Federated learning contribution calculation method and federated learning contribution calculation and profit-sharing system

Abstract

To provide a method and system for calculating contributions in federated learning and a system for performing profit-sharing on the basis of the calculated contributions.SOLUTION: A method comprises the steps of: collaboratively developing, by all of participants in federated learning, a federated aggregation model by federated learning, using their own local datasets; excluding participation of at least one first member in all participants, and then the remained participants excluding the at least one first member collaboratively developing a first contribution model by a federated learning method; and, comparing the value of the first contribution model with the value of the federated aggregation model to obtain the contribution of the first member. The method of the present invention is capable of calculating the contribution of each of the members or group contributions of some members in the federated learning by few additional information and few additional calculations, so as to achieve profit sharing according to the contributions.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a method for calculating contribution in federated learning and a system for calculating contribution and distributing profits in federated learning, and in particular to a method and system that enables calculation of the contribution of each member in federated learning, and a system that enables profit distribution based on the calculated contribution. [Background technology]

[0002] With the advancement of medical engineering technology, the application of regenerative medicine in clinical treatment of diseases is becoming more and more diverse, and its research and application are attracting more and more attention from all walks of life. Currently, regenerative medicine is mainly applied to organ repair, immune cell therapy, and stem cell therapy. Cell therapy involves the procedure of culturing or processing human cells outside the body, and then transplanting the cultured cells into an individual's body for use. Therefore, how to effectively culture cells is a crucial element in regenerative medicine. Previous research has shown that the cells in each person's body have unique characteristics, and it is necessary to customize cells suitable for treatment according to each situation, which increases the complexity and difficulty of cell processes. On the other hand, when there is a large amount of cell data and bioprocess data, machine learning can generate predictive models that evaluate the effects of each cell culture condition and process parameter. Thus, machine learning can reduce the complexity and difficulty of customizing cell processes. However, most of these cell data and bioprocess data belong to very sensitive and highly confidential personal data, such as cell data of patients in various hospitals or laboratories. Conventional aggregated machine learning methods require all bioprocessing data to be aggregated in a central server for analysis and training, which may be impractical due to constraints such as regulations, privacy protection, and commercial considerations. In addition, aggregated machine learning methods may face similar problems not only in biomedical engineering, but also in any technical field where data is highly sensitive and confidential.

[0003] Federated learning is a type of machine learning technology, also known as collaborative learning, in which an algorithm is trained on multiple distributed edge devices or servers with local data samples. This method is significantly different from traditional aggregated machine learning technology, in which all local datasets are uploaded to one server for machine learning training. On the other hand, federated learning is composed of multiple members. Each member performs machine learning training on its own dataset to generate its own trained model, and uploads the trained model parameters to a central server. The central server assigns weights to all the trained model parameters and aggregates them to form a federated model. This federated model can be redistributed to each member for use. The central server only aggregates the trained model parameters, and does not directly perform machine learning analysis or processing on each member's dataset. Therefore, the highly sensitive and confidential data samples and information in each member's dataset are not leaked to the outside, and the confidentiality of the data is protected.

[0004] The federated model finally generated by federated learning is completed by the models provided by each member through training. Therefore, each member has his / her own contribution to the federated model. In addition, if the federated model is commercially used and profitable due to its reliability and prediction accuracy, each member may try to obtain a certain profit share according to his / her own contribution. In addition, the contribution of each member also represents the influence of each member on the federated model, and can be used to evaluate whether the weight of each member in the federated model is appropriate. However, currently, there is no method in federated learning that can calculate the contribution of each member.

[0005] Therefore, it is necessary to develop a method and system that can efficiently calculate the contribution of each member in federated learning, for application to member profit distribution and other functions based on contribution. Summary of the Invention [Problem to be solved by the invention]

[0006] In view of the above, and in order to solve the above-mentioned problems in the related art, the present invention provides a method for calculating a contribution in associative learning. [Means for solving the problem]

[0007] According to a specific embodiment of the present invention, a method for calculating contribution in federated learning includes the steps of: a central server accepting a plurality of trained model parameters, the trained model parameters being generated by machine learning a plurality of local datasets, respectively; a central server running an aggregation algorithm on the trained model parameters to generate a federated model and storing the federated model; a central server running an aggregation algorithm on the trained model parameters excluding or simulating excluding at least one first trained model parameter to generate a first contribution model; and a central server comparing a value of the first contribution model with a value of the federated model to generate a contribution of the first trained model parameter.

[0008] The trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, the central server generates a federated model by running an aggregation algorithm on each round of machine learning training, and the central server generates a first contribution model by running an aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in a last machine learning training.

[0009] The trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, and the central server generates a federated model by running an aggregation algorithm on each round of machine learning training, and generates a first contribution model by running an aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in each round of machine learning training in a simulated manner.

[0010] The trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset. The central server generates a federated model by executing the aggregation algorithm in each round of machine learning training. The central server generates a first contribution model by executing the aggregation algorithm on the trained model parameters simulating the exclusion of at least one first trained model parameter in the kth machine learning training, counting down from the last, where k is an integer between 1 and the number of times the first trained model parameter has participated in the aggregation algorithm of the federated model.

[0011] The trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset. The central server generates a federated model corresponding to each round of machine learning by executing an aggregation algorithm in each round of machine learning training. The central server generates a first contribution model corresponding to each round of machine learning by executing an aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in each round of machine learning training. The central server also compares the value of the first contribution model in each round of machine learning with the value of the federated model, and performs an average or weighted average to generate an average contribution as the contribution of the first trained model parameter.

[0012] The method further includes a step of the central server determining whether the contribution of at least one first trained model parameter is a positive value or a negative value, and if the contribution is a negative value, the central server replacing the federated model with the first contribution model as an updated federated model.

[0013] In addition, the present invention further solves the above-mentioned conventional problems and provides a contribution calculation and profit distribution system in federated learning that can be applied to the distribution of profits among members in federated learning.

[0014] According to a specific embodiment of the present invention, the contribution calculation and profit distribution system in federated learning includes a plurality of artificial intelligence model training devices and a central server. The central server further includes a calculation module, a contribution calculation module, and a profit distribution module. The plurality of artificial intelligence model training devices generate a plurality of trained model parameters by machine learning a plurality of local data sets, respectively. The central server is connected to each artificial intelligence model training device and receives the trained model parameters. The calculation module of the central server generates a federated model by executing an aggregation algorithm on the received trained model parameters. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters excluding or simulating the excluding at least one first trained model parameter. Then, the value of the first contribution model is compared with the value of the federated model to generate a contribution of the first trained model parameter. The profit distribution module is connected to the contribution calculation module to receive the contribution of the first trained model parameter, and calculates a profit distribution ratio of the first trained model parameter in the federated model based on the contribution.

[0015] The artificial intelligence model training device performs multiple rounds of machine learning training on the local dataset, and the calculation module generates a federated model by executing an aggregation algorithm in each round of machine learning training, and the contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in the last machine learning training.

[0016] The artificial intelligence model training device performs multiple rounds of machine learning training on the local data set, and the calculation module generates a federated model by executing an aggregation algorithm in each round of machine learning training. In addition, the contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in each round of machine learning training in a simulated manner.

[0017] The artificial intelligence model training device performs multiple rounds of machine learning training on the local data set. The calculation module generates the federated model by executing an aggregation algorithm in each round of machine learning training. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters simulating the exclusion of at least one first trained model parameter in the machine learning training up to the kth round, counting down from the last, where k is an integer between 1 and the number of times the first trained model parameter has participated in the aggregation algorithm of the federated model.

[0018] The artificial intelligence model training device performs multiple rounds of machine learning training on the local data set. The calculation module executes the aggregation algorithm in each round of machine learning training to generate a federated model corresponding to each round of machine learning. The contribution calculation module executes the aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in each round of machine learning training in a simulation manner to generate a first contribution model corresponding to each round of machine learning. The contribution calculation module compares the value of the first contribution model in each round of machine learning with the value of the federated model, and performs an average or weighted average to generate an average contribution as the contribution of the first trained model parameter. Effect of the Invention

[0019] In summary, the method and system for calculating the contribution in federated learning of the present invention can calculate the contribution in the federated model for the model trained by the members participating in the federated learning. This can also achieve the function of distributing profits to the members. Alternatively, the weight of the model in the federated model can be adjusted based on the contribution of each member, so that the federated model generated by the federated learning can have a more accurate prediction effect. [Brief description of the drawings]

[0020] [Figure 1] FIG. 1 shows a flowchart of the steps of a method for calculating a contribution in associative learning according to a specific embodiment of the present invention. [Diagram 2] FIG. 2 shows a schematic diagram of a contribution calculation and profit distribution system in federated learning according to another specific embodiment of the present invention. [Diagram 3] FIG. 3 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. [Figure 4] FIG. 4 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. [Diagram 5] FIG. 5 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. [Figure 6] FIG. 6 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. [Figure 7] FIG. 7 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. [Figure 8] FIG. 8 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] In order to make the advantages, spirits and features of the present invention easier and clearer to understand, the present invention will be described and discussed in detail with reference to the drawings, using specific examples. It should be noted that these specific examples are merely representative specific examples of the present invention, and the specific methods, devices, conditions, materials, etc. illustrated do not limit the present invention or the corresponding specific examples. In addition, each component in the drawings is used only to represent their relative positions, and is not described based on the actual ratio. In addition, the step numbers of the present invention are only used to distinguish different steps, and do not represent the order of steps. The above will be explained in advance.

[0022] Please refer to Figures 1 and 2. Figure 1 shows a flowchart of the steps of a method for calculating contribution in federated learning based on a specific embodiment of the present invention. Figure 2 shows a schematic diagram of a system 2 for calculating contribution and sharing profits in federated learning based on another specific embodiment of the present invention. It should be noted that each step of the method in Figure 1 can be performed by different units of the system in Figure 2. Therefore, in the following, each step of the method in Figure 1 will be described using the system in Figure 2.

[0023] As shown in FIG. 2, the contribution calculation and profit sharing system 2 in the federated learning of this specific embodiment includes a plurality of artificial intelligence model training devices 20 and a central server 22 connected to these artificial intelligence model training devices 20. Each artificial intelligence model training device 20 performs machine learning on its local data set to generate a trained model. In practice, the artificial intelligence model training device 20 is a distributed node capable of independently executing a complete machine learning algorithm, and for example, in the above-mentioned field of regenerative medicine, it can be a computer device capable of executing a machine learning algorithm in each hospital or laboratory. In addition, the local data set is a sample data set for training, verification, and testing held by each hospital or laboratory. Each artificial intelligence model training device 20 can transmit the model parameters after training to the central server. Note that, although only two artificial intelligence model training devices 22 are illustrated in FIG. 2, in practice, this is not limited to this, and any number of artificial intelligence model training devices and trained model parameters generated by the devices can be used in the present invention. The central server 22 can receive the trained model parameters trained by these artificial intelligence model training devices 20, and includes a calculation module 220, a contribution calculation module 222, and a profit sharing module 224. The contribution calculation module 222 is connected to the calculation module 220, and the profit sharing module 224 is connected to the contribution calculation module 222. In practice, the calculation module 220, the contribution calculation module 222, and the profit sharing module 224 are built into a computing unit of the central server 22, such as a central processing unit of a computer device or a server, or a computing chip. Alternatively, the functions of the calculation module 220, the contribution calculation module 222, and the profit sharing module 224 may be respectively performed by different computing units connected to each other.

[0024] The calculation module 220 can generate the federated model by executing an aggregation algorithm for all the received trained model parameters. In the actual operation of the aggregation algorithm, the calculation module 220 assigns weights to each trained model parameter before the parameters can be aggregated. It should be noted that the aggregation algorithm for federated learning is common knowledge in the art, and will not be described in detail here. It should be noted that any aggregation algorithm for generating a federated model can be applied to the present invention. The contribution calculation module 222 firstly simulates excluding at least one first trained model parameter, which is at least one of the trained model parameters, from all the trained model parameters, and then executes a similar aggregation algorithm for the other trained model parameters after excluding or simulating excluding the at least one first trained model parameter, to obtain a first contribution model. Then, the value of the first contribution model is compared with the value of the federated model, so that the contribution of the first trained model parameter to the federated model can be known. In detail, if the first trained model parameters contribute to the federated model, when the first trained model parameters are excluded from all trained model parameters, the value of the first contribution model generated by the same aggregation algorithm should be reduced compared to the federated model. Therefore, it is possible to estimate the contribution of the first trained model parameters from the reduced value. The value of the above first contribution model and federated model can be estimated from the accuracy of each model. For example, when a value function is defined in which the relationship between the value and the accuracy of the model is linear, it is shown that the higher the accuracy of the model, the higher the value. In practice, the value function can be jointly determined by all federated members and is not limited to the linear relationship described above.

[0025] The profit sharing module 224 can determine the profit sharing ratio of the artificial intelligence model training device that provided the first trained model based on the contribution or collective contribution of at least one first trained model parameter. Therefore, when the federated model generates profit or profit, the members of the artificial intelligence model training device that generated the first trained model parameter can obtain profit based on the profit sharing ratio. In practice, the profit sharing module 224 uses a profit sharing ratio calculation algorithm based on the Shapley value or marginal contribution, which is a criterion of fairness, to determine the profit sharing ratio of the first trained model parameter or all trained model parameters based on the contribution.

[0026] As described above, the contribution calculation module 222 excludes or simulates the exclusion of at least one first trained model parameter. When the at least one first trained model parameter is a set of first trained model parameters, i.e., when the participation of one federation member is simulated, the contribution calculated after comparing the generated first contribution model with the value of the federation model is the individual contribution of the excluded federation member. In contrast, when the at least one first trained model parameter is a set of first trained model parameters, i.e., when the participation of multiple federation members is simultaneously excluded or simulated, the contribution calculated after comparing the generated first contribution model with the value of the federation model is the collective contribution of the excluded multiple federation members. It should be noted that in practice, the collective contribution calculated by simultaneously excluding multiple federation members is not necessarily equal to the sum of the individual contributions calculated when the federation members are individually excluded in a simulated manner.

[0027] The method of calculating the contribution degree in the above-mentioned contribution degree calculation / profit distribution system 2 in the federated learning is as shown in Fig. 1. In the specific embodiment of Fig. 1, the method of calculating the contribution degree in the federated learning includes the following steps.

[0028] Step S10: The central server 22 receives a number of trained model parameters, which are generated by machine learning of a number of local data sets, respectively.

[0029] Step S12: The central server 22 generates and stores a federated model by running an aggregation algorithm on all the trained model parameters.

[0030] Step S14: The central server 22 generates a first contribution model by executing an aggregation algorithm on other trained model parameters excluding or simulating excluding at least one first trained model parameter.

[0031] Step S16: The central server 22 compares the value of the first contribution model with the value of the federated model to generate a contribution of at least one first trained model parameter (i.e., generate a contribution of the federated members of the at least one first trained model parameter).

[0032] In step S10, the central server 22 can receive trained model parameters from a plurality of artificial intelligence model training devices 20 as shown in FIG. 2. In step S12, the central server 22 can generate a federated model by using the calculation module 220 to weight and aggregate each trained model parameter using an aggregation algorithm. In step S14, the contribution calculation module 222 of the central server 22 can first simulate and exclude at least one first trained model parameter among all trained model parameters. Next, the contribution calculation module 222 obtains a first contribution model by executing an aggregation algorithm on other trained model parameters from which the at least one first trained model parameter has been simulated. In step S16, the contribution calculation module 222 can compare the value of the first contribution model with the value of the federated model, and reverse-estimate the individual contribution or collective contribution of at least one first trained model parameter based on the degree of decrease in value.

[0033] It should be noted that in the above specific embodiment, the trained model parameters provided by each member in the federated learning are aggregated by the central processing unit, but the central processing unit is merely a representative computing device for aggregating and generating the federated model and the first contribution model in the present invention, and is not limited to the central processing unit in the narrow sense. In other words, in practice, the central processing unit may be a cloud server. In this case, the trained model parameters of the federated members are received through the Internet, aggregated into the federated model and the first contribution model, and then the federated model is redistributed to each federated member through the Internet. Then, the contribution of each federated member is obtained by performing the above calculation.

[0034] The contribution calculation and profit sharing system 2 in federated learning in FIG. 2 can calculate the individual contribution or collective contribution of the trained model generated by each member or multiple members in federated learning by the method of calculating the contribution in federated learning in FIG. 1, so that the profit sharing ratio of each member or multiple members can be determined based on the contribution. However, the calculated contribution of each trained model is not only usable for the function of profit sharing. Now, refer to FIG. 3. FIG. 3 shows a flowchart of steps of a method of calculating the contribution in federated learning based on another specific embodiment of the present invention. Each step of the method in FIG. 3 can also be performed by different units of the system in FIG. 2. Therefore, each step of the method in FIG. 3 will be described below using the system in FIG. 2. As shown in FIG. 3, this specific embodiment is different from the above-mentioned specific embodiment in the following points. That is, the method of this specific embodiment further includes the following steps.

[0035] Step S17: The central server 22 adjusts the weight of the first trained model parameter in the federated model based on the contribution of the first trained model parameter.

[0036] The contribution is determined by the value of the trained model parameters to the federation model, and the higher the contribution, the greater the impact on the value of the federation model, while the opposite is true, the smaller the impact on the value. Therefore, the weights can be readjusted based on the contribution of each trained model parameter to make the federation model more accurate. It should be noted that the other steps in this specific embodiment are almost the same as the corresponding steps in the specific embodiment described above, and are not described in detail here.

[0037] In addition, if the contribution of the trained model is calculated to be a negative value, it means that the trained model has a negative effect on the federated model, and should be excluded. Now, refer to FIG. 4. FIG. 4 shows a flowchart of the steps of a method for calculating the contribution in federated learning according to another specific embodiment of the present invention. Each step of the method in FIG. 4 can be performed by different units of the system in FIG. 2. Therefore, in the following, each step of the method in FIG. 4 will be described using the system in FIG. 2. As shown in FIG. 4, the method of this specific embodiment further includes the following steps:

[0038] Step S180: The central server 22 determines whether the contribution of the first trained model parameter is a positive value or a negative value.

[0039] Step S182: If the contribution is a negative value, the central server replaces the association model with the first contribution model as the updated association model.

[0040] As described above, if the contribution calculation module 222 of the central server 22 calculates that the individual contribution or collective contribution of at least one first trained model parameter is a negative value, it means that adding the at least one first trained model parameter to the federated model will reduce the value of the federated model. Therefore, the central server 22 uses the first contribution model obtained by the aggregation algorithm from the other trained model parameters after excluding the at least one first trained model parameter as a new federated model, so that the federated model can have higher value and higher accuracy.

[0041] Typically, in order to continuously update an artificial intelligence model, machine learning training is periodically performed on a data set including old and new data, and the same is true for federated learning. In practice, each artificial intelligence model training device can update the trained model by performing machine learning multiple times on a local data set, and can transmit all trained model parameters completed in each machine learning round to a central server. The central server then executes an aggregation algorithm on all trained model parameters after each machine learning training round to generate a federated model for that round. For example, each artificial intelligence model training device (member of federated learning) may be set to update the trained model once every few months. In this case, all artificial intelligence model training devices upload the trained model parameters updated in that round to the central server, and the aggregation algorithm is executed to generate a federated model for that round. Therefore, by adjusting the weight of each trained model parameter when the next aggregation algorithm is executed based on the contribution of each trained model parameter calculated by the above method, a more optimized and more accurate federated model can be generated. In addition, when a new member joins the federated learning, it is possible to join directly from the next machine learning training and aggregation algorithm stage.

[0042] In practice, after a federated model is generated by an aggregation algorithm from trained model parameters generated in one machine learning stage, the federated model is redistributed to each federated member (artificial intelligence model training device). Therefore, in the next machine learning stage, even if each federated member uses its own local data set to train machine learning, the parameters of the completed federated model may be updated. Therefore, when calculating the contribution of the first trained model parameters, each machine learning and aggregation algorithm stage must be taken into account. Now, refer to FIG. 5. FIG. 5 shows a flowchart of steps of a method for calculating contribution in federated learning according to another specific embodiment of the present invention. Each step of the method in FIG. 5 can be performed by different units of the system in FIG. 2. Therefore, hereinafter, each step of the method in FIG. 5 will be described using the system in FIG. 2.

[0043] As shown in FIG. 5, the method of this specific embodiment includes the following steps:

[0044] Step S30: The central server 22 receives a number of trained model parameters, which are generated by performing machine learning multiple times (e.g., r times) on a number of local datasets, respectively.

[0045] Step S32: The central server 22 generates and stores a federated model by executing an aggregation algorithm on all trained model parameters in each machine learning training.

[0046] Step S340: The central server 22 generates a first contribution model by executing an aggregation algorithm on other trained model parameters, excluding the first trained model parameters in each machine learning training in a simulated manner.

[0047] Step S36: The central server 22 compares the value of the first contribution model with the value of the federated model to generate a contribution of the first trained model parameters.

[0048] In the method of this specific embodiment, when constructing the first contribution model, the federation member that generated the first trained model parameter is excluded from the beginning, and the remaining federation members perform local and global calculations completely all times (r times). Therefore, the final first contribution model is one in which the first trained model parameter does not participate in federated learning at all, and the difference in value from the generated federation model is the contribution of the first trained model parameter. In practice, if the artificial intelligence model training device 20 and the central server 22 both try to go back to the initial state and then completely perform all local and global calculations to get as close as possible to the actual contribution when constructing the first contribution model, the computational resources and costs will double. Therefore, in this specific embodiment, the simulation-based exclusion method (i.e., the simulation-based retroactive method) is used, but it may still consume considerable computational resources and costs. This is particularly noticeable when calculating individual contributions for each federation member.

[0049] As described above, when the first trained model parameters are excluded from the beginning in a simulated manner, the obtained contribution is closest to the actual situation, but still consumes a large amount of computational resources and costs. Therefore, in practice, an appropriate simulated backtracking can be determined based on efficiency and computational performance. Now, refer to FIG. 6. FIG. 6 shows a flowchart of steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. Each step of the method in FIG. 6 can be performed by different units of the system in FIG. 2. Therefore, in the following, each step of the method in FIG. 6 will be described using the system in FIG. 2. As shown in FIG. 6, this specific embodiment is different from the above specific embodiment in the following points. That is, the method of this specific embodiment further includes the following:

[0050] Step S342: The central server 22 generates a first contribution model by executing an aggregation algorithm on the other trained model parameters excluding the first trained model parameters in the last machine learning training.

[0051] In this specific embodiment, the central server 22 can obtain the first contribution model by simply running the aggregation algorithm after excluding the first trained model parameters from all the trained model parameters received last. In this case, the federated members do not need to go back in time in a simulation manner or perform an extra machine learning process. That is, in the method of this specific embodiment, the central server 22 only needs to run the aggregation algorithm one more time, and the computational resources and costs consumed are extremely small. However, since each federated member still uses the parameters of the previous federated model as a base in each machine learning stage, the contribution calculated by excluding the first trained model parameters only at the end is somewhat dissociated from the actual contribution.

[0052] Also refer to FIG. 7. FIG. 7 shows a flowchart of the steps of a method for calculating contribution in associative learning according to another specific embodiment of the present invention. As shown in FIG. 7, this specific embodiment is different from the above specific embodiment in the following points. That is, the method of this specific embodiment further includes:

[0053] Step S344: The central server 22 executes an aggregation algorithm on other trained model parameters, excluding the first trained model parameter in the kth machine learning training from the last, in a simulated manner, to generate a first contribution model, where k is an integer between 1 and the number of times the first trained model parameter has participated in the aggregation algorithm of the federated model.

[0054] Therefore, the first contribution model can be generated by executing the aggregation algorithm for other trained model parameters excluding the first trained model parameters from the last machine learning stage to the kth machine learning stage in a simulated manner. That is, the simulation is performed by counting down from the last machine learning and aggregation algorithm stage to the kth machine learning and aggregation algorithm stage. Then, the remaining federation members other than the federation member that generated the first trained model parameters and the central server 22 continue to complete the remaining k machine learning and aggregation algorithms, thereby finally generating a first contribution model excluding the first trained model parameters in the last k machine learning stages. k can be defined by oneself as necessary. In detail, the closer the value of k is to the total number of machine learning stages of the first trained model parameters, the closer the contribution obtained by calculation is to the actual contribution, but the more computational resources and costs are consumed. Conversely, the closer the value of k is to 1, the fewer computational resources and costs are consumed, but the more the contribution obtained by calculation deviates from the actual contribution. It should be noted that the other steps in the method of this specific embodiment are substantially the same as the corresponding steps in the above specific embodiment, and therefore will not be described in detail again here.

[0055] In the above-mentioned specific embodiment, various methods for simulating the exclusion of the first trained model parameters are presented, which may be selected according to actual needs. In short, the contribution obtained by the method of excluding the first trained model parameters from the initial machine learning stage is closer to the actual contribution, but the computational resources and costs consumed are high. However, when calculating the contribution, it is also possible to reduce the consumption of computational resources and costs while approaching the actual contribution by averaging or weighted averaging. Here, reference is made to FIG. 8. FIG. 8 shows a flowchart of the steps of the method for calculating the contribution in federated learning according to another specific embodiment of the present invention. Each step of the method in FIG. 8 can be performed by different units of the system in FIG. 2. Therefore, each step of the method in FIG. 8 will be described below using the system in FIG. 2. As shown in FIG. 8, the method in this specific embodiment includes the following:

[0056] Step S30: The central server 22 receives a number of trained model parameters, which are generated by performing machine learning multiple times (e.g., r times) on a number of local datasets, respectively.

[0057] Step S320: The central server 22 generates and stores multiple federated models by executing an aggregation algorithm on all trained model parameters in each machine learning training.

[0058] Step S346: In each machine learning training, the central server 22 excludes each first trained model parameter, and then runs an aggregation algorithm on the remaining trained model parameters to generate multiple first contribution models.

[0059] Step S360: The central server 22 compares the value of the first contribution model generated after each round of machine learning aggregation with the value of the federated model generated by each round of aggregation, and generates contributions of multiple first trained model parameters.

[0060] Step S362: The contributions of all the first trained model parameters are averaged or weighted averaged to obtain an average contribution, which is the contribution of the first trained model parameter.

[0061] In this specific embodiment, the central server 22 accepts the federated model generated by executing the aggregation algorithm after each machine learning stage, and not only redistributes it to the federated members as the basis for the parameters of the next machine learning, but also stores it in the central server 22. In addition, after accepting the trained model parameters transmitted from all the federated members in each machine learning stage, the central server 22 can directly exclude the first trained model parameter in that stage and execute the aggregation algorithm for the remaining trained model parameters to generate the first contribution model in that machine learning stage. Therefore, the contribution of the first trained model parameter in each machine learning stage is calculated by executing the aggregation algorithm twice, and the average contribution can be obtained by averaging or weighting all the calculated contributions. This average contribution is close to the actual contribution because it takes into account the contribution of the first trained model parameter in each machine learning stage to the federated model. In addition, in this specific embodiment, all the federated members do not need to go back and do machine learning again, and the central server 22 only needs to execute the aggregation algorithm one more time in each machine learning stage. Therefore, the computational resources and costs consumed are extremely low. In addition, there is no need to simulate a situation in which the first trained model parameters are not included in each machine learning training by going back in time in a simulation.

[0062] In practice, the step of calculating the contribution of the federated members described above can be achieved by the algorithm of the following specific embodiment. According to the specific embodiment, the method of calculating the contribution in federated learning can obtain a federated model (global model) by further optimizing the parameters of the trained model (local model) after each machine learning by the FedAvg algorithm. In this specific embodiment, the algorithm of calculating the contribution of the federated members has the following steps:

[0063] 1. Define a set N formed by many participants, and obtain the value of the global model of set N by performing r calculations.

[0064]

number

[0065] It is called.

[0066] 2. Exclude participants in set M from participants in set N, and calculate the value of the model obtained by rk times.

[0067]

number

[0068] Set it as follows.

[0069] 3. For all i ∈ N,

[0070]

number

[0071] where i represents each participant in N. This allows us to calculate the marginal contribution of each participant to the global model (i.e.,

[0072]

number

[0073] ) can be calculated.

[0074] It should be noted that in this specific embodiment, the calculation is performed using FedAvg, but in practice, modifications related to FedAvg, such as FedOpt, FedProx, and SCAFFOLD, may also be applied to the present invention.

[0075] In this specific example,

[0076]

number

[0077] is the value of the model obtained by calculating for set N, so it represents the case where all participants participate in federated learning.

[0078]

number

[0079] represents the case where all participants jointly participate in rk rounds of associative learning calculations, then participants in set M are removed and the remaining participants complete the final k rounds of associative learning calculations.

[0080]

number

[0081] represents the change in the value of the model when participant i is absent, and can be defined as the marginal contribution of participant i. In the above calculation, the marginal contribution in the entire process (when the participant does not participate in any of the r federated learnings) is estimated when one participant or a set of participants (|M| participants) is excluded and participates in k out of r federated learnings. The value of k can be determined as needed. In general, the larger the value of k is or the closer it is to r, the closer the calculated marginal contribution will be to the result after the actual exclusion, but the calculation cost will also be higher accordingly.

[0082] In this specific example, the marginal contribution of participant i

[0083]

number

[0084] represents the case where only the last one of r associative learnings is excluded. In other words, represents the case where k is 1. If the calculation performance allows, k may be freely adjusted and may be set to a maximum of r. In other words, participant i may be excluded from the entire process. Therefore, the marginal contribution of participant i calculated by excluding k times is

[0085]

number

[0086] In addition, participant i is one in set N. That is, the above algorithm for the marginal contribution considers the case where only one participant is excluded. If it is considered that all participants in set M are excluded at the same time, the marginal contribution is

[0087]

number

[0088] It becomes.

[0089]

number

[0090] The calculation of FedAvg is performed by going back to the rkth FedAvg calculation and then only NM participants perform the subsequent k FedAvg calculations. Therefore, the larger the k value, the higher the cost of calculation.

[0091] In addition to the method of calculating the final marginal contribution for one participant or multiple participants by excluding the participation of k times as described above, the value of a model excluding the participant or multiple participants may be calculated in each round of associative learning (a total of r times), and the values ​​of the models calculated in each round of associative learning may be averaged or weighted averaged. The value of a model excluding the participant or multiple participants in each round of associative learning may be calculated as follows:

[0092]

number

[0093] Then, the average value of the model excluding the participant or the multiple participants in r times of associative learning is

[0094]

number

[0095] Or, more generally,

[0096]

number

[0097] Here, wj is a weighting coefficient, for example,

[0098]

number

[0099] or other specific arrangements such as increasing, decreasing, partially setting to 0, etc. By defining it in this way, it becomes possible to grasp the contribution of each participant from different angles in a more holistic way.

[0100] In summary, the method for calculating the contribution in federated learning and the system for calculating the contribution and profit distribution in federated learning of the present invention can calculate the contribution to the federated model for the members participating in the federated learning. In detail, the trained model parameters trained by the members are excluded, and the other trained model parameters after the exclusion are aggregated to form a contribution model. Then, by comparing the value of the contribution model with that of the federated model, the contribution of the member and its trained model parameters can be estimated. In addition, based on the contribution of the member calculated by the system and method of the present invention, the profit distribution ratio of the member can be determined. Alternatively, the weight of the trained model parameters of the member in the federated model can be adjusted based on the contribution, thereby making the federated model more accurate. Furthermore, if the trained model parameters of the member have a negative effect on the federated model, it is also possible to grasp and process it from the contribution calculated by the method and system of the present invention. For example, if the contribution of the trained model parameters of the member to the federated model is calculated to be a negative value, the method and system of the present invention can also eliminate the negative effect of the member by excluding the member and the trained model parameters and regenerating the federated model.

[0101] The detailed description of the preferred specific embodiments above is intended to more clearly describe the features and spirit of the present invention, and is not intended to limit the scope of the present invention to the preferred specific embodiments disclosed above. Rather, it is intended that various modifications and equivalent configurations be included within the scope of the claims that are to be claimed in the present invention. Therefore, the scope of the claims that are to be claimed in the present invention should be interpreted in the broadest possible manner based on the above description so as to include all possible modifications and equivalent configurations. [Explanation of symbols]

[0102] 2. Contribution calculation and profit sharing system in associative learning 20. Artificial Intelligence Model Training Device 22 Central Server 220 Calculation Module 222 Contribution Calculation Module 224 Profit Sharing Module S10~S16,S17,S180~S182,S30~S36,S340~S346,S360~S362 Process steps

Claims

1. a central server receiving a plurality of trained model parameters, the trained model parameters being generated by machine learning a plurality of local data sets respectively; the central server generating a federated model by running an aggregation algorithm on the trained model parameters and storing the federated model; generating a first contribution model by running the aggregation algorithm on the trained model parameters excluding or simulating excluding at least one first trained model parameter; A method for calculating contribution in federated learning, comprising the step of the central server comparing the value of the first contribution model with the value of the federated model to generate a contribution of the first trained model parameter.

2. 2. The method of claim 1 , wherein the trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, the central server generates the federated model by running an aggregation algorithm on the trained model parameters generated in each round of machine learning training, and the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters excluding the at least one first trained model parameter in a last machine learning training.

3. 2. The method of claim 1, wherein the trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, the central server generates the federated model by running the aggregation algorithm in each round of machine learning training, and the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters that have been simulated to exclude the at least one first trained model parameter in each round of machine learning training.

4. The method of claim 1, wherein the trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, the central server generates the federated model corresponding to each round of machine learning by executing the aggregation algorithm in each round of machine learning training, the central server generates the first contribution model corresponding to each round of machine learning by executing the aggregation algorithm on the trained model parameters excluding the at least one first trained model parameter in each round of machine learning training, and the central server compares the value of the first contribution model with the value of the federated model in each round of machine learning and performs an average or weighted average to generate an average contribution as the contribution of the first trained model parameter.

5. The method of claim 1, wherein the trained model parameters are generated by performing multiple rounds of machine learning training on the local dataset, the central server generates the federated model by running the aggregation algorithm in each round of machine learning training, and the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters that have been simulated to exclude the at least one first trained model parameter in the machine learning training up to the kth round, counting down from the last, where k is an integer between 1 and the number of times the first trained model parameter has participated in the aggregation algorithm of the federated model.

6. the central server determining whether the contribution of the first trained model parameter is a positive value or a negative value; The method of claim 1 , further comprising: if the contribution is a negative value, the central server replaces the federated model with the first contribution model as the updated federated model.

7. A plurality of artificial intelligence model training devices that generate a plurality of trained model parameters by performing machine learning on a plurality of local data sets, respectively; a central server connected to the artificial intelligence model training device for accepting the trained model parameters; The central server comprises: a computation module that generates a federated model by running an aggregation algorithm on the trained model parameters; a contribution calculation module connected to the calculation module, for generating a first contribution model by executing the aggregation algorithm on the trained model parameters excluding or simulating excluding at least one first trained model parameter, and for comparing the value of the first contribution model with the value of the federated model to generate a contribution of the first trained model parameter; A contribution calculation and profit sharing system in federated learning, comprising: a profit sharing module connected to the contribution calculation module to receive the contribution and calculate a profit sharing ratio of the first trained model parameter in the federated model based on the contribution.

8. The system of claim 7, wherein the artificial intelligence model training device performs multiple rounds of machine learning training on the local dataset, the calculation module generates the federated model by running an aggregation algorithm in each round of machine learning training, and the contribution calculation module generates the first contribution model by running the aggregation algorithm on the trained model parameters excluding the at least one first trained model parameter in the last machine learning training.

9. The system of claim 7, wherein the artificial intelligence model training device performs multiple rounds of machine learning training on the local dataset, the calculation module generates the federated model by executing an aggregation algorithm in each round of machine learning training, and the contribution calculation module generates the first contribution model by executing the aggregation algorithm on the trained model parameters that have been simulated to exclude at least one first trained model parameter in each round of machine learning training.

10. The system of claim 7, wherein the artificial intelligence model training device performs multiple rounds of machine learning training on the local dataset, the calculation module generates the federated model by executing an aggregation algorithm in each round of machine learning training, and the contribution calculation module generates the first contribution model by executing the aggregation algorithm on the trained model parameters that have been simulated to exclude at least one first trained model parameter in the machine learning training up to the kth round, counting down from the last, where k is an integer between 1 and the number of times the first trained model parameter has participated in the aggregation algorithm of the federated model.

11. The artificial intelligence model training device performs multiple rounds of machine learning training on the local data set, and the calculation module executes the aggregation algorithm in each round of machine learning training to generate the federated model corresponding to each round of machine learning, and the contribution calculation module executes the aggregation algorithm on the trained model parameters excluding the at least one first trained model parameter in each round of machine learning training to generate the first contribution model corresponding to each round of machine learning, and compares the value of the first contribution model in each round of machine learning with the value of the federated model, and performs an average or weighted average to generate an average contribution as the contribution of the first trained model parameter.

Images