Federated learning method and system with dynamic personalization and generalization

Abstract

One embodiment of the present invention provides a federated learning system that dynamically reflects the degree of personalization and generalization, comprising a plurality of clients and servers, wherein the client model provided by each client comprises: a body based on an artificial neural network that extracts features from input data; a bridge based on an artificial neural network that converts the features extracted from the body into personalized features according to the characteristics of the client; a head based on an artificial neural network that performs classification based on the features converted by the bridge; and a control unit that controls the neural network learning of the body, bridge, and head.

Description

Technology Field

[0001] The present invention relates to the field of machine learning, and more specifically, to a federated learning method and system in which multiple clients cooperate to perform learning without sharing data. Background Technology

[0003] In general, the quantity and quality of training data are critical factors that determine model performance in the field of machine learning. Early machine learning began with a centralized approach where all training data was collected and processed on a single central server. Subsequently, as the scale of data grew and the need for distributed processing emerged, distributed learning methods were introduced, which involve distributing data across multiple servers for parallel processing.

[0004] Meanwhile, as awareness of personal information protection has increased, various legal frameworks have been established. The European Union's General Data Protection Regulation (GDPR), effective since 2018, applies strict regulations to the collection, processing, and transfer of personal data, and the U.S. California Consumer Privacy Act (CCPA) protects consumers' data rights in a similar context. Korea also protects data privacy through the Personal Information Protection Act.

[0005] Against this backdrop, federated learning was proposed as a method that enables collaborative learning while protecting data privacy. In federated learning, each client trains a model locally without sharing its data externally. The entire training process proceeds by a central server aggregating only the model parameters from each client to create a global model, which is then distributed back to the clients.

[0006] Federated learning is being actively researched, particularly in fields where data security is critical, such as healthcare, finance, and autonomous driving. In the medical field, it enables the training of disease diagnosis models without directly sharing patient data from multiple hospitals; in the financial sector, it allows for the development of financial service models while protecting customers' financial information. In the field of autonomous driving, it enables the training of safe driving models by utilizing driving data collected from various vehicles.

[0007] The related market was valued at approximately $114.8 million in 2022 and is expected to grow at an average annual rate of 10.6% until 2030, driven by the proliferation of the Internet of Things (IoT) and increasing demand for edge computing. In particular, the practical application of federated learning technology is expanding, centered around medical AI companies and autonomous driving companies. The problem to be solved

[0009] The technical problem that the present invention aims to solve is to provide a federated learning method and system capable of effectively learning generalized patterns from data of all clients while reflecting the characteristics of each client, even in non-independent and non-identical distribution (Non-IID) data environments and environments with a limited number of data.

[0010] The technical problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art to which the present invention belongs from the description below. means of solving the problem

[0012] To achieve the above technical objectives, one embodiment of the present invention provides a federated learning system that dynamically reflects the degree of personalization and generalization, comprising a plurality of clients and servers, wherein the client model provided by each client comprises: a body based on an artificial neural network that extracts features from input data; a bridge based on an artificial neural network that converts the features extracted from the body into personalized features according to the characteristics of the client; a head based on an artificial neural network that performs classification based on the features converted by the bridge; and a control unit that controls the neural network learning of the body, bridge, and head.

[0013] In an embodiment of the present invention, the control unit fixes the parameters of the body and the bridge when training the head, generates a global prototype from features extracted by the body, generates a local prototype from features transformed by the bridge, and stores each in the memory of the client; when training the bridge, fixes the parameters of the body and the head, trains the bridge based on a loss function between the local prototype stored in the memory and the current bridge output, generates a new local prototype from the trained bridge and stores it in the memory; when training the body, fixes the parameters of the bridge and the head, calculates a global-local similarity weight based on the similarity between the global prototype and the local prototype, applies the global-local similarity weight to a loss function to adjust the degree of reflection of the corresponding client's characteristics, stores the global-local similarity weight in the memory and uses it for training the bridge in the next round, and the server integrates and generalizes the parameters of the body collected from the plurality of clients, and the generalized parameters It can be distributed to the above multiple clients.

[0014] In an embodiment of the present invention, the control unit activates a personalization early termination indicator when the rate of change of the body's parameters decreases in a preset number of consecutive rounds, and when the personalization early termination indicator is activated, the bridge learning can be omitted in subsequent rounds and the application of the global-local similarity weighting can be excluded during the body learning.

[0015] In an embodiment of the present invention, the control unit calculates the mean squared error for each class for the global prototype and the local prototype to calculate the global-local similarity weight, and generates the global-local similarity weight for each class by multiplying the mean squared error for each class by a coefficient reflecting dataset characteristics, wherein the coefficient reflecting dataset characteristics may be determined based on the number of classes of the dataset and the distribution characteristics between classes.

[0016] To achieve the above technical objective, another embodiment of the present invention provides a federated learning method that dynamically reflects the degree of personalization and generalization between a plurality of clients and servers, wherein each client model is configured to extract features from input data through an artificial neural network-based body, convert the features into personalized features according to the characteristics of the client through an artificial neural network-based bridge, and perform classification based on the personalized features through an artificial neural network-based head, wherein the head performs classification based on the features converted by the bridge, and a control unit trains the head; a feature conversion step in which the bridge converts features extracted from the body into personalized features according to the characteristics of the client, and the control unit trains the bridge; and a feature extraction step in which the body extracts features from input data and the control unit trains the body. Effects of the invention

[0018] According to an embodiment of the present invention, the following effects can be obtained.

[0019] First, the balance between personalization and generalization can be effectively achieved through bridge layers and step-by-step learning.

[0020] Second, the characteristics of each client can be dynamically reflected through global-local similarity weights and personalization early termination metrics.

[0021] Third, through the dynamic adjustment mechanism of the P-Stop Indicator and GL-Sim Weight, the risk of overfitting is reduced by optimizing the degree of personalization, and the model accuracy performance of users with extremely different data distributions is significantly improved.

[0022] Fourth, personalized AI models can be effectively trained in real industrial sectors where data security is important, such as healthcare, finance, and autonomous driving.

[0023] The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the configuration of the invention described in the detailed description of the invention or the claims. Brief explanation of the drawing

[0025] FIG. 1 is a block diagram showing the overall configuration of a federated learning system that dynamically reflects the degree of personalization and generalization according to one embodiment of the present invention. Figure 2 is a diagram illustrating the operation of the federated learning system of Figure 1. FIG. 3 is a flowchart illustrating a federated learning method that dynamically reflects the degree of personalization and generalization according to another embodiment of the present invention. Figure 4 is a diagram illustrating the classification stage for training the head. Figure 5 is a diagram illustrating the prototyping stage for training the bridge. Figure 6 is a diagram illustrating the feature extraction stage (Localization Stage) for training the body. Figure 7 is a table showing the results of comparing the performance of the present invention (FedSDP) and existing federated learning methods. Figure 8 is a graph showing the change in accuracy per communication round for the CIFAR10 dataset in Pathological Setting (extreme situation). Figure 9 is a graph showing the change in accuracy per communication round for the CIFAR10 dataset in a practical setting. Figure 10 is a graph showing the change in accuracy per communication round for the CIFAR100 dataset in Pathological Setting. Figure 11 is a graph showing the change in accuracy per communication round for the CIFAR100 dataset in a practical setting. Figure 12 is a table comparing the performance of the present invention (FedSDP) and existing federated learning methods according to client participation rate (ρ). Figure 13 is a table comparing the performance variance between clients of the present invention (FedSDP) and existing federated learning methods in various datasets using the coefficient of variance (CV). FIG. 14 is a table comparing the performance of the present invention according to the presence or absence of a personalized early termination indicator (P-Stop Indicator) and the number of consecutive decreases (n) that serve as the judgment criteria. Figure 15 is a table comparing the performance of the present invention according to the global-local similarity weighting (GL-Sim Weight) and fixed weighting methods. FIG. 16 is a table showing experimental results analyzing the effect of each component of the present invention on performance. Specific details for implementing the invention

[0026] The present invention is susceptible to various modifications and may take various forms; therefore, specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the invention to the specific disclosed forms, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. Similar reference numerals have been used for similar components in the description of each drawing.

[0027] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0028] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the attached drawings.

[0029] FIG. 1 is a block diagram showing the overall configuration of a federated learning system that dynamically reflects the degree of personalization and generalization according to an embodiment of the present invention. FIG. 2 is a diagram for explaining the operation of the federated learning system of FIG. 1.

[0030] Referring to FIGS. 1 and 2, a federated learning system (hereinafter referred to as the federated learning system) (10) that dynamically reflects the degree of personalization and generalization of the present embodiment includes a server (100) and a plurality of clients (200). Each client (200) may include a neural network with a three-stage structure that sequentially extracts features from input data and transforms them to perform final classification, a control unit (250) that controls the same, and a memory (240) that stores a prototype.

[0031] The client (200) is a neural network structure and includes a body (210), a bridge (220), and a head (230). The body (210) is a feature extractor that extracts basic features from input data, the bridge (220) is a feature converter that converts the extracted features to suit the characteristics of the client, and the head (230) is a classifier that performs final classification based on the converted features.

[0032] Referring to Fig. 2, F θ (x) corresponds to a body (layer) that acts as a feature encoder. The body extracts relevant features from input data x. The representation vector z generated from the body (210) is a bridge (layer) P h (z) passes through. This layer, designed for local prototype generation, takes feature z as a local feature space. Convert to. The head layer acting as the final classifier It is entered as.

[0033] The control unit (250) controls the learning of this three-stage neural network step by step.

[0034] Head (230) training is a process focused on setting classification boundaries (criteria) of a local model (client) tailored to user data. The control unit (250) trains the head (230) (updates the head's parameters) using a cross-entropy loss function. When the control unit (250) trains the head (230) (Classification Stage), as shown in FIG. 2, the parameters of the body (210) and the bridge (220) are fixed (the body and the bridge are indicated by dotted lines), and from the features extracted by the body (210), a global prototype ( Generates a local prototype () from features transformed by the bridge (220) A prototype is generated and stored in memory (240) for the next step. i represents the client and k represents the class. In this embodiment, the generation and utilization of the prototype constitute the core of the learning process. The purpose of generating a prototype during head training is to first establish classification criteria and then optimize feature extraction and transformation based on them. The global prototype generated in the body represents general feature patterns, while the local prototype generated in the bridge represents client-specific feature patterns. These generated prototypes are stored in memory and used as reference points for learning in subsequent steps.

[0035] The training of the bridge (220) is a process focused on enhancing the personalization of the local prototype. The control unit (250) trains the bridge using an MSE loss function with Global-Local Similarity Weights (GL-Sim Weights) applied, which will be described later. When the control unit (250) trains the bridge (220) (Prototyping Stage), as shown in FIG. 2, the parameters of the body (210) and the head (230) are fixed (the body and the head are indicated by dotted lines), and the local prototype ( stored in memory (240) ) and current bridge (220) output ( The bridge (220) is trained by comparing it with ). That is, the parameters of the bridge (220) are updated. This is to ensure that the bridge (220) gradually develops while consistently maintaining the client's unique characteristics. If the current output of the bridge (220) differs significantly from the stored local prototype, strong training is performed, and if the difference is small, only fine-tuning is performed. In addition, the local prototype from the previous step (head training) is updated so that the updated local prototype ( Saves ).

[0036] The body (210) balances personalization and generalization through the learning (Localization Stage). Before learning the body (210), the control unit (250) calculates Global-Local Similarity Weights (GL-Sim Weights) based on the similarity between the global prototype and the local prototype stored in memory (240). This will be described later. For body updates, the control unit (250) uses both CE loss and MSE loss with applied Global-Local Similarity Weights, and the two loss functions are applied via different paths. First, the Cross Entropy Loss is applied via the bridge (P h ) and head(G φ It is applied to the final output that has passed through to improve class distinction ability. This is akin to evaluating and improving the accuracy of a final diagnosis in a medical system. For example, it ensures the ability to accurately distinguish between pneumonia and bronchitis. Secondly, the Mean Squared Error (MSE) Loss with global-local similarity weights is applied to the body (F θ It is used to directly compare the output of the body with the local prototype stored in memory. This evaluates how well the features generated by the body reflect the characteristics of the hospital. At this time, global-local similarity weights are applied, allowing for personalization of different intensities for each class. Through this dual-path structure, the system can simultaneously optimize diagnostic accuracy (Cross Entropy Loss) and the reflection of hospital characteristics (Weighted MSE Loss). The body is updated by taking both of these goals into account. The control unit (250) fixes the parameters of the bridge (220) and the head (230) (the bridge and the head are indicated by dotted lines) as shown in FIG. 2 and proceeds with the body update.

[0037] As mentioned above, fixing the parameters of a specific layer at each training step is a strategy for the stability and efficiency of training. If you train all layers at once The roles of each layer can become ambiguous or interfere with one another. For example, when training the head, fixing the body and bridge allows for the optimization of pure classification boundaries while maintaining feature extraction and transformation.

[0038] The global-local similarity weights can be stored in memory (240) for use in the next round of bridge learning. The global-local similarity weights serve as an important measure for determining the strength of personalization in the next round of bridge learning. A large weight indicates that the characteristics of the current client differ significantly from the general pattern, so the bridge attempts stronger personalization in the next round. Conversely, a small weight indicates that it is close to the general pattern, so only weak personalization is performed.

[0039] Subsequently, if sufficient personalization is achieved through the personalization early termination indicator, the process transitions to a generalization process driven by the server. The personalization early termination indicator will be discussed later.

[0040] The server (100) performs the role of collecting only the parameters of the body (210) from each client, integrating and generalizing them, and then distributing them to all clients. The server integrates only the parameters of the body to clearly separate personalization and generalization. Since the body is responsible for extracting basic features, sharing this knowledge allows all clients to possess robust basic capabilities. On the other hand, since the bridge and head reflect the unique characteristics of each client, integrating them can dilute individuality.

[0041] Through this cyclical process, generalized knowledge can be learned from the entire dataset while preserving the individual characteristics of clients. In other words, as these mechanisms operate organically, the system accumulates generalized knowledge as a whole while preserving the specificity of each client. In particular, by dynamically adjusting the degree of personalization, it becomes possible to find the optimal balance point for each client.

[0042] The aforementioned global-local similarity weights are further explained as follows.

[0043] The control unit (250) precisely calculates global-local similarity weights. Specifically, it calculates the mean squared error for each class for the global prototype and the local prototype, and generates global-local similarity weights for each class by multiplying this by a coefficient reflecting dataset characteristics. At this time, the coefficient reflecting dataset characteristics is determined by considering the total number of classes and the characteristics of the data distribution between classes. This is intended to achieve a more accurate personalization-generalization balance by reflecting the complexity and imbalance of the data in the weight calculation.

[0044] Global-Local Similarity Weight (GL-Sim Weight ; The mathematical definition of ) is as follows.

[0045]

[0046] Here is the global prototype for the k-th class of client i, represents a local prototype. λ is a coefficient reflecting dataset characteristics and is set differently depending on the complexity of the dataset. Specifically, λ = 2.85 is used for relatively simple datasets such as CIFAR10 or STL-10, while λ = 3.3 can be used for complex datasets such as CIFAR100, CINIC*, and TinyImageNet. Embodiments of the present invention have been verified using various benchmark datasets. Here, 'CIFAR10' refers to a dataset of 60,000 32x32 color images consisting of 10 classes, and 'CIFAR100' refers to an image dataset of the same size consisting of 100 classes. 'STL-10' is a high-resolution (96x96) color image dataset of 10 classes, and 'CINIC*' refers to a portion of the ImageNet dataset processed into a format compatible with CIFAR10. 'TinyImageNet' is a reduced version of the ImageNet dataset, consisting of 200 classes of 64x64 color images.

[0047] These datasets can be broadly divided into two groups based on their complexity. Relatively simple datasets, such as CIFAR10 or STL-10, have a small number of classes and relatively low difficulty in classification tasks, whereas complex datasets, such as CIFAR100, CINIC*, and TinyImageNet, are characterized by a large number of classes and more difficult classification.

[0048] These global-local similarity weights are calculated and applied independently for each class. For example, in a medical imaging diagnostic system, if a hospital has strong specificity regarding a particular disease such as pneumonia (i.e., significant differences between global and local prototypes), a higher global-local similarity weight is assigned to that class, resulting in stronger personalization. Conversely, for diseases like the common cold where differences between hospitals are not significant, a lower global-local similarity weight is assigned, allowing for greater utilization of generalized knowledge.

[0049] In particular, global-local similarity weights change dynamically as training progresses. Initially, they generally have high values ​​to drive strong personalization, but as training continues, they gradually decrease to strike a balance with generalization. This dynamic adjustment mechanism enables each client to automatically find the optimal level of personalization suited to its data characteristics.

[0050] In addition, the aforementioned personalization early termination indicators are further explained as follows.

[0051] The control unit (250) monitors the rate of change of parameters of the body (210) to control the degree of personalization. Specifically, if the rate of change of parameters of the body (210) decreases in a preset number of consecutive rounds (e.g., 2 times), an early personalization termination indicator is activated. When this indicator is activated, bridge (220) learning is omitted in subsequent rounds, and the application of global-local similarity weights is excluded even during body (210) learning. This is a mechanism that allows the client to focus on generalization without further personalization when it is determined that the client has reached a sufficient level of personalization.

[0052] The mathematical definition of the Personalized Early Stop Indicator (P-Stop Indicator) is as follows.

[0053]

[0054] Here, △t represents the amount of change in body parameters at the t-th round, and specifically, △ t = ||θ t - θ t-1 It is calculated as ||. θ t is the body parameter value at the t-th round. n is the number of rounds used as a criterion for determining continuous reduction, and experimental verification confirmed that optimal performance is achieved when n=2.

[0055] The reason this early personalization termination mechanism is actually effective is that the gradual decrease in changes to body parameters signifies that the client's personalization has entered a stabilization phase. A continuous decrease over two consecutive rounds can be viewed as a genuine sign of convergence rather than a coincidence. At this point, continuing personalization could lead to overfitting; therefore, bridge learning and the application of global-local similarity weights are discontinued, and the focus shifts to generalization. In particular, experiments using the CIFAR10 dataset demonstrated its effectiveness, achieving an accuracy of 82.13% (±0.32) when the early personalization termination metric was applied, compared to 80.14% (±0.07) when it was not. This shows that this mechanism plays a key role in finding the appropriate balance between personalization and generalization.

[0056] The process of applying a federated learning system that dynamically reflects the degree of personalization and generalization of this embodiment to a real-world setting, using the medical field as an example, is as follows.

[0057] When the trained system is deployed to an actual hospital, each hospital will have its own specialized diagnostic system. When a new patient's medical image is input, the body (210) first extracts basic medical features from the image. For example, it captures basic visual features such as shadow patterns, tissue texture, and singularities in an X-ray image.

[0058] Next, the bridge (220) reinterprets these basic characteristics extracted by the body to suit the expertise of the hospital. For example, in the case of a hospital specializing in the elderly, the transformation can be made to emphasize characteristics related to geriatric diseases, and in the case of a hospital specializing in children, to emphasize characteristics related to pediatric diseases. This reflects the specialized perspective of the hospital established during the learning process.

[0059] Finally, the head (230) performs a final diagnosis based on the characteristics that reflect the hospital's characteristics. The diagnosis at this stage is based on general medical knowledge, but also fully reflects the expertise of the hospital. Through this three-stage processing, the system is able to provide a medically reliable diagnosis result (generalization) while preserving the expertise of each hospital (personalization).

[0060] In particular, the federated learning system of this embodiment can be continuously developed. Through regular communication with the server, the body of each hospital receives updates that reflect the experiences of other hospitals, and by processing these through its own bridge and head, the development of overall medical knowledge and the expertise of individual hospitals are harmonized.

[0061] FIG. 3 is a flowchart illustrating a federated learning method that dynamically reflects the degree of personalization and generalization according to another embodiment of the present invention.

[0062] Referring to FIG. 3, the federated learning method of the present invention performs a classification step (S10) for training a head, a feature transformation step (S20) for training a bridge, and a feature extraction step (S30) for training a body at each client, sequentially performs a parameter generalization step (S40) through a server, and may selectively omit the bridge training step (S20) depending on whether personalization is terminated early. Information processed at each step is stored in memory, and the processing result of each step is used as input for the next step. The federated learning method is explained through the federated learning system of the present invention described above.

[0063] Figure 4 is a diagram illustrating the classification stage for training the head.

[0064] Referring to FIGS. 3 and 4, in the classification step (S10) for training the head, only the head is trained using Cross Entropy Loss while keeping the parameters of the body and bridge fixed, in order to obtain decision boundaries that fit local features. At this time, a global prototype is generated using the average of features extracted from the body, and a local prototype is generated using the average of features transformed from the bridge, and each is stored in memory. This is similar to the process of simultaneously establishing general diagnostic criteria and hospital-specific diagnostic criteria in a medical system.

[0065] The loss function of the head is as follows.

[0066]

[0067] F θ (x) corresponds to a body (layer) that acts as a feature encoder. The body extracts relevant features from input data x. The representation vector z generated from the body (210) is a bridge (layer) P h (z) passes through. This layer, designed for local prototype generation, takes feature z as a local feature space. Convert to. The head layer acting as the final classifier It is input as. Here, CE is the cross-entropy loss function, which is particularly effective for class classification through the probabilistic interpretation of local decision boundaries. After the head is updated, the features generated by the body and bridge. and Global prototype approximated by and local prototype It is extracted and stored in memory for use in the next step.

[0068] Next, the personalization early termination indicator is checked. If the rate of change of the body parameter decreases in consecutive rounds, it means that sufficient personalization has been achieved, so the bridge learning step is skipped and the process proceeds directly to the body learning step. Otherwise, the bridge learning step (S20) is performed. The personalization early termination indicator is as described above.

[0069] Figure 5 is a diagram illustrating the prototyping stage for training the bridge.

[0070] In the bridge training phase (S20), with the body and head parameters fixed, the bridge is trained by comparing the local prototype stored in memory with the current bridge output. This can be viewed as a process that further strengthens the hospital's expertise. As the communication rounds progress, the local body repeatedly aggregates the server's body parameters to generate a generalized representation. During this process, the local model updates the bridge to personalize the local prototype. In the prototyping phase (bridge training), with θ and φ fixed, the global-local similarity weights (i.e., We use the MSE loss with ) applied. The loss function is as follows.

[0071]

[0072] The bridge is updated by considering class-specific distances using w obtained from the previous communication round. Similar to the head, once the update is complete, the updated local prototypes are stored in memory for use in the next step. Meanwhile, in the above loss function is the aforementioned personalization early termination metric. If the personalization early termination metric is 1, it is determined that sufficient personalization has not yet been achieved, personalization learning is continued, and the MSE Loss is applied normally. On the other hand, if the personalization early termination metric is 0, it is determined that sufficient personalization has been achieved, personalization learning is stopped, the application of the MSE Loss is stopped, and the focus is then placed on generalization learning.

[0073] To explain using a hospital example, if the early personalization termination indicator is 1, it means "the hospital's expertise has not yet been sufficiently reflected, so personalization will continue," and if the early personalization termination indicator is 0, it means "the hospital's expertise has now been sufficiently reflected, so focus on improving general medical knowledge."

[0074] Figure 6 is a diagram illustrating the feature extraction stage (Localization Stage) for training the body.

[0075] In the body training step (S30), the body is trained while keeping the parameters of the bridge and the head fixed. At this time, global-local similarity weights are calculated based on the similarity between the global prototype and the local prototype, and these are applied to the loss function to adjust the degree of reflection of the client's characteristics. The calculated global-local similarity weights are stored in memory to be used for bridge training in the next round.

[0076] For body updates, both CE loss and MSE loss with global-local similarity weights are used. CE loss effectively maintains the discriminative ability based on class decision boundaries and enables the body to learn clear distinctions between classes, thereby facilitating better generalization. Meanwhile, MSE loss ensures that the representation vectors generated by the body become closer to the updated local prototypes, thus aligning more closely with the local space and contributing to personalization. In this process, φ and h are fixed. The loss function used for the body is as follows.

[0077]

[0078] Here, φ and h are the head and bridge parameters, respectively, and is the aforementioned personalization early termination indicator, and is GL-Sim Weight by class, is the updated local prototype, is the feature vector generated by the body. Also, y i represents the actual ground truth label. This value indicates the actual class of the input data. The remaining variables or parameters are as described above.

[0079] Finally, in the Body Parameter Aggregation step (S40), the server collects, integrates, and generalizes the body parameters of each client, then distributes them back to all clients. Subsequently, a new round begins from the classification step, where the head is trained again. Through this cyclical process, clients can learn generalized knowledge from the entire dataset while maintaining their specificity.

[0080] Training the head first follows the same logical sequence as a doctor establishing diagnostic criteria first. This is because personalization or generalization of features is meaningless without basic classification capabilities. Additionally, global and local prototypes are generated during the head training process, serving as reference points for the subsequent training of the bridge and body. In other words, it adopts a top-down approach where "what results are desired" is defined first, and then feature extraction and transformation are adjusted accordingly.

[0081] It is a highly strategic design of the present invention that the decision to terminate personalization early is located immediately before the bridge learning step. Since the bridge layer is a core component responsible for personalization, omitting this step is equivalent to halting the entire personalization process. If this decision were located elsewhere, unnecessary personalization could have already proceeded (if located further back) or be stopped before personalization is sufficiently completed (if located further ahead). Furthermore, the decision at this location allows for sufficient time to observe changes in body parameters.

[0082] Integrating only body parameters is a key strategy for a clear separation between personalization and generalization. Since the body is responsible for extracting basic features, sharing knowledge in this area enables all clients to possess robust basic feature extraction capabilities. On the other hand, because the bridge and head reflect the unique characteristics of each client, integrating them could actually dilute the specificity of individual clients. This is the same principle as the medical system, where basic diagnostic techniques are shared, but each hospital maintains its own professional interpretation methods and final diagnostic criteria independently.

[0083] Figure 7 is a table showing the results of comparing the performance of the present invention (FedSDP) and existing federated learning methods.

[0084] Referring to FIG. 7, the present invention achieved the best performance in both experimental environments (Pathological Setting and Practical Setting).

[0085] First, in the Pathological Setting (extreme situation), an accuracy of 83.67% (±0.06) was achieved on the CIFAR10 dataset, 65.37% (±0.18) on the CIFAR100 dataset, and 32.42% (±0.12) on the TinyImageNet dataset.

[0086] This demonstrates that the system operates effectively even in situations where each hospital handles only completely different specialties (e.g., Hospital A specializes in geriatric diseases, Hospital B specializes in pediatric diseases).

[0087] Next, in the Practical Setting, an accuracy of 82.13% (±0.32) was achieved on the CIFAR10 dataset, 53.94% (±0.27) on the CIFAR100 dataset, and 33.19% (±0.22) on the TinyImageNet dataset.

[0088] This proves that the system operates stably even in situations where there is a focus area but a diverse range of patients are treated, just like in a real hospital environment (e.g., Hospital A has 70% elderly patients and 30% general patients).

[0089] Particularly noteworthy is that the present invention demonstrates higher accuracy compared to existing methods across all datasets, maintains a relative advantage even as data complexity increases (CIFAR10 → CIFAR100), and exhibits excellent performance stability with a very small standard deviation (±0.06 to ±0.32).

[0090] These results demonstrate that the bridge layer and dynamic weighting mechanism proposed by the present invention effectively achieve a balance between personalization and generalization.

[0091] Figure 8 is a graph showing the change in accuracy per communication round for the CIFAR10 dataset in Pathological Setting (extreme situation).

[0092] Referring to Fig. 8, the present invention (FedSDP) exhibits a fast learning speed from the beginning and maintains a high accuracy of over 80% stably from about 25 rounds onwards. This is similar to a situation where each specialized hospital quickly acquires expertise in its specific medical field (e.g., geriatrics, pediatrics) and maintains it stably.

[0093] Figure 9 is a graph showing the change in accuracy per communication round for the CIFAR10 dataset in a practical setting.

[0094] Referring to FIG. 9, the present invention demonstrates excellent performance even in a real hospital environment treating various patient groups. Initially, it shows a learning curve similar to FedROD, but after about 50 rounds, it achieves higher accuracy and maintains it stably.

[0095] Figure 10 is a graph showing the change in accuracy per communication round for the CIFAR100 dataset in Pathological Setting.

[0096] Referring to Fig. 10, the present invention shows steady performance improvement even in a situation where the number of diseases to be diagnosed increases (CIFAR100). In particular, unlike FedCAC which shows a rapid decline in performance in the early stages, the present invention maintains a stable learning curve.

[0097] Figure 11 is a graph showing the change in accuracy per communication round for the CIFAR100 dataset in a practical setting.

[0098] Referring to FIG. 11, the present invention demonstrates superior performance compared to other methods even in a complex diagnostic environment (CIFAR100) and a situation where various patient groups are mixed. In particular, unlike the unstable learning patterns shown by FedProto or FedCAC, it achieves consistent and stable performance improvement.

[0099] Figure 12 is a table comparing the performance of the present invention (FedSDP) and existing federated learning methods according to client participation rate (ρ).

[0100] Referring to Fig. 12, the present invention demonstrates the best performance under all participation rate conditions. Even under low participation rates (ρ = 0.2), the present invention achieved a high accuracy of 82.35% (±0.01), which means that stable performance is maintained even when only 20% of all hospitals participate. In particular, it showed a lower standard deviation compared to FedROD (82.09%±0.16), which showed the closest performance, proving its superiority in terms of performance stability.

[0101] In an environment with a moderate participation rate (ρ = 0.6), it recorded an accuracy of 82.76% (±0.10), showing a gradual improvement in performance with increasing participation. This suggests that sufficient performance can be achieved with only partial cooperation from hospitals. Furthermore, it was observed that the performance gap with other methods widened further.

[0102] In a full participation scenario (ρ = 1.0) where all clients participate, the highest accuracy of 82.94% (±0.17) was achieved. Notably, the performance under full participation did not differ significantly from that under low participation rates. This is a very important characteristic demonstrating that the system can operate effectively even in real-world medical environments where the participation of all hospitals is difficult.

[0103] On the other hand, in the case of FedProto, the performance variation according to the participation rate was very large, ranging from 29.40% to 66.05%, and FedCAC showed limitations in that it could only operate under full participation conditions. These comparison results demonstrate that the present invention can provide stable and excellent performance even in environments with various participation rates.

[0104] Figure 13 is a table comparing the performance variance between clients of the present invention (FedSDP) and existing federated learning methods in various datasets using the coefficient of variance (CV).

[0105] Referring to Fig. 13, the present invention shows the lowest CV value across all datasets, which means the best balance of performance among clients. From the perspective of a medical environment, this indicates that there is no significant difference in overall diagnostic capabilities while each hospital maintains its expertise.

[0106] In particular, on the MNIST dataset, a very low CV value of 2.25 was recorded, which is a better result than FedROD (2.31), which showed the second best performance. This trend was maintained in the more complex dataset CIFAR10, where the present invention showed a CV value of 13.94, showing a significant difference from other methods. This demonstrates that the performance balance between clients is well maintained even as data complexity increases.

[0107] It is noteworthy that the present invention maintains a low CV value of 10.08 even in complex environments with a large number of classes, such as CIFAR100. This implies that there is not a significant difference in performance between medical departments, much like when diagnosing various diseases in a general hospital. On the other hand, FedCAC showed a relatively high CV value of 16.59, revealing the difficulty in maintaining performance balance between clients in complex environments.

[0108] In particular, traditional federated learning methods such as FedAvg and FedProx exhibit very high CV values ​​across all datasets, suggesting that it is difficult to achieve performance balance between clients by simply averaging models without personalization mechanisms.

[0109] FIG. 14 is a table comparing the performance of the present invention according to the presence or absence of a personalized early termination indicator (P-Stop Indicator) and the number of consecutive decreases (n) that serve as the judgment criteria.

[0110] Referring to Figure 14, the accuracy was 80.14% (±0.07) when the personalized early termination metric was not used, whereas the performance improved significantly to 82.13% (±0.32) when it was introduced and n=2. To draw an analogy to the medical environment, this means that each hospital can achieve better diagnostic capabilities when it stops strengthening its own expertise at an appropriate time and focuses on acquiring generalized medical knowledge.

[0111] When examining the effect of the number of consecutive decreases (n), when n=1, the performance was 81.63% (±0.25), which is interpreted as being due to the failure to secure sufficient expertise because personalization was stopped too early. On the other hand, when n was set to 3, the performance actually decreased to 81.22% (±0.06), which can be seen as a result of the acquisition of generalized knowledge being hindered by excessive personalization.

[0112] The experimental results show that n=2 is the optimal setting, suggesting that two consecutive observations of performance decline can serve as an appropriate criterion for finding the balance between personalization and generalization.

[0113] Figure 15 is a table comparing the performance of the present invention according to the global-local similarity weighting (GL-Sim Weight) and fixed weighting methods.

[0114] Referring to Fig. 15, the GL-Sim Weight method of the present invention achieved an accuracy of 82.13% (±0.32), demonstrating superior performance compared to all fixed weight methods. To use a medical environment analogy, this means that dynamically adjusting the balance between the characteristics of each hospital and general medical knowledge is more effective than adjusting it at a fixed ratio.

[0115] Looking at the cases using fixed weights, the lowest performance was observed at 80.26% (±0.03) when low weights (α = 0.1) were applied, which is interpreted as a result of not sufficiently reflecting the characteristics of the hospital. On the other hand, when high weights (α = 1) were applied, performance improved to 81.50% (±0.27), but it still fell short of the dynamic weighting method.

[0116] In particular, weights that decrease exponentially over time (α = e (-0.1·x) Even when applying ), it showed an accuracy of 81.37% (±0.030), indicating that it is difficult to achieve optimal performance with simple time-based weight adjustment. This suggests that the balance point between personalization and generalization must change dynamically depending on the situation of each hospital and the characteristics of the data.

[0117] FIG. 16 is a table showing experimental results analyzing the effect of each component of the present invention on performance.

[0118] Referring to Figure 16, the highest accuracy of 82.13% (±0.32) was achieved when all components of the present invention were applied. When no MSE loss function was used (w / o MSE), performance deteriorated significantly to 79.23% (±0.50), which indicates that it is difficult to obtain sufficient performance with only generalized learning without a mechanism that reflects the characteristics of the hospital.

[0119] When the MSE loss function was used only in the localization stage (MSE-LOC), a relatively high accuracy of 81.76% (±0.16) was observed, but when the MSE loss function was used only in the prototyping stage (MSE-PRO), a somewhat lower performance of 80.41% (±0.35) was observed. This suggests that feature reflection in the body learning stage may be more important than feature reflection in the bridge learning stage.

[0120] When the early personalization stop metric was not used (w / o P-Stop), the accuracy was 80.14% (±0.07). This indicates that if excessive personalization is not properly controlled, overall performance may actually deteriorate. To draw an analogy to the medical environment, this means that if each hospital focuses solely on strengthening its own expertise and neglects the learning of generalized medical knowledge, its overall diagnostic ability may decline.

[0121] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0122] The scope of the present invention is defined by the claims set forth below, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention. Explanation of the symbols

[0124] Federation learning system (10) Server (100) Client (200) Body (210) Bridge (220) Head (230) Memory (240) control unit (250)

Claims

Claim 1 A federated learning system that dynamically reflects the degree of personalization and generalization, comprising a plurality of clients and servers, wherein the client model provided by each client comprises: an artificial neural network-based body that extracts features from input data; an artificial neural network-based bridge that converts the features extracted from the body into personalized features tailored to the characteristics of the client; an artificial neural network-based head that performs classification based on the features converted by the bridge; and a control unit that controls the neural network learning of the body, bridge, and head. Claim 2 In claim 1, the control unit fixes the parameters of the body and the bridge when training the head, generates a global prototype from features extracted by the body, generates a local prototype from features transformed by the bridge, and stores each in the memory of the client; when training the bridge, fixes the parameters of the body and the head, trains the bridge based on a loss function between the local prototype stored in the memory and the current bridge output, generates a new local prototype from the trained bridge and stores it in the memory; when training the body, fixes the parameters of the bridge and the head, calculates a global-local similarity weight based on the similarity between the global prototype and the local prototype, applies the global-local similarity weight to a loss function to adjust the degree of reflection of the corresponding client's characteristics, stores the global-local similarity weight in the memory and uses it for training the bridge in the next round; and the server integrates and generalizes the parameters of the body collected from the plurality of clients, and the generalized A federated learning system that dynamically reflects the degree of personalization and generalization, characterized by distributing parameters to the aforementioned multiple clients. Claim 3 A federated learning system that dynamically reflects the degree of personalization and generalization according to claim 2, wherein the control unit activates a personalization early termination indicator when the rate of change of the body's parameters decreases in a preset number of consecutive rounds, and when the personalization early termination indicator is activated, omits the bridge learning in subsequent rounds and excludes the application of the global-local similarity weight during body learning. Claim 4 A federated learning system that dynamically reflects the degree of personalization and generalization according to claim 2, wherein the control unit calculates the mean squared error for each class for the global prototype and the local prototype to calculate the global-local similarity weight, and generates the global-local similarity weight for each class by multiplying the mean squared error for each class by a dataset characteristic reflection coefficient, wherein the dataset characteristic reflection coefficient is determined based on the number of classes of the dataset and the distribution characteristics between classes. Claim 5 A federated learning method that dynamically reflects the degree of personalization and generalization among multiple clients and servers, wherein a client model equipped by each client is configured to extract features from input data through an artificial neural network-based body, convert the features into personalized features according to the characteristics of the client through an artificial neural network-based bridge, and perform classification based on the personalized features through an artificial neural network-based head; a classification step in which the head performs classification based on the features converted by the bridge and a control unit trains the head; a feature conversion step in which the bridge converts features extracted from the body into personalized features according to the characteristics of the client and the control unit trains the bridge; and a feature extraction step in which the body extracts features from input data and the control unit trains the body.

Images