System and method for optimized client selection in federated learning

Abstract

The present invention relates to a system and method for selecting optimal clients in federated learning. The present invention relates to a technique that enables faster convergence and improved global model accuracy by selecting an optimal set of clients based on clients' statistical resources and system resources, thereby ensuring that the most capable clients participate in each learning round.

Description

Technology Field

[0001] The present invention relates to a system and method for selecting an optimal client in federated learning, and more specifically, to a system and method for selecting an optimal client in federated learning that can optimize participating clients in each training round of federated learning using a genetic algorithm (GA). Background Technology

[0002] Federated Learning (FL) is a machine learning method in which local models are trained individually on multiple distributed clients (devices or servers), and the parameters (weights) of these trained models are transmitted to a central server to be combined. In this process, since data from each client remains local and is not transmitted to a central server, it is emerging as a promising paradigm capable of training machine learning models on distributed data from multiple clients while protecting data privacy. In other words, Federated Learning has brought innovation to distributed machine learning by enabling multiple distributed clients to collaborate and train a global model without sharing raw data, thereby preserving personal information; it is highly useful in environments where security and privacy are critical.

[0003] However, such federated learning faces ongoing problems due to statistical heterogeneity, where the size and distribution of datasets differ for each client, and system heterogeneity, caused by differences in computing and communication capacities between clients. These factors hinder the efficient training of federated learning, cause synchronization delays, and degrade the performance of global models.

[0004] Conventionally, because clients are selected randomly, the aforementioned problem is not effectively solved, resulting in a slow convergence speed and reduced scalability of the federated learning system.

[0005] More specifically, conventional random client selection methods aim to address specific aspects of heterogeneity; some prioritize statistical utility centered on data quality or training loss, while others emphasize system constraints such as computing power or communication bandwidth. In other words, these conventional approaches fail to achieve optimal performance in federated learning environments because they do not simultaneously consider statistical heterogeneity and system heterogeneity. This underscores the need for a robust client selection mechanism capable of balancing efficiency and accuracy while accommodating diverse client characteristics.

[0006] In this regard, Chinese patent document (CN 113191484 B) proposes a deep reinforcement learning-based method for client selection in federated learning and discloses a technology that can improve the performance of federated learning by selecting an optimal group of clients by considering the data quality, size, and price of the clients. Prior art literature

[0007] (Patent Document 0001) CN 113191484 B(2022.10.14.) The problem to be solved

[0008] Accordingly, the present invention has been devised to solve the problems of the prior art as described above. The objective of the present invention is to provide a system and method for selecting an optimal client in federated learning that can improve federated learning efficiency by selecting an optimal client using a genetic algorithm. means of solving the problem

[0009] A system for selecting an optimal client in federated learning according to the present invention for achieving the above-mentioned purpose comprises: a central server that receives resource metric information from each of a plurality of linked clients at a predetermined interval, applies the received resource metric information to a stored artificial intelligence algorithm to select an optimal set of clients, and distributes weight values ​​based on a global model stored in the selected optimal set of clients; and a plurality of clients that perform local learning by applying the distributed weight values ​​using locally stored data, wherein each of the clients transmits an updated weight value according to the learning processing result to the central server, and the central server preferably performs an update of the global model using the received updated weight values.

[0010] Furthermore, the central server uses GA (Genetic Algorithm), calculates the fitness for each client using resource metric information received from each of the multiple clients, and selects at least two sets of clients corresponding to the parent generation based on the calculated fitness.

[0011] Furthermore, it is desirable for the central server to randomly cross clients included in each of the selected at least two client sets to generate at least two new client sets, and to set the newly generated client sets as child generations.

[0012] Furthermore, it is desirable for the central server to perform a mutation operation to change the set of clients of the child generation by replacing a randomly selected client among the clients included in each of the newly generated client sets.

[0013] Furthermore, it is desirable for the central server to combine the set of clients of the parent generation and the set of clients of the child generation to select the optimal set of clients.

[0014] Another invention for achieving the above-mentioned purpose is a method for selecting an optimal client in federated learning, wherein each step is performed by a computational processing means, and preferably comprises: an input step (S100) in which resource metric information is received from each of a plurality of linked clients at a predetermined period at a central server; an optimal selection step (S200) in which an optimal set of clients is selected based on the received resource metric information using a Genetic Algorithm (GA) at the central server; a weight distribution step (S300) in which weight values ​​based on a global model stored in the optimal set of clients are distributed at the central server; a learning processing step (S400) in which local learning is performed by applying the distributed weight values ​​using local data stored in each of the plurality of clients; a transmission step (S500) in which an updated weight value according to the learning processing result is transmitted from each client to the central server; and an update step (S600) in which an update of the global model is performed using the received updated weight values ​​at the central server.

[0015] Furthermore, the optimal selection step (S200) preferably includes a parent selection step (S210) that calculates the fitness for each client using resource indicator information received from each of the multiple clients and selects at least two client sets corresponding to the parent generation based on the calculated fitness; a crossover step (S220) that randomly crosses clients included in each of the at least two client sets selected by the parent selection step (S210) to generate at least two new client sets and sets the newly generated client sets as the child generation; a mutation step (S230) that performs a mutation operation to change the client set of the child generation by replacing a randomly selected client among the clients included in each of the client sets newly generated by the crossover step (S220); and a set update step (S240) that combines the client set of the parent generation from the parent selection step (S210) and the client set of the child generation from the mutation step (S230) to select the optimal client set.

[0016] Furthermore, it is preferable that the method for selecting the optimal client in the federated learning be performed repeatedly during a predefined federated learning round or until the global model updated by the update step (S600) reaches a desired predetermined performance. Effects of the invention

[0017] According to the present invention, a system and method for selecting optimal clients in federated learning, unlike conventional techniques that randomly select clients, select an optimal set of clients based on the statistical resources of the clients and system resources, thereby allowing the most capable clients to participate in each learning round, and thus achieving faster convergence and improved global model accuracy.

[0018] In particular, by selecting clients with sufficient computing power and data resources, the system can exclude low-performance clients that may slow down training speed or clients that do not contribute effectively, thereby reducing computing overhead, minimizing communication costs, and improving the efficiency of the federated learning process.

[0019] Furthermore, through a genetic algorithm-based selection mechanism, it is possible to dynamically adapt to client performance trends and balance exploration and utilization, thereby preventing underperforming clients from degrading the global model.

[0020] In actual federated learning deployments, client availability is unpredictable due to resource variability and connection constraints. However, the present invention allows for the dynamic selection of clients based on their current resource status, thereby effectively handling resource variability and offering the advantage of being suitable for large-scale heterogeneous federated learning environments.

[0021] In addition, conventional federated learning tends to overuse specific high-performance clients, resulting in an uneven distribution of workload; however, the present invention diversifies the selection process to prevent resource exhaustion in frequently selected clients and has the advantage of ensuring fairer participation while maintaining high performance.

[0022] Furthermore, as federated learning networks grow, client selection management becomes increasingly complex. The present invention is designed to enable efficient scalability by utilizing the evolutionary optimization of genetic algorithms, and has the advantage of maintaining robust performance across various data distributions while handling a large number of clients.

[0023] In addition, since the present invention relies solely on resource metrics for selection and does not access raw data, it can optimize client participation while maintaining data privacy. This offers the advantage of providing a privacy-conscious solution for federated learning that complies with strict data protection regulations such as GDPR. Brief explanation of the drawing

[0024] FIG. 1 is an example diagram of operation showing a system for selecting an optimal client in federated learning according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating a method for selecting an optimal client in federated learning according to one embodiment of the present invention. Specific details for implementing the invention

[0025] Hereinafter, a system and method for selecting an optimal client in federated learning according to the present invention, having the configuration as described above, will be explained in detail with reference to the attached drawings. The drawings presented below are provided as examples to ensure that the concept of the present invention is sufficiently conveyed to those skilled in the art. Accordingly, the present invention is not limited to the drawings presented below and may be embodied in other forms. In addition, throughout the specification, the same reference numerals indicate the same components.

[0026] Unless otherwise defined, technical and scientific terms used herein have the meaning commonly understood by those skilled in the art to which this invention pertains, and descriptions of known functions and configurations that could unnecessarily obscure the essence of the invention are omitted in the following description and accompanying drawings.

[0027] Furthermore, a system refers to a set of components, including devices, mechanisms, and means, that are organized and interact regularly to perform necessary functions.

[0028] A system and method for optimal client selection in federated learning according to one embodiment of the present invention introduces a client optimal selection algorithm called FedCSGA (Federated Client Selection using Genetic Algorithm) to overcome the limitations of the aforementioned prior art. Through this, optimal client selection based on resource indicators of each client is guaranteed while maintaining data privacy, and the efficiency of federated learning can be improved by enhancing the convergence of the global model.

[0029] That is, the system and method for selecting an optimal client in federated learning according to one embodiment of the present invention can present an innovative solution to the problem of heterogeneity (statistical heterogeneity / system heterogeneity) in federated learning by providing a balanced compromise that utilizes high-quality data and minimizes delays caused by resource-constrained clients, by optimizing participating clients in each training round of federated learning using a genetic algorithm (GA). By using a genetic algorithm to encode a client subset into a chromosome and iteratively improving the fitness based on statistical utility and system constraints, FedCSGA identifies an optimal client subset that improves model performance and convergence speed in federated learning.

[0030] Broadly speaking, a system for selecting an optimal client in federated learning according to one embodiment of the present invention comprises a central server (Federated Server) and a plurality of clients.

[0031] Each client stores and processes local training data while simultaneously transmitting resource metric information (e.g., CPU, dataset size, and previous loss value) to a central server. Additionally, it performs local model training and returns model updates by transmitting updated weight values ​​to the central server after each federated training round (FL round).

[0032] The central server performs optimal client selection and global model aggregation. It coordinates the federated learning process by collecting resource metric information transmitted from each client and applying genetic algorithms to select the optimal set of clients.

[0033] The global model receives updated weight values ​​from the selected optimal client set and iteratively improves performance. The central server aggregates the models learned from each client using a weighted average and redistributes the updated models to the optimal client set, thereby ensuring efficient training between training rounds. In particular, by selecting clients based on genetic algorithms, the present invention allows only high-resource clients to participate in the training rounds, thereby improving convergence speed and overall model accuracy.

[0034] FIG. 1 is an example diagram illustrating the operation of a system for selecting an optimal client in federated learning according to an embodiment of the present invention. As shown in FIG. 1, the system for selecting an optimal client in federated learning according to an embodiment of the present invention preferably includes one central server and a plurality of clients. It is preferable that each component is configured to perform operations either individually in a plurality of computational processing means including a CPU or integrated into a single computational processing means.

[0035] Let's take a closer look at each component.

[0036] It is desirable for the central server above to receive resource metric information from each of the multiple clients at predetermined intervals. At this time, it is desirable for the resource metric information to include CPU, data set size, and previous loss value, and it is desirable for this to be metric information capable of verifying statistical heterogeneity and system heterogeneity for each client.

[0037] It is desirable for the central server to select the optimal set of clients by applying the resource indicator information received from each client to the genetic algorithm (GA), which is a pre-stored artificial intelligence algorithm.

[0038] Afterwards, it is desirable for the central server to distribute weight values ​​based on a global model that are pre-stored to the selected optimal client set.

[0039] As described above, it is desirable for the central server to select the optimal set of clients to participate in the corresponding learning round from among all connected clients for each learning round.

[0040] Table 1 below is pseudo code describing the operation algorithm of a system for selecting an optimal client in federated learning according to an embodiment of the present invention. As described in Table 1 below, when each learning round (t) begins, the central server selects an optimal set of clients (among all clients (C) You end up choosing ).

[0041]

[0042] Referring to FIG. 1, the system for selecting an optimal client in federated learning according to one embodiment of the present invention is such that, for each learning round, the central server receives resource metrics (R_es) from all clients (C).

[0043] As described in row 18 of Table 1, the above central server initializes a population consisting of a set of multiple clients configured by random selection, using the received resource metric ( ...becomes

[0044] Afterwards, the central server repeatedly improves the entity set through generations (G), as described in rows 19-35 of Table 1.

[0045] More specifically, the central server uses resource indicator information received from each client configured in each client set, and each client ( )( It is desirable to calculate the fitness for ).

[0046] The above fit is calculated by the following mathematical formula 1, as indicated in rows 37-38 of Table 1.

[0047]

[0048] (Here, , Each client ( It represents the statistical and system resources of ), , is a weighting coefficient.)

[0049] According to the above mathematical formula 1, the chromosome representing a client with higher statistical and system resources will have a better fitness value.

[0050] Afterwards, it is desirable for the central server to select at least two sets of clients (P1, P2) corresponding to the parent generation based on each calculated fitness level, as described in row 20 of Table 1.

[0051] More specifically, in the current generation, that is to say, a certain proportion of the client with the highest fitness among a set of multiple clients randomly selected from the entire client pool ( It is desirable to ensure that ) is selected as the parent of the next generation. Here, a predetermined ratio ( ) determines the proportion of clients selected as parents in the previous generation.

[0052] Afterwards, it is desirable for the central server to randomly cross clients included in at least two client sets (P1, P2) selected as parent generations to generate at least two new client sets, and to set the newly generated client sets (Y1, Y2) as child generations.

[0053] Specifically, as described in rows 21-28 of Table 1, a predetermined proportion of clients with the highest fitness value among at least two sets of clients (P1, P2) selected as parent generations ( Clients corresponding to a new child generation possessing ) are configured. Here, a predetermined ratio ( ) specifies the ratio of new clients generated through crossover. With reference to FIG. 1, clients included in at least two sets of clients (P1, P2), each selected as the parent generation according to a randomly selected crossover index (L), are randomly crossed to generate at least two new sets of clients, and the newly generated sets of clients (Y1, Y2) are set as the child generation. Through this, diversity is secured while preserving high-quality genetic material.

[0054] Each child generation undergoes a mutation with a low probability. To this end, the central server performs a mutation operation and changes the client set of the child generation by replacing a randomly selected client among the clients included in the newly created client sets (Y1, Y2), respectively.

[0055] Specifically, as described in rows 29-33 of Table 1, it is desirable to replace a randomly selected client among the clients included in the newly created client sets (Y1, Y2) with another client. This prevents premature convergence and facilitates the exploration of the solution space.

[0056] Subsequently, as described in row 34 of Table 1, the central server combines the set of clients of the parent generation (P1, P2) and the set of clients of the child generation (Y1, Y2) to obtain an object set. It is updated. Through this, clients with high fitness are preserved. The central server repeats this process for a predefined federated learning round or until no further significant improvement in the fitness of the updated set of objects is observed.

[0057] As described in row 36 of Table 1, the above central server comprises a set of clients consisting of the clients with the highest fitness value after the update of the aforementioned set of objects is completed for each training round ( ) is selected as the optimal client set mentioned above, and federated learning is performed to participate in the local training of the global model.

[0058] That is, the above optimal client set ( Each client corresponding to ) performs local learning by applying the weight values ​​of the distributed global model using pre-stored local data.

[0059] Subsequently, each client transmits an updated weight value to the central server based on the learning processing result, and the central server performs an update of the global model using the received updated weight values.

[0060] Specifically, as described in rows 10-16 of Table 1, the optimal client set ( Each client corresponding to ) has the weight value of the above global model ( The local dataset stored after receiving the ) Local learning is performed using ).

[0061] After this, the client updates the local model's weight values ​​( ) is transmitted back to the central server, and the central server performs an update of the global model using a weighted average of the received updated weight values. The updated local model's weight values ​​( If we define ), it is equal to the following mathematical formula 2.

[0062]

[0063] (Here, is the dataset size of client(i).

[0064] This entire process is repeated during predefined federated learning rounds or until the updated global model reaches a desired predetermined performance, and thereafter, the final global model ( It is returned as ).

[0065] FIG. 2 is a sequence example diagram illustrating a method for selecting an optimal client in federated learning according to an embodiment of the present invention. As shown in FIG. 2, the method for selecting an optimal client in federated learning according to an embodiment of the present invention includes an input step (S100), an optimal selection step (S200), a weight distribution step (S300), a learning processing step (S400), a transmission step (S500), and an update step (S600). The above steps are performed by a system for selecting an optimal client in federated learning according to the present invention, which is operated by a computation processing means.

[0066] Let's take a closer look at each step.

[0067] The above input step (S100) receives resource metric information from each of the multiple clients at a preset interval from the central server. At this time, it is desirable that the resource metric information includes CPU, data set size, and previous loss value, and it is desirable that this is metric information capable of verifying statistical heterogeneity and system heterogeneity for each client.

[0068] The optimal selection step (S200) selects an optimal set of clients by applying the resource indicator information received from each client to a genetic algorithm (GA), which is a genetic algorithm that is a pre-stored artificial intelligence algorithm, at the central server.

[0069] The optimal selection step (S200) above includes a parent selection step (S210), a crossover step (S220), a mutation step (S230), and a set update step (S240), as shown in FIG. 2.

[0070] The above parent selection step (S210) uses resource indicator information received from each client configured in each client set, and each client ( )( The fitness for ) is calculated. Based on each calculated fitness, at least two sets of clients (P1, P2) corresponding to the parent generation are selected.

[0071] More specifically, in the current generation, that is to say, a certain proportion of the client with the highest fitness among a set of multiple clients randomly selected from the entire client pool ( It is desirable to ensure that ) is selected as the parent of the next generation. Here, a predetermined ratio ( ) determines the proportion of clients selected as parents in the previous generation.

[0072] The above crossing step (S220) randomly crosses clients included in each of the at least two client sets selected by the parent selection step (S210) to create at least two new client sets and sets the newly created client sets as child generations.

[0073] In other words, the above crossing step (S220) randomly crosses the clients included in each of the at least two client sets (P1, P2) selected as parent generations to generate at least two new client sets, and sets the newly generated client sets (Y1, Y2) as child generations.

[0074] More specifically, a predetermined proportion of clients with the highest fitness value among at least two sets of clients (P1, P2) selected as parent generations ( Clients corresponding to a new child generation possessing ) are configured. Here, a predetermined ratio ( ) specifies the ratio of new clients generated through crossover. With reference to FIG. 1, clients included in at least two sets of clients (P1, P2), each selected as the parent generation according to a randomly selected crossover index (L), are randomly crossed to generate at least two new sets of clients, and the newly generated sets of clients (Y1, Y2) are set as the child generation. Through this, diversity is secured while preserving high-quality genetic material.

[0075] The mutation step (S230) performs a mutation operation to replace a randomly selected client among the clients included in the client set newly created by the crossover step (S220), thereby changing the client set of the child generation.

[0076] More specifically, it is desirable to replace a randomly selected client from among the clients included in the newly generated client sets (Y1, Y2) with another client. This prevents premature convergence and facilitates the exploration of the solution space.

[0077] The set update step (S240) combines the set of clients of the parent generation from the parent selection step (S210) and the set of clients of the child generation from the mutation step (S230) to select the optimal set of clients.

[0078] That is, by combining the set of clients of the parent generation (P1, P2) and the set of clients of the child generation (Y1, Y2), the object set It is updated. Through this, clients with high fitness are preserved. The central server repeats this process for a predefined federated learning round or until no further significant improvement in the fitness of the updated set of objects is observed.

[0079] The above central server, for each learning round, after the update of the aforementioned set of objects is completed, consists of a set of clients composed of the clients with the highest fitness value ( ) is selected as the optimal client set mentioned above, and federated learning is performed to participate in the local training of the global model.

[0080] That is, the weight distribution step (S300) above is performed at the central server, the optimal client set ( The weight values ​​of the global model, which are stored in advance, are distributed to each client corresponding to ).

[0081] The above learning processing step (S400) is performed on a plurality of clients, more specifically, the optimal client set ( In each client corresponding to ), local learning is performed by applying the weight values ​​of the distributed global model using pre-stored local data.

[0082] The transmission step (S500) transmits the updated weight values ​​from each client to the central server according to the respective learning processing results.

[0083] The above update step (S600) performs an update of the global model using the updated weight values ​​received from the central server.

[0084] More specifically, the above optimal client set ( Each client corresponding to ) has the weight value of the above global model ( The local dataset stored after receiving the ) Local learning is performed using ).

[0085] After this, the client updates the local model's weight values ​​( ) is transmitted back to the central server, and the central server performs an update of the global model using a weighted average of the received updated weight values.

[0086] The entire step of the method for selecting an optimal client in federated learning according to one embodiment of the present invention is repeated during a predefined federated learning round or until the updated global model reaches a desired predetermined performance, and thereafter, the final global model ( It is returned as ).

[0087] A system and method for selecting an optimal client in federated learning according to one embodiment of the present invention will be described with examples as follows.

[0088] Total available clients (C = {c1, c2, c3, c4, c5, c6}), number of clients participating in each training round (m = 2), number of generations of the genetic algorithm (G = 3), size of the population being initialized (P = 4) (i.e., 4 sets of clients are initialized, each set containing 2 individual clients), parent chromosomes fraction ( = 0.5), child chromosomes fraction ( When limited to = 0.5), as indicated in row 15 of Table 1, the initial object set (Initial Population ) consists of a set of 4 clients including 2 clients, and this It is initialized as follows. Each client set ( ) includes 2 clients, and this , , , It can be defined as ).

[0089] Each client is evaluated based on its fitness, and the fitness is calculated as shown in Equation 1 above.

[0090] As indicated in row 17 of Table 1, subsequently, the chromosome ratio of the parental generation ( Following this, clients corresponding to the top 50% of fitness from the initial object set are selected as the parent generation. The two client sets with the highest fitness , If you say that, It becomes included in the next generation's set of objects.

[0091] As indicated in rows 18–25 of Table 1, subsequently, the chromosome ratio of the offspring generation ( ), Accordingly, cross-operations are performed.

[0092] A crossover point (L = 1) is randomly selected, dividing the chromosomes (containing clients) of the two parent client sets in half and performing a crossover. That is, = P1 = {c2, c5}, In P2 = {c1, c6}, performing the crossover operation generates Y1 = {c2, c6} and Y2 = {c1, c5}. Accordingly, the new child generation client set is It corresponds to.

[0093] As indicated in rows 26-30 of Table 1, a mutation operation is subsequently performed, and in this embodiment, a mutation operation is performed to replace c2 with c3 in Y1. At this time, the mutation probability is It is desirable, and as a result of the mutation operation, the client set of the child generation changes to Y1 = {c3, c6} and Y2 = {c1, c5}.

[0094] As indicated in row 31 of Table 1, subsequently, a new set of objects ( ) updates by combining the set of clients from the parent generation and the set of clients from the child generation.

[0095] in other words, is, and as a result, It corresponds to.

[0096] As indicated in row 36 of Table 1, thereafter, for each training round, after the update of the aforementioned object set is completed, a client set consisting of the client with the highest fitness value ( ) is selected as the optimal client set mentioned above, and federated learning is performed to participate in the local training of the global model.

[0097] The system and method for selecting an optimal client in federated learning according to one embodiment of the present invention, when applied to the medical field, can facilitate the collaborative training of medical AI models without sharing sensitive patient data between hospitals and research institutions. By selecting a client based on resource availability, it ensures seamless model training even in heterogeneous environments where medical providers with different computational capabilities exist. This is particularly useful for disease diagnosis, medical image analysis, and personalized treatment recommendations.

[0098] Furthermore, when applied to the financial sector, it can improve fraud detection and risk assessment models by enabling institutions to collaborate and train models while maintaining data privacy in financial applications. By optimizing client selection, it ensures that only trustworthy and resource-efficient institutions contribute to training, thereby creating more accurate and robust financial models.

[0099] Furthermore, when applied to the Smart Cities & IoT sector, it can improve federated learning in smart cities. For instance, high-performance edge devices, such as surveillance cameras and traffic sensors, are selected and involved in training. This enhances real-time anomaly detection, predictive maintenance, and traffic optimization, while ensuring efficient energy consumption and network bandwidth utilization.

[0100] In addition, when applied to the field of Mobile & Edge Computing, it enables efficient federated learning in mobile applications on smartphones, wearable devices, and other edge devices. By selecting clients with sufficient computing power and network stability, it improves the training efficiency of personalized AI models for applications such as voice assistants, recommendation systems, and health monitoring apps.

[0101] Furthermore, when applied to the field of Telecommunications & Network Optimization, telecommunications service providers can optimize network performance monitoring and user experience prediction models. By dynamically selecting clients based on resource availability, it ensures that only the most suitable network nodes participate in training, thereby reducing communication costs and improving model accuracy.

[0102] As such, the system and method for optimal client selection in federated learning according to one embodiment of the present invention improve federated learning efficiency through resource-aware client selection, thereby solving major industrial problems and providing commercially valuable innovations.

[0103] Meanwhile, a system and method for selecting an optimal client in federated learning according to one embodiment of the present invention may be implemented in the form of program instructions that can be executed through various means of electronically processing information and recorded on a storage medium. The storage medium may include program instructions, data files, data structures, etc., either individually or in combination.

[0104] Program instructions recorded on a storage medium may be those specifically designed and configured for the present invention, or they may be those known and available to those skilled in the art of software. Examples of storage media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a device that processes information electronically using an interpreter, such as a computer.

[0105] As described above, the present invention has been explained with specific details such as specific constituent elements and limited exemplary drawings; however, this is provided merely to aid in a more comprehensive understanding of the invention, and the invention is not limited to the above-mentioned exemplary embodiment. Those skilled in the art can make various modifications and variations from this description.

[0106] Accordingly, the scope of the present invention should not be limited to the described embodiments, and all things equivalent to or having equivalent variations to the claims set forth below, as well as the claims set forth below, shall be considered to fall within the scope of the concept of the present invention.

Claims

Claim 1 A system for selecting optimal clients in federated learning, comprising: a central server that receives resource metric information from each of a plurality of linked clients at predetermined intervals, applies the received resource metric information to a stored artificial intelligence algorithm to select an optimal set of clients, and distributes weight values ​​based on a stored global model to the selected optimal set of clients; and a plurality of clients that perform local learning by applying the distributed weight values ​​using stored local data, wherein each of the clients transmits an updated weight value according to the learning processing result to the central server, and the central server performs an update of the global model using the received updated weight values. Claim 2 A system for selecting an optimal client in federated learning, wherein the central server uses a Genetic Algorithm (GA), calculates a fitness for each client using resource indicator information received from each of a plurality of clients, and selects at least two sets of clients corresponding to a parent generation based on the calculated fitness. Claim 3 A system for selecting optimal clients in federated learning, wherein the central server randomly crosses clients included in each of at least two selected client sets to generate at least two new client sets and sets the newly generated client sets as child generations. Claim 4 A system for optimal client selection in federated learning, wherein the central server performs a mutation operation to replace a randomly selected client among the clients included in each of the newly generated client sets, thereby changing the child generation client set. Claim 5 In claim 4, the central server combines the set of clients of the parent generation and the set of clients of the child generation to select the optimal set of clients, a system for optimal client selection in federated learning. Claim 6 A method for selecting an optimal client in federated learning, wherein each step is performed by a computational processing means, comprising: an input step (S100) in which a central server receives resource metric information from each of a plurality of linked clients at a predetermined interval; an optimal selection step (S200) in which a central server selects an optimal set of clients based on the received resource metric information using a Genetic Algorithm (GA); a weight distribution step (S300) in which a central server distributes weight values ​​based on a global model stored in the optimal set of clients; a learning processing step (S400) in which a plurality of clients perform local learning by applying the distributed weight values ​​using local data stored in each of the clients; a transmission step (S500) in which each client transmits an updated weight value according to the learning processing result to the central server; and an update step (S600) in which a central server performs an update of the global model using the received updated weight values. Claim 7 In claim 6, the optimal selection step (S200) calculates the fitness for each client using resource indicator information received from each of a plurality of clients, and selects at least two client sets corresponding to the parent generation based on each calculated fitness; the parent selection step (S210) randomly crosses the clients included in each of the at least two client sets selected by the parent selection step (S210) to generate at least two new client sets and sets the newly generated client sets as the child generation; the mutation step (S230) performs a mutation operation to change the client set of the child generation by replacing a randomly selected client among the clients included in each of the client sets newly generated by the mutation step (S220); and the set update step (S240) combines the client set of the parent generation from the parent selection step (S210) and the client set of the child generation from the mutation step (S230) to select the optimal client set. Claim 8 In claim 7, the method for selecting an optimal client in federated learning is performed repeatedly during a predetermined federated learning round or until the global model updated by the update step (S600) reaches a desired predetermined performance.

Images