribution is realized.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a federated learning service method and system for data circulation. Background Art

[0002] The centralized federated learning method is a method in which clients of multiple participants collaborate to solve machine learning problems under the coordination of an aggregation server. The basic principle is that the aggregation server initializes the global model and distributes it to each client. Each client uses a local data set to train the global model, and then uploads the updated parameters or gradients to the aggregation server. The aggregation server aggregates these parameters or gradients to update the global model, and repeats the above process based on the updated global model until the performance of the global model meets the requirements.

[0003] Data circulation is the process of promoting the effective sharing and utilization of data resources under the premise of ensuring data security and privacy protection. In this process, data security and privacy protection are core issues. Although data is not directly shared during federated learning, it still faces multiple security challenges.

[0004] First of all, how to ensure the trustworthiness of the identities of the participants and the security of the data when the participants interact with the aggregation server to update the parameters of the model is an important issue. This involves technologies such as identity authentication and data transmission encryption.

[0005] Secondly, the credibility of the aggregation server also needs to be guaranteed. As a central node, the aggregation server has the model update information uploaded by all participants. How to prevent it from abusing or leaking this information is a challenge.

[0006] Third, it is a technical difficulty to evaluate the quality of the corpus dataset provided by the participants without leaking the plaintext data. This requires the design of a special encryption evaluation mechanism.

[0007] Fourth, how to avoid model poisoning attacks and collaborative fraud that may occur when participants submit model update parameters to ensure that the performance of the final model is not affected is a key security issue in federated learning.

[0008] Finally, how to fairly and reasonably evaluate the contributions of each participant in the entire process to ensure the rationality of reward distribution is crucial to maintaining the healthy development of the federated learning ecosystem.

[0009] In view of the above problems, the existing technology needs to be improved urgently. Summary of the invention

[0010] The purpose of this application is to provide a federated learning service method and system for data circulation, which can improve data security, protect the privacy of participants, prevent model poisoning attacks, and fairly evaluate the contributions of participants.

[0011] In a first aspect, the present application provides a federated learning service method for data circulation, comprising the steps of: A1. In response to a task request initiated by a consumer client, review the task request; A2. After the task request is reviewed, the task is issued to multiple participating clients based on the task request; A3. Perform qualification verification on the client of the participant who submitted the request to participate in the task, so as to verify whether the corpus data set held by the corresponding client of the participant meets the quality requirements, and determine the client of the participant who has passed the qualification verification as the participant of this task; the qualification verification process is based on data exchange based on encryption algorithm; A4. Use the aggregation server and the participating terminals to perform federated learning to generate a model that meets the task requirements and count the contribution value of each participating terminal; during the federated learning process, the aggregation server and the participating terminals exchange data based on an encryption algorithm, and the aggregation server screens the model update parameters uploaded by each participating terminal before performing aggregation calculation each time; A5. Send the model to the consumer client for acceptance and settlement; A6. After the acceptance and settlement are completed, rewards will be issued to each participating terminal according to the contribution value of each participating terminal.

[0012] By introducing mechanisms such as task review, qualification verification, encrypted data interaction, model update parameter screening, and contribution-based reward distribution, the data privacy and security of the participants are effectively protected, potential malicious attacks are prevented, and fair and reasonable benefit distribution is achieved.

[0013] Preferably, step A1 comprises: Receiving a task request including task requirements sent by the consumer client; Assess whether the stated mission requirements can be met; Otherwise, it is determined that the task request has not passed the review; If so, it is determined that the task request has passed the review, and the verification data set provided by the consumer client is received.

[0014] By clarifying the task requirements, assessing the feasibility of the requirements, making review decisions, and obtaining verification data sets, we can achieve effective review of task requests. This review mechanism can ensure that the accepted tasks are feasible, which helps to improve the efficiency and success rate of the subsequent federated learning process. At the same time, by obtaining the verification data set, it provides the necessary data support for subsequent model verification, which helps to ensure the quality of the final model.

[0015] Preferably, step A3 comprises: Based on the encryption algorithm, the zero-knowledge proof technology is used to verify the quality of the corpus data set held by each of the participating client terminals that submits the task participation request, so as to verify whether each of the corpus data sets meets the quality requirements; The participant client corresponding to the corpus data set that meets the quality requirements is determined to have passed the qualification verification, and the participant client that has passed the qualification verification is determined as a participant terminal participating in this task.

[0016] In this way, the system can screen out corpus data sets that meet the quality requirements and identify the corresponding participating client as the participating end that is eligible to participate in the task. This not only ensures the data quality of participating in federated learning, but also protects the data privacy of the participants, laying the foundation for the subsequent federated learning process.

[0017] Preferably, step A4 comprises: A401. The aggregation server initializes global parameters according to task requirements and sends the global parameters to each participating terminal based on an encryption algorithm; the global parameters include model type and model parameters; A402. Each of the participating terminals performs model optimization training based on the global parameters using the corpus data set it holds, obtains model update parameters, and uploads the model update parameters to the aggregation server based on an encryption algorithm; A403. After all the participating terminals have completed uploading the model update parameters, the aggregation server screens the model update parameters of each participating terminal, performs aggregation calculation with the screened valid model update parameters to obtain the aggregation parameters, and calculates the single-round contribution of each participating terminal; A404. The aggregation server updates the model parameters according to the aggregation parameters, and uses the verification data set to verify whether the model after the updated model parameters meets the task requirements. If not, the updated model parameters are sent to each participating terminal based on the encryption algorithm, and the process returns to execute step A402. If it is satisfied, the contribution value of each participating terminal is calculated based on the single-round contribution degree, and the iteration ends.

[0018] Preferably, in step A403, the aggregation server screening the model update parameters of each of the participating terminals includes: B1. Using the verification data set to verify the impact of the model update parameters submitted by each participating terminal in this round on the model performance, and screening out the model update parameters that cause the model performance to deteriorate; B2. Compare the remaining model update parameters with the aggregation parameters calculated in the previous round, and filter out the model update parameters that differ too much from the aggregation parameters calculated in the previous round; B3. Take the remaining model updating parameters as effective model updating parameters.

[0019] Preferably, after step B2 and before step B3, the method further comprises the following steps: The remaining model update parameters are further screened based on the Multi-Krum algorithm.

[0020] Preferably, in step A403, calculating the single-round contribution of each participating terminal includes: Using the Shapley value algorithm, according to the model update parameters of each participating terminal, the single-round contribution marginal value ratio of each participating terminal is calculated; Based on the screening results of the model update parameters, the proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated; Calculate the proportion of the single-round efficiency contribution evaluation value of each participating terminal according to the time consumed by each participating terminal in the current round of learning; The single-round contribution degree of each participating terminal is calculated according to the proportion of the single-round contribution marginal value, the proportion of the single-round reputation contribution evaluation value and the proportion of the single-round efficiency contribution evaluation value.

[0021] Preferably, based on the screening results of the model update parameters, the proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated, specifically including: The attenuation index of each participating terminal is calculated according to the following penalty rule formula: ; ; in, is the learning round number of this round of learning, is the attenuation index of the i-th participating terminal in the t-th round of learning, e is the base of the natural logarithm, is the decay rate index, is the learning round number bias of the i-th participating terminal, is the learning round number when the model update parameters of the i-th participating terminal were last screened out, is the set of participating terminals corresponding to the effective model update parameters in the tth round of learning, Indicates that the i-th participating terminal belongs to The participating end, Indicates that the i-th participating terminal does not belong to The participating end, is the attenuation window period; The proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated according to the following formula: ; in, is the proportion of the single-round reputation contribution evaluation value of the i-th participant in the t-th round of learning, is the attenuation index of the jth participating terminal in the tth round of learning, and K is the total number of participating terminals.

[0022] Preferably, according to the time consumed by each of the participating terminals in the current round of learning, the proportion of the single-round efficiency contribution evaluation value of each of the participating terminals is calculated, which specifically includes: The efficiency index of each participating terminal is calculated according to the following formula: ; in, is the learning round number of this round of learning, is the efficiency index of the i-th participant in the t-th round of learning, is the time consumed by the i-th participant in the t-th round of learning, is the time consumed by the jth participating terminal in the tth round of learning, K is the total number of participating terminals, is the set of participating terminals corresponding to the effective model update parameters in the tth round of learning, Indicates that the i-th participating terminal belongs to The participating end, Indicates that the i-th participating terminal does not belong to Participating end in The single-round efficiency contribution evaluation value ratio of each participating terminal is calculated according to the following formula: ; in, is the proportion of the single-round efficiency contribution evaluation value of the i-th participant in the t-th round of learning, is the efficiency index of the j-th participating terminal in the t-th round of learning.

[0023] In a second aspect, the present application provides a federated learning service system for data circulation, including a business server and a consumer client, an aggregation server, and multiple participant clients that are communicatively connected to the business server; The consumer client is used to initiate a task request to the business server; The business server is used for: In response to a task request initiated by a consumer client, review the task request; After the task request is reviewed, the task is issued to multiple participating clients based on the task request; Perform qualification verification on the client of the participating party that submitted the request to participate in the task, so as to verify whether the corpus data set held by the corresponding client of the participating party meets the quality requirements, and determine the client of the participating party that has passed the qualification verification as the participating end of this task; the qualification verification process is based on data exchange based on encryption algorithm; The aggregation server and the participating terminals are used to perform federated learning to generate a model that meets the task requirements and to count the contribution value of each participating terminal. During the federated learning process, the aggregation server and the participating terminals exchange data based on an encryption algorithm, and the aggregation server screens the model update parameters uploaded by each participating terminal before performing aggregation calculation each time. Sending the model to the consumer client for acceptance and settlement; After the acceptance and settlement are completed, rewards will be issued to each participating terminal according to the contribution value of each participating terminal.

[0024] Beneficial effects: The federated learning service method and system for data circulation provided by this application effectively protects the data privacy and security of participants, prevents potential malicious attacks, and achieves fair and reasonable distribution of benefits by introducing mechanisms such as task review, qualification verification, encrypted data interaction, model update parameter screening, and reward distribution based on contribution value. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of a federated learning service method for data circulation provided in an embodiment of the present application.

[0026] Figure 2 A schematic diagram of the structure of a federated learning service system for data circulation provided in an embodiment of the present application.

[0027] Explanation of reference numerals: 100, business server; 200, consumer client; 300, aggregation server; 400, participant client. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0029] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0030] Centralized federated learning solves machine learning problems with the collaboration of multiple participating clients and aggregation servers. However, in the process of data circulation, there are multiple challenges in ensuring data security and privacy protection while promoting the effective sharing and utilization of data resources. Specifically, these challenges include the verification of the credibility of the identity of the participants, the security of data interaction, the credibility of the aggregation server, data quality assessment, prevention of model poisoning attacks and collaborative fraud, and fair evaluation of the contributions of each participant to reasonably distribute rewards. These issues directly affect the security, effectiveness and fairness of the federated learning system, and a comprehensive solution is urgently needed.

[0031] To do this, please refer to Figure 1 , Figure 1 It is a federated learning service method for data circulation in some embodiments of the present application, comprising the steps of: A1. Respond to the task request initiated by the consumer client and review the task request; A2. After the task request is reviewed, the task is published to multiple participating clients based on the task request; A3. Perform qualification verification on the client that submits the task participation request to verify whether the corpus dataset held by the corresponding client meets the quality requirements, and determine the client that passes the qualification verification as the participating end of this task; the qualification verification process is based on data exchange based on encryption algorithm; A4. Use the aggregation server and the participating terminals to perform federated learning to generate a model that meets the task requirements and count the contribution value of each participating terminal. During the federated learning process, the aggregation server and the participating terminals exchange data based on an encryption algorithm, and the aggregation server screens the model update parameters uploaded by each participating terminal before performing each aggregation calculation. A5. Send the model to the consumer client for acceptance and settlement; A6. After acceptance and settlement, rewards will be distributed to each participant based on their contribution value.

[0032] By introducing mechanisms such as task review, qualification verification, encrypted data interaction, model update parameter screening, and contribution-based reward distribution, the data privacy and security of the participants are effectively protected, potential malicious attacks are prevented, and fair and reasonable benefit distribution is achieved.

[0033] Among them, the federated learning service method for data circulation can be applied to Figure 2 The business server 100 in the federated learning service system for data circulation is shown.

[0034] The working principle of this application can be described in detail as follows: First, when a consumer client initiates a task request, the system will review the request. The review process includes evaluating the legality and feasibility of the task requirements and ensuring that the task meets the system's specifications and capabilities. After the task request passes the review, the system will publish the task information to multiple potential participating clients.

[0035] Next, the system verifies the qualifications of the client that submitted the request to participate in the task. This step uses encryption algorithms for data interaction to protect the data privacy of the participants. The verification process focuses on whether the corpus dataset held by the participant (i.e., the training dataset used for model training) meets the quality requirements (for example, zero-knowledge proof technology can be used to achieve this). The client of the participant that passes the qualification verification is determined as the participant of this task.

[0036] In the process of federated learning, all data interactions between the aggregator and the participating terminals are based on encryption algorithms to ensure the security of data transmission. The aggregator first initializes the global model parameters and distributes them to each participating terminal. The participating terminal uses local data to train the model, and then uploads the updated model parameters or gradients to the aggregator server in encrypted form.

[0037] Before each aggregation calculation, the aggregation server will first filter the model update parameters uploaded by each participating end. This step can effectively prevent model poisoning attacks and collaborative fraud. The screening can be based on a variety of strategies, such as outlier detection, historical contribution evaluation, etc. The filtered valid model update parameters are used for aggregation calculation to update the global model.

[0038] During the entire learning process, the system will count the contribution value of each participant. The calculation of contribution value can take into account multiple factors, such as model performance improvement, data quality, computing resource investment, etc. This provides an objective basis for subsequent reward distribution.

[0039] After the learning process is completed, the system sends the generated model to the consumer client for acceptance and settlement. This step ensures that the final model meets the needs of consumers. After the acceptance and settlement are completed, the system will issue rewards based on the contribution values ​​of each participating terminal counted previously to achieve fair and reasonable distribution of benefits.

[0040] Specifically, step A1 includes: A101. Receive a task request including task requirements sent by a consumer client; A102. Evaluate whether mission requirements can be met; A103. Otherwise, it is determined that the task request has not passed the review; A104. If so, the task request is determined to have passed the review, and the verification data set provided by the consumer client is received.

[0041] By clarifying the task requirements, assessing the feasibility of the requirements, making review decisions, and obtaining verification data sets, we can achieve effective review of task requests. This review mechanism can ensure that the accepted tasks are feasible, which helps to improve the efficiency and success rate of the subsequent federated learning process. At the same time, by obtaining the verification data set, it provides the necessary data support for subsequent model verification, which helps to ensure the quality of the final model.

[0042] Among them, task request refers to a request initiated by a consumer client that contains specific machine learning task requirements. Task requirements may include some or all of model attribute requirements, model quality requirements, and delivery requirements; but are not limited to these. Among them, model attribute requirements may include one or more of model type requirements, model structure requirements, model function requirements, etc.; model quality requirements may include one or more of performance indicators such as accuracy requirements, recall requirements, and F1 score requirements; delivery requirements may include one or more of delivery time requirements, delivery form requirements, etc.

[0043] After receiving the task request, this application will evaluate whether the task requirements can be met. This evaluation process may involve multiple aspects, such as some or all of the following items: 1. Check whether there are sufficient computing resources to support the required model training; 2. Assess whether there are enough participants to provide data that meets the requirements; 3. Analyze whether the task can be completed within the specified time; 4. Determine whether the required model performance indicators can be achieved.

[0044] If the evaluation results show that the task requirements cannot be met (i.e., at least one requirement is not met), the task request is considered to have failed the review. This may be due to insufficient resources, too strict time constraints, or too high performance requirements. In this case, the specific reasons can be fed back to the consumer client so that they can adjust their requirements or find other solutions.

[0045] If the evaluation results show that the task requirements can be met, the task request is considered to have passed the review. At this point, the application will receive the verification data set provided by the consumer client. The role of the verification data set is to evaluate whether the performance of the model meets the expected requirements in the subsequent federated learning process.

[0046] The receipt of the validation data set can be based on an encryption algorithm to ensure the confidentiality and integrity of the data. After receiving the validation data set, it can be preliminarily checked to ensure that the data format is correct, the quantity is sufficient, and it matches the task requirements.

[0047] Through this review mechanism, this application can effectively filter out infeasible or difficult-to-achieve task requests, thereby improving the efficiency of the entire federated learning system. At the same time, by obtaining a verification data set, it provides the necessary foundation for the subsequent model training and verification process, helping to ensure that the quality of the final model meets the requirements of consumers. Specifically, it has the following advantages: More comprehensive assessment of task requirements: Existing technologies may only focus on model performance requirements, while ignoring factors such as computing resources and time constraints. The solution of this application considers requirements in multiple dimensions, including model attributes, quality indicators, and delivery requirements, so as to more comprehensively assess the feasibility of the task. Obtain validation data sets in advance: Many existing systems require validation data only after model training is completed; this application obtains validation data sets during the task review phase, which enables the system to evaluate performance during training and adjust strategies in a timely manner, improving the efficiency of the entire process; Higher system efficiency: By conducting strict review before task execution, this application can effectively avoid the situation where infeasible tasks occupy system resources, thereby improving the operating efficiency of the entire federated learning system; Better user experience: For tasks that fail the review, the system can provide specific feedback on the reasons to help users adjust their requirements or find alternatives, which provides a better user experience than simply rejecting the task.

[0048] More flexible task management: The review mechanism of this application allows the system to dynamically adjust the task acceptance strategy according to the current resource status. This flexibility is especially valuable when processing high-concurrency task requests.

[0049] In some possible implementations, in step A104, while or after receiving the verification data set provided by the consumer client, a deposit collection operation is also performed to collect the deposit paid by the consumer client. The deposit collection operation reduces the probability of consumers refusing to pay the incentive after the training is completed, thereby protecting the interests of the participants.

[0050] Among them, qualification verification refers to the process of quality assessment of the corpus data set held by the participating client. For example, it can be implemented by using zero-knowledge proof technology to verify the data quality without leaking the original data.

[0051] In some embodiments, step A3 comprises: A301. Based on encryption algorithms, zero-knowledge proof technology is used to verify the quality of the corpus datasets held by the participating clients that submit the task request to verify whether each corpus dataset meets the quality requirements; A302. Determine whether the participant client corresponding to the corpus data set that meets the quality requirements has passed the qualification verification, and determine the participant client that has passed the qualification verification as the participant terminal participating in this task.

[0052] The problem of evaluating the quality of the corpus datasets provided by the participants without leaking plaintext data is solved through encryption algorithms and zero-knowledge proof technology. The encryption algorithm is used to ensure the security of the data transmission process during the verification process. Zero-knowledge proof technology is used to verify the quality of the corpus dataset. It can verify whether the dataset meets the quality requirements without directly accessing or exposing the original data (the corpus dataset does not need to be uploaded to the aggregation server). The combination of these two technologies not only protects the data privacy of the participants, but also realizes the effective verification of data quality.

[0053] In this way, the system can screen out corpus data sets that meet the quality requirements and identify the corresponding participating client as the participating end that is eligible to participate in the task. This not only ensures the data quality of participating in federated learning, but also protects the data privacy of the participants, laying the foundation for the subsequent federated learning process. Compared with the traditional method of directly checking data, it has higher security and privacy protection capabilities, while ensuring data quality, providing an innovative solution to the data security and privacy protection issues in federated learning.

[0054] In the present application, the use of encryption algorithms can be implemented in a variety of ways. For example, a symmetric encryption algorithm such as SM3 (Advanced Encryption Standard) or an asymmetric encryption algorithm such as SM2 can be used.

[0055] Taking the asymmetric encryption algorithm based on the SM2 algorithm as an example, the system and each participating client (as well as the consumer client) each generate a pair of asymmetric encryption key pairs (including public key and private key) based on the SM2 algorithm, and send the public key to the other party through a secure channel. When transmitting data to the other party, it is first encrypted using the other party's public key before sending it. After receiving the encrypted data sent by the other party, it is decrypted using its own private key.

[0056] For zero-knowledge proof technology, methods such as zk-SNARK (zero-knowledge succinct non-interactive argument of knowledge) or zk-STARK (scalable transparent argument of knowledge) can be used.

[0057] For example, in some embodiments, a zk-SNAKR based on the ZEN architecture (ZEN architecture is a microprocessor architecture) and using the Groth16 protocol (Groth16 is a concise non-interactive knowledge argumentation protocol used in zero-knowledge proof systems, a way of representing calculations, which operates on arithmetic circuits, using addition and multiplication gates) is used. The specific process includes: The system constructs a data brusher according to the task requirements and the verification data set (the specific process is an existing technology and will not be described in detail here), and sends it to each participant client (referring to the participant client that submits the task request, and the sending process can be based on an encryption algorithm); the data brusher is used to verify whether the corpus data set meets the quality requirements (such as one or more of the requirements for data duplication, data missing, data tilt, data anomaly, and conditional filtering requirements related to the task requirements). When the corpus data set meets the quality requirements, the corresponding output of the data brusher is 1, otherwise the corresponding output is 0; The system converts the data selector into an arithmetic circuit F according to security parameters (set according to actual needs), generates a proof key pk and a verification key vk corresponding to the arithmetic circuit F (the specific process is prior art and will not be described in detail here), and sends the proof key pk (which can be sent based on an encryption algorithm) to the client of each participant; The Pedersen commitment C of the participating client using the random blinding factor r to generate its own corpus dataset i , and uses the Prove algorithm based on the ZEN architecture to generate a zero-knowledge proof π to prove whether its corpus data satisfies the arithmetic circuit F according to its own corpus data set, proof key pk, random blinding factor r and data selector. i (C i and π i The generation process is prior art and will not be described in detail here), and then the Pedersen commitment C i And zero-knowledge proof π i Send back to the system (the sending process can be based on an encryption algorithm); Pedersen commitment C sent by the system using the verification key vk and the participating client i And zero-knowledge proof π i As input to the arithmetic circuit F to verify the knowledge proof π i When the output result of the arithmetic circuit F is 1, it proves that the corresponding corpus data set can satisfy the arithmetic circuit F and thus meets the quality requirements. When the output result of the arithmetic circuit F is 0, the corresponding corpus data set does not satisfy the arithmetic circuit F and thus does not meet the quality requirements.

[0058] The above method not only solves the problem of data quality verification, but also brings additional benefits. Since participants do not need to directly expose their data, this increases the willingness to participate in federated learning and may attract more high-quality data providers to participate. At the same time, since the verification process is automated, it can greatly improve verification efficiency and enable the system to handle large-scale participation requests.

[0059] Among them, federated learning is the process of multiple participants collaborating to train machine learning models under the premise of protecting data privacy; and the federated learning in step A4 refers to central federated learning, which requires the aggregation server and multiple participants to collaborate to complete the training of the machine learning model.

[0060] Preferably, step A4 comprises: A401. The aggregation server initializes global parameters according to task requirements and sends them to each participating end based on an encryption algorithm; global parameters include model type and model parameters; A402. Each participating terminal uses its own corpus data set to perform model optimization training based on global parameters, obtains model update parameters, and uploads the model update parameters to the aggregation server based on an encryption algorithm; A403. After all participating terminals have completed uploading the model update parameters, the aggregation server screens the model update parameters of each participating terminal, performs aggregation calculation with the screened valid model update parameters to obtain the aggregation parameters, and calculates the single-round contribution of each participating terminal; A404. The aggregation server updates the model parameters according to the aggregation parameters, and uses the verification data set to verify whether the model after the updated model parameters meets the task requirements. If not, the updated model parameters are sent to each participating terminal based on the encryption algorithm, and the process returns to step A402. If it is satisfied, the contribution value of each participating terminal is calculated based on the single-round contribution degree, and the iteration ends.

[0061] This technical solution effectively solves the problems of model training, parameter updating and contribution calculation in federated learning through a series of carefully designed steps. It realizes multi-party collaborative model training while protecting data privacy, and improves the quality and robustness of the model through parameter screening and aggregation mechanisms. The iterative training and verification mechanism ensures that the model can meet the task requirements, while the contribution calculation provides a scientific basis for the fair distribution of rewards. The encryption measures throughout the process further enhance the security of data interaction.

[0062] In step A401, the aggregation server can initialize and send global parameters in a variety of ways. For example, an appropriate model architecture (i.e., model type) can be selected according to the task type, such as a neural network model such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a linear model. Model parameters (such as at least one of weights, gradients, etc.) need to be initialized before the training process can be started. The initialization method can be random initialization, pre-trained model migration, etc.

[0063] The global parameters may also include hyperparameters and data preprocessing instructions. The hyperparameters may include at least one of a learning rate, a batch size, a number of iterations, an optimizer type and its configuration, a regularization coefficient, etc. The data preprocessing instructions may include at least one of operations such as data normalization, feature selection, and feature scaling.

[0064] In step A402, during the local model training process, the participating end can use a variety of optimization algorithms, such as stochastic gradient descent (SGD), Adam, RMSprop, etc. When the optimizer type is specified in the global parameters, the corresponding optimizer is used for optimization. In order to further protect privacy, a differential privacy mechanism can be introduced in the local training process to prevent the model from reversely deducing the original data by adding noise to the gradient.

[0065] Among them, the model update parameters refer to the model parameters or gradients obtained by the participating terminals after local training, so that the aggregation parameters correspond to the aggregated model parameters or aggregated gradients, which can be uploaded to the aggregation server in an encrypted transmission manner.

[0066] Furthermore, in step A403, the aggregation server screens the model update parameters of each participating terminal, including: B1. Use the validation data set to verify the impact of the model update parameters submitted by each participant in this round on the model performance, and filter out the model update parameters that cause the model performance to deteriorate; B2. Compare the remaining model update parameters with the aggregate parameters calculated in the previous round, and filter out the model update parameters that differ too much from the aggregate parameters calculated in the previous round; B3. Take the remaining model update parameters as valid model update parameters.

[0067] This technical solution solves the quality control problem of model update parameters in federated learning through a multi-step screening mechanism. First, the validation data set is used to verify the impact of the model update parameters uploaded by each participating end on the model performance, and those parameters that cause the model performance to deteriorate are screened out. This step can effectively prevent the negative impact of poisoning attacks or low-quality parameters on model performance by malicious participants. Secondly, by comparing the difference between the remaining parameters and the aggregation parameters of the previous round, those parameters with too large differences are further screened out, which helps to maintain the stability of model training and prevent drastic fluctuations in model updates. Finally, the parameters after two rounds of screening are used as valid parameters for subsequent aggregation calculations. This multiple screening mechanism can effectively improve the quality of model update parameters in the federated learning process, thereby improving the performance and reliability of the final model.

[0068] In step B1, the impact of model update parameters on model performance can be verified by using the validation data set in the following way: the aggregation server can use the validation data set to evaluate the model update parameters submitted by each participating terminal, calculate the performance indicators (such as accuracy, F1 score, etc.) of the updated model on the validation data set, and compare these performance indicators with the performance of the current global model (for the first round of learning, the current global model is the initialized model; for non-first round of learning, the current global model is the model obtained after the previous round of learning). If the model update parameters of a participating terminal cause the performance of the current global model to deteriorate, the model update parameters will be screened out.

[0069] For the performance indicator accuracy, we will explain this as an example. For each participating end’s model update parameter, first copy the current global model as the comparison model (each participating end’s model update parameter corresponds to a comparison model, and these comparison models are independent of each other), and use the model update parameter of a single participating end to update the model parameters of the comparison model (if the model update parameter is a model parameter, then directly use the model update parameter to replace the model parameter of the comparison model; if the model update parameter is a gradient, then use the model parameter update formula to calculate the model parameter update value according to the gradient, and then use the model parameter update value to replace the model parameter of the comparison model. The model parameter update formula is , To compare the current model parameters of the model, Update values ​​for model parameters, is the learning rate, which can be set according to actual needs or extracted from global parameters. The current global model and the updated comparison model are verified using the verification data set to obtain the accuracy of the current global model and the updated comparison model. If the accuracy of the updated comparison model is lower than that of the current global model, the corresponding participating end may transmit invalid or malicious gradients, thereby screening out the corresponding model update parameters.

[0070] In step B2, for the first round of learning, the remaining model update parameters are not screened out. For non-first round learning, the norm or Euclidean distance between each remaining model update parameter and the aggregated parameter calculated in the previous round can be calculated as a deviation value. If the deviation value exceeds the deviation threshold, the corresponding model update parameter is screened out. The deviation threshold can be a preset fixed value or can be determined according to the following learning rate annealing formula: ; in, is the learning round number of this round of learning, is the deviation threshold in the t-th round of learning, is the preset initial value of the deviation threshold (can be set between 0.1 and 1), e is the base of the natural logarithm, is the decay rate exponent (which can be set between 0.001 and 0.1). Determining the deviation threshold through the above learning rate annealing formula is helpful to ensure the rationality of setting the deviation threshold in the entire learning process.

[0071] Preferably, after step B2 and before step B3, the method may further include the following steps: B4. Further screen the remaining model update parameters based on the Multi-Krum algorithm.

[0072] After step B2 and before step B3, a step of further screening the remaining model update parameters based on the Multi-Krum algorithm is introduced. This technical feature plays a key role in solving the problem of further screening model update parameters. The Multi-Krum algorithm can effectively identify and exclude abnormal or malicious model update parameters, and has an effective defense effect against perturbation attacks (such perturbation attacks are usually difficult to detect through the methods of steps B1 and B2), thereby improving the robustness and security of the federated learning system.

[0073] Specifically, step B4 includes: Calculate the Euclidean distance between the remaining model update parameters; the remaining model update parameters are the remaining model update parameters; For each remaining model update parameter, the sum of the minimum nf-2 (i.e., the first nf-2 after sorting from small to large) Euclidean distances related to the corresponding remaining model update parameter is calculated as the relevant score of the corresponding remaining model update parameter; n is the total number of remaining model update parameters, f is the number of suspicious participating terminals among the participating terminals corresponding to the remaining model update parameters (wherein f is set according to historical experience, and can also be set to the number of participating terminals whose model update parameters have been screened out in this federated learning process among the participating terminals corresponding to each remaining model update parameter); the Euclidean distance related to the corresponding remaining model update parameter refers to the Euclidean distance between the corresponding remaining model update parameter and other remaining model update parameters (for example, the Euclidean distance dij between the i-th and j-th remaining model update parameters is the Euclidean distance related to the i-th remaining model update parameter, and is also the Euclidean distance related to the j-th remaining model update parameter); The remaining model update parameters are sorted in ascending order according to the relevant scores, and the remaining model update parameters with sorting numbers exceeding m are screened out; m is a preset positive integer, and m≤nf-2.

[0074] Furthermore, in step A403, the screened effective model update parameters are used to perform aggregation calculation to obtain an aggregate parameter. Specifically, an average value of the effective model update parameters may be calculated as the aggregate parameter.

[0075] The contribution value refers to the contribution of each participating end to the improvement of the final model performance during the federated learning process. The sum of the single-round contributions of each participating end in each round of learning can be calculated as the contribution value of each participating end.

[0076] In some implementations, in step A403, calculating the single-round contribution of each participating terminal includes: Using the Shapley value algorithm, update the parameters according to the model of each participating end and calculate the marginal value of each participating end's single-round contribution; Based on the screening results of the model update parameters, calculate the proportion of the single-round reputation contribution evaluation value of each participating terminal; According to the time consumed by each participating terminal in this round of learning, calculate the proportion of the single-round efficiency contribution evaluation value of each participating terminal; The single-round contribution of each participating terminal is calculated based on the proportion of single-round contribution marginal value, single-round reputation contribution evaluation value and single-round efficiency contribution evaluation value.

[0077] By comprehensively considering the contribution margin, reputation contribution, and efficiency contribution, the method of this application can more comprehensively and fairly evaluate the contribution of each participant in the federated learning process. This evaluation method helps solve the problem of how to fairly distribute rewards, and can also motivate participants to provide high-quality data and computing resources.

[0078] Among them, the Shapley value algorithm is used to calculate the contribution margin ratio of each participant. The Shapley value algorithm can fairly evaluate the contribution of each participant to the final result. Specifically including: A combination set N is constructed based on all valid model update parameters; the combination set N includes subsets s formed by all possible combinations of valid model update parameters. Specifically, assuming there are m valid model update parameters, the combination set N includes subsets consisting of a single valid model update parameter (a total of m), subsets consisting of two valid model update parameters (the number of subsets is determined according to the permutation and combination formula determine), … and a subset of m valid model update parameters (a total of one); According to the verification data set, the parameter correctness index of each subset s in N is calculated; for example, for each subset s, the current global model can be copied as a reference model (each subset s corresponds to a reference model, and each reference model is independent), and the average value of the valid model update parameters in the subset s is calculated to update the model parameters of the reference model. The current global model and the updated reference model are verified with the verification data set respectively to obtain the accuracy of the current global model and the updated reference model, and the accuracy of the updated reference model is divided by the accuracy of the current global model as the parameter correctness index of the subset s; The contribution weighting factor of each subset s is calculated according to the following formula: ; in, is the contribution weighting factor of subset s, is the length of subset s, ! is the factorial symbol; The single-round contribution margin value of each participating terminal is calculated according to the following formula: ; in, The marginal value contributed by the i-th participant in the t-th round of learning, is the set consisting of all subsets of the combined set N that contain the i-th participating terminal, is the parameter correctness index of subset s, represents the parameter correctness index after excluding the effective model update parameters of the i-th participant in the subset s (after excluding the effective model update parameters of the i-th participant in the subset s, it is calculated according to the parameter correctness index calculation method in the previous text. If the subset s becomes an empty set after excluding the effective model update parameters of the i-th participant in the subset s, then is 0), is the set of participating terminals corresponding to the effective model update parameters in the tth round of learning, Indicates that the i-th participating end belongs to The participating end, Indicates that the i-th participant does not belong to Participating end in The marginal value of a single round contribution of each participating terminal is calculated according to the following formula: ; in, is the marginal value ratio of the single round contribution of the i-th participant in the t-th round of learning, is the marginal value contributed by the jth participant in the tth round of learning, and K is the total number of participants.

[0079] Furthermore, based on the screening results of the model update parameters, the proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated, including: The attenuation index of each participating terminal is calculated according to the following penalty rule formula: ; ; in, is the decay index of the ith participant in the tth round of learning, is the learning round number bias of the ith participant, is the learning round number when the model update parameters of the i-th participant were last screened out, is the attenuation window period (can be adjusted according to actual needs); The proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated according to the following formula: ; in, is the proportion of the single-round reputation contribution evaluation value of the i-th participant in the t-th round of learning, is the decay index of the jth participant in the tth round of learning, and K is the total number of participants.

[0080] in, and Represents the intensity of punishment for dishonest behavior. When a participant behaves dishonestly in a round of learning (its model update parameters are screened out), in the subsequent rounds During the round of learning, the decay exponent will decrease (less than 1), and the amount of decrease will decrease as the number of learning rounds increases.

[0081] Furthermore, according to the time consumed by each participating terminal in this round of learning, the proportion of the single-round efficiency contribution evaluation value of each participating terminal is calculated, including: The efficiency index of each participating terminal is calculated according to the following formula: ; in, is the efficiency index of the ith participant in the tth round of learning, is the time consumed by the i-th participant in the t-th round of learning, is the time consumed by the jth participant in the tth round of learning; The single-round efficiency contribution evaluation value ratio of each participating terminal is calculated according to the following formula: ; in, is the proportion of the single-round efficiency contribution evaluation value of the i-th participant in the t-th round of learning, is the efficiency index of the jth participant in the tth round of learning.

[0082] Further, the single-round contribution of each participating terminal is calculated according to the proportion of the single-round contribution marginal value, the proportion of the single-round reputation contribution evaluation value and the proportion of the single-round efficiency contribution evaluation value, specifically including calculating the single-round contribution of each participating terminal according to the following formula: ; in, is the single-round contribution of the i-th participant in the t-th round of learning, , , is a weight value (which can be set according to actual needs, for example, it can be set when the system creates a task, and each participating client is notified when the task is released, so that users of the participating clients can weigh whether to participate).

[0083] In step A404, when updating the model parameters according to the aggregated parameters, if the aggregated parameters are aggregated model parameters, the model parameters of the current global model are directly replaced by the aggregated parameters; if the aggregated parameters are aggregated gradients, the model parameter update formula (see above for details) is used to calculate the model parameter update value according to the aggregated gradient, and then the model parameter update value is used to replace the model parameters of the current global model.

[0084] In step A404, the validation data set is used to verify whether the performance indicators (such as accuracy, F1 score, etc.) of the model after the model parameters are updated meet the requirements.

[0085] In step A404, the sum of the single-round contribution of each participating terminal in each round of learning during this federated learning process may be calculated as the contribution value of each participating terminal.

[0086] Specifically, in step A5, the model obtained after completing federated learning is sent to the consumer client, which verifies it using the verification data set. When the verification is passed, settlement is performed.

[0087] Furthermore, in step A6, the reward amount of each participating terminal can be calculated according to the following formula, and the reward can be distributed according to the calculation result: ; in, is the reward amount of the i-th participant, is the contribution value of the ith participant, is the contribution value of the jth participant, The total settlement amount (which can be the total amount paid by the consumer client, or the product of the total amount paid by the consumer client and a preset percentage value greater than 0 and less than 1. If it is the latter, the part not distributed to the participating end is the platform commission).

[0088] refer to Figure 2 , the present application provides a federated learning service system for data circulation, including a business server 100 and a consumer client 200, an aggregation server 300 and a plurality of participant clients 400 that are communicatively connected to the business server 100; The consumer client 200 is used to initiate a task request to the business server 100; The service server 100 is used for: In response to the task request initiated by the consumer client 200, review the task request (refer to step A1 above for details); After the task request is reviewed, the task is issued to multiple participating clients 400 based on the task request (refer to step A2 above for details); Perform qualification verification on the participant client 400 that submits the request to participate in the task, to verify whether the corpus data set held by the corresponding participant client 400 meets the quality requirements, and determine the participant client 400 that passes the qualification verification as the participant terminal participating in this task; the qualification verification process is based on the encryption algorithm to exchange data (for details, refer to step A3 above); The aggregation server 300 and the participating terminals are used to perform federated learning to generate a model that meets the task requirements and to count the contribution value of each participating terminal. During the federated learning process, the aggregation server 300 and the participating terminals exchange data based on an encryption algorithm, and the aggregation server 300 screens the model update parameters uploaded by each participating terminal before performing each aggregation calculation (refer to step A4 above for details). Send the model to the consumer client 200 for acceptance and settlement (refer to step A5 above for details); After the acceptance and settlement are completed, rewards will be distributed to each participant based on their contribution value (please refer to step A6 above for details).

[0089] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0090] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0091] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0092] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0093] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A federated learning service method for data circulation, characterized in that: Includes steps: A1. In response to a task request initiated by a consumer client, review the task request; A2. After the task request is reviewed, the task is issued to multiple participating clients based on the task request; A3. Perform qualification verification on the client of the participant who submitted the request to participate in the task, so as to verify whether the corpus data set held by the corresponding client of the participant meets the quality requirements, and determine the client of the participant who passes the qualification verification as the participant of this task; The qualification verification process is based on encryption algorithms for data interaction; A4. Use the aggregation server and the participating terminals to perform federated learning to generate a model that meets the task requirements and count the contribution value of each participating terminal; during the federated learning process, the aggregation server and the participating terminals exchange data based on an encryption algorithm, and the aggregation server screens the model update parameters uploaded by each participating terminal before performing aggregation calculation each time; A5. Send the model to the consumer client for acceptance and settlement; A6. After the acceptance and settlement are completed, rewards will be issued to each participating terminal according to the contribution value of each participating terminal.

2. The method for providing a federated learning service for data circulation according to claim 1, characterized in that: Step A1 includes: Receiving a task request including task requirements sent by the consumer client; Assess whether the stated mission requirements can be met; Otherwise, it is determined that the task request has not passed the review; If so, it is determined that the task request has passed the review, and the verification data set provided by the consumer client is received.

3. The method for providing a federated learning service for data circulation according to claim 1, characterized in that: Step A3 includes: Based on the encryption algorithm, the zero-knowledge proof technology is used to verify the quality of the corpus data set held by each of the participating client terminals that submits the task participation request, so as to verify whether each of the corpus data sets meets the quality requirements; The participant client corresponding to the corpus data set that meets the quality requirements is determined to have passed the qualification verification, and the participant client that has passed the qualification verification is determined as a participant terminal participating in this task.

4. The method for providing a federated learning service for data circulation according to claim 2, characterized in that: Step A4 includes: A401. The aggregation server initializes global parameters according to task requirements and sends the global parameters to each participating terminal based on an encryption algorithm; the global parameters include model type and model parameters; A402. Each of the participating terminals performs model optimization training based on the global parameters using the corpus data set it holds, obtains model update parameters, and uploads the model update parameters to the aggregation server based on an encryption algorithm; A403. After all the participating terminals have completed uploading the model update parameters, the aggregation server screens the model update parameters of each participating terminal, performs aggregation calculation with the screened valid model update parameters to obtain the aggregation parameters, and calculates the single-round contribution of each participating terminal; A404. The aggregation server updates the model parameters according to the aggregation parameters, and uses the verification data set to verify whether the model after the updated model parameters meets the task requirements. If not, the updated model parameters are sent to each participating terminal based on the encryption algorithm, and the process returns to execute step A402. If it is satisfied, the contribution value of each participating terminal is calculated based on the single-round contribution degree, and the iteration ends.

5. The method for providing a federated learning service for data circulation according to claim 4, characterized in that: In step A403, the aggregation server screens the model update parameters of each of the participating terminals, including: B1. Using the verification data set to verify the impact of the model update parameters submitted by each participating terminal in this round on the model performance, and screening out the model update parameters that cause the model performance to deteriorate; B2. Compare the remaining model update parameters with the aggregation parameters calculated in the previous round, and filter out the model update parameters that differ too much from the aggregation parameters calculated in the previous round; B3. Take the remaining model updating parameters as effective model updating parameters.

6. The method for providing a federated learning service for data circulation according to claim 5, characterized in that: After step B2 and before step B3, the method further includes the following steps: The remaining model update parameters are further screened based on the Multi-Krum algorithm.

7. The method for providing a federated learning service for data circulation according to claim 4, characterized in that: In step A403, calculating the single-round contribution of each participating terminal includes: Using the Shapley value algorithm, according to the model update parameters of each participating terminal, the single-round contribution marginal value ratio of each participating terminal is calculated; Based on the screening results of the model update parameters, the proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated; Calculate the proportion of the single-round efficiency contribution evaluation value of each participating terminal according to the time consumed by each participating terminal in the current round of learning; The single-round contribution degree of each participating terminal is calculated according to the proportion of the single-round contribution marginal value, the proportion of the single-round reputation contribution evaluation value and the proportion of the single-round efficiency contribution evaluation value.

8. The method for federated learning services for data circulation according to claim 7, characterized in that: Based on the screening results of the model update parameters, the proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated, specifically including: The attenuation index of each participating terminal is calculated according to the following penalty rule formula: ； ； in, is the learning round number of this round of learning, is the attenuation index of the i-th participating terminal in the t-th round of learning, e is the base of the natural logarithm, is the decay rate index, is the learning round number bias of the i-th participating terminal, is the learning round number when the model update parameters of the i-th participating terminal were most recently screened out, is the set of participating terminals corresponding to the effective model update parameters in the tth round of learning, Indicates that the i-th participating terminal belongs to The participating end, Indicates that the i-th participating terminal does not belong to The participating end, is the attenuation window period; The proportion of the single-round reputation contribution evaluation value of each participating terminal is calculated according to the following formula: ； in, is the proportion of the single-round reputation contribution evaluation value of the i-th participant in the t-th round of learning, is the attenuation index of the jth participating terminal in the tth round of learning, and K is the total number of participating terminals.

9. The method for providing a federated learning service for data circulation according to claim 7, characterized in that: According to the time consumed by each of the participating terminals in this round of learning, the proportion of the single-round efficiency contribution evaluation value of each of the participating terminals is calculated, specifically including: The efficiency index of each participating terminal is calculated according to the following formula: ； in, is the learning round number of this round of learning, is the efficiency index of the i-th participant in the t-th round of learning, is the time consumed by the i-th participant in the t-th round of learning, is the time consumed by the jth participating terminal in the tth round of learning, K is the total number of participating terminals, is the set of participating terminals corresponding to the effective model update parameters in the tth round of learning, Indicates that the i-th participating terminal belongs to The participating end, Indicates that the i-th participating terminal does not belong to Participating end in The single-round efficiency contribution evaluation value ratio of each participating terminal is calculated according to the following formula: ； in, is the proportion of the single-round efficiency contribution evaluation value of the i-th participant in the t-th round of learning, is the efficiency index of the j-th participating terminal in the t-th round of learning.

10. A federated learning service system for data circulation, characterized in that: It includes a business server and a consumer client, an aggregation server and a plurality of participant clients that are communicatively connected to the business server; The consumer client is used to initiate a task request to the business server; The business server is used for: In response to a task request initiated by a consumer client, review the task request; After the task request is reviewed, the task is issued to multiple participating clients based on the task request; Perform qualification verification on the client of the participant who submitted the request to participate in the task, so as to verify whether the corpus data set held by the corresponding client of the participant meets the quality requirements, and determine the client of the participant who passes the qualification verification as the participant of this task; The qualification verification process is based on encryption algorithms for data interaction; The aggregation server and the participating terminals are used to perform federated learning to generate a model that meets the task requirements and to count the contribution value of each participating terminal. During the federated learning process, the aggregation server and the participating terminals exchange data based on an encryption algorithm, and the aggregation server screens the model update parameters uploaded by each participating terminal before performing aggregation calculation each time. Sending the model to the consumer client for acceptance and settlement; After the acceptance and settlement are completed, rewards will be issued to each participating terminal according to the contribution value of each participating terminal.

Images