Associative learning systems, associative learning methods, and programs

Abstract

This system provides a highly reliable federated reinforcement learning environment where only users who have registered their biometric data can participate through biometric authentication. [Solution] According to one embodiment, the federated learning system includes multiple edge terminals and a central server. Each edge terminal includes a biometric authentication unit that acquires user-specific biometric data from the user's body parts, a biometric data storage unit that stores the biometric data acquired by the biometric authentication unit, a learning unit that executes the learning process for an AI model, and a management unit that causes the learning unit to execute the learning process if the user's biometric data is registered on the edge terminal. The central server includes a biometric data existence confirmation unit that checks whether biometric data is registered on the edge terminal, a distribution unit that distributes the AI ​​model to multiple edge terminals, a collection unit that collects learning results from multiple edge terminals, and an update unit that updates the AI ​​model based on the learning results collected from multiple edge terminals.

Description

Technical Field

[0001] The present invention relates to a federated learning system with a biometric authentication function, particularly trusted federated reinforcement learning. It is a system for providing a federated learning environment in which only users who have registered biometric authentication can participate, suppressing the mixing of poisoned data, and generating a safe and privacy-protected federated learning model. Furthermore, since the user's owned terminal will be used as an edge device, the federated learning process is performed while checking the usage status of the terminal resources by an AI agent. In addition, it includes a mechanism for providing token rewards to the participating users of the federated learning, becoming a federated learning system that provides convenience and economy to the users.

Background Art

[0002] Federated learning in which users participate has great advantages in terms of privacy protection and decentralization, but at the same time has several problems. Federated learning by edge terminals is a technology for jointly constructing a machine learning model without sharing the data owned by multiple individuals, but this technology is accompanied by serious problems regarding data privacy and security. In federated learning, establishing trust, together with data privacy, model fairness, and algorithm transparency, is a major issue, such as the possibility that other participants will provide illegal data or models. These concerns have become a major obstacle to the popularization of federated learning, and the development of new technologies and methodologies for establishing trust in federated learning is required. Therefore, technological innovation for improving the safety and reliability of federated learning is an urgent task.

[0003] For example, Japanese Patent Publication No. 2023-2338 (Patent Document 1 below) proposes a learning system having multiple client terminals, a learning data management server, and a learning server, wherein the system updates a model based on the learning data, the learning server manages a common model, the client terminals and the learning data management server manage personal data, a different individual model is generated for each individual from the common model and the personal data, the common model and the individual model are shared with the learning server, the common model is received from the learning server, the common model and the individual model are updated based on the personal data, the updated common model and individual model are sent to the learning server, and the learning server classifies the common model and the individual model sent from the multiple client terminals based on the individual model sent from the learning data management server, and updates the common model and the individual model according to the classification result. In this learning system, biometric data used for biometric authentication is used as data to be learned in order to reduce the difference in authentication accuracy between individuals with different attributes through federated learning. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2024-2338 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] Federative learning, which uses user-owned devices such as smartphones as edge devices, offers significant advantages in terms of privacy protection and decentralization, but it also faces several major challenges. In particular, data privacy, security, and trust establishment are major issues. Specifically, concerns such as the possibility of participants' data being misused, the possibility of learning results favoring certain participants, the possibility of the learning process being a black box and the unclear process of how the results were derived, and the possibility of other participants providing fraudulent data or models are major obstacles to the widespread adoption of feasible learning.

[0006] Furthermore, using users' devices as edge devices presents technical challenges such as communication environment constraints, limited computing resources, and concerns about battery consumption. For example, smartphones operate in communication environments with limited bandwidth, such as Wi-Fi® and mobile networks, which can become a bottleneck when exchanging large amounts of model parameters. In addition, unstable communication environments can delay model updates and reduce the efficiency of learning. Moreover, the processing power of CPUs and GPUs in edge devices may be insufficient for training large-scale models, and while sufficient memory capacity is required to process large amounts of data, memory is limited. In addition to this, federative learning is computationally intensive, rapidly draining the device's battery, and the user experience deteriorates when the device becomes unusable due to a dead battery.

[0007] This invention was conceived in view of the above points, and aims to provide a highly reliable federative reinforcement learning environment in which only users who have registered their biometric data using a biometric authentication function can participate. Furthermore, by equipping the edge terminal with an AI agent that manages the learning, it aims to provide a trusted federative learning system that allows users to participate in federative learning while maintaining convenience and comfort for the user. [Means for solving the problem]

[0008] According to one embodiment, the federated learning system is a federated learning system including a plurality of edge terminals and a central server, wherein the edge terminal includes a biometric authentication unit 11 that acquires user-specific biometric data from the user's biological parts, a biometric data storage unit that stores the biometric data acquired by the biometric authentication unit, a learning unit that executes the learning process of an AI model, and a management unit that causes the learning unit to execute the learning process if the user's biometric data is registered in the edge terminal, and the central server includes a biometric data existence confirmation unit that confirms whether biometric data is registered in the edge terminal, a distribution unit that distributes the AI ​​model to the plurality of edge terminals, a collection unit that collects learning results from the plurality of edge terminals, and an update unit that updates the AI ​​model based on the learning results collected from the plurality of edge terminals. [Effects of the Invention]

[0009] Federative learning using user-owned devices as edge devices, as described in this invention, is still largely in the research stage. However, its potential is being explored in various fields, and several demonstration experiments and specific use cases have been reported. In the medical field in particular, by learning in a distributed manner without aggregating patient data held by each hospital, it is possible to build more accurate disease diagnostic models and realize personalized medicine that predicts the optimal drug for each individual patient. In the mobile app field, it is possible to build word prediction models using keyboard input data on smartphones, and to develop highly accurate speech recognition models without sharing voice data from each device. Furthermore, in the fields of traffic prediction and smart homes, it is expected to be possible to develop models that predict real-time traffic conditions using smartphone location data, and models that realize comfortable living spaces by integrating smart home appliance data from each household. Therefore, federative learning using user-owned devices is attracting attention as a new technology that balances the protection of personal information and the utilization of data, and its application in various fields is expected in the future. In particular, its use is expected in areas with a large social impact, such as the medical field and smart city field, and its industrial applicability is very large. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of a federated learning system according to the first embodiment. [Figure 2] This is a block diagram illustrating the general outline of the learning process. [Figure 3] This is an explanatory diagram illustrating examples of sensor usage. [Figure 4] This is an explanatory diagram illustrating an example of an image from which palm print and vein information has been extracted. [Figure 5] This is a block diagram illustrating the process for generating biometric data. [Figure 6] This is a block diagram illustrating the general outline of the learning process. [Figure 7] This figure shows an example of a federated learning system according to the third embodiment. [Figure 8] This figure shows an example of a federated learning system according to the fourth embodiment. [Figure 9] This figure shows an example of a federated learning system according to the fifth embodiment. [Figure 10] This figure shows an example of a federated learning system according to the sixth embodiment. [Modes for carrying out the invention]

[0011] Hereinafter, one embodiment of the present invention will be described with reference to the attached drawings.

[0012] <First Embodiment> (System Configuration) First, Figure 1 shows the system configuration of this embodiment.

[0013] The edge terminal 1 of this embodiment includes a biometric authentication unit 11 that acquires user-specific biometric data from a user's biological part. Biometric authentication often involves capturing a biological body (such as a palm, face, eyes, whole body, etc.) to obtain a biometric image. In this embodiment, as shown in FIG. 3, the palm is photographed with the camera 20 of the smartphone. The biometric features unique to the user are extracted from the biometric image and stored in the biometric data storage unit 12. The biometric features are also used as template data for biometric authentication described later.

[0014] Furthermore, the edge terminal 1 includes a learning unit 14 that executes the federated learning process of the AI model, and a management unit 13 that enables the learning unit 14 to execute the federated learning process safely and efficiently. The above-mentioned personal information and the biometric data associated with the personal information are transmitted to the identifier generation device through, for example, a communication path.

[0015] The central server 4 includes a biometric registration confirmation unit that checks whether biometric data is registered in the edge terminal 1. This system provides a federated learning environment in which only users who have performed biometric authentication can participate, suppresses the mixing of poisoned data, and generates a safe and privacy-protected federated learning model. Therefore, before starting the federated learning, it is checked whether biometric data is registered. If it is registered, the flow of the federated learning process described later starts. If it is not registered, the processing for the edge terminal 1 ends.

[0016] Furthermore, the central server 4 includes a distribution unit 43 that distributes the AI model to the plurality of edge terminals 1, a collection unit 42 that collects the learning results from the plurality of edge terminals 1, and an update unit 44 that updates the AI model based on the learning results collected from the plurality of edge terminals 1.

[0017] The above-mentioned edge terminal 1 and the central server 4 communicate necessary information and data through, for example, a communication path.

[0018] In this embodiment, the edge terminal 1 is the smartphone described above, but various IoT devices such as smart speakers, smartwatches, and wearable devices can be used for federated learning, allowing for the development of new functions such as collecting sensor data and user behavior data to provide personalized services. Furthermore, automobiles such as autonomous vehicles and connected cars can also be used as edge terminals 1, enabling federated learning that supports safer and more comfortable driving by utilizing traffic conditions and driving data collected by in-vehicle cameras and sensors. In addition, sensors and control devices installed in factories and plants, although not for personal use, can also become edge terminals 1 and be subjects of federated learning.

[0019] Next, we will explain the details of the trusted federative learning system. Here, edge device 1 is a smartphone, and the camera images stored on the smartphone are the target of federative learning. Biometric authentication will be palm recognition using the smartphone's camera.

[0020] (Biometric authentication section 11) A typical biometric authentication unit 11 in this embodiment is a camera 20 mounted on a smartphone, which is an edge terminal 1. In this embodiment, the palm is photographed by the imaging device, and the biological body part is the palm 21. In order to acquire biological data of the palm, it is necessary to irradiate the palm of the human body with light that includes at least red light in the visible light region, so in this embodiment, the light emitted from the smartphone's display 22 is used.

[0021] (Biological feature extraction unit) The biofeature extraction method extracts the palm print and vein patterns from palm images obtained using a smartphone camera. The biofeature extraction method of this embodiment is described in detail below.

[0022] Image processing is performed on the captured image of the palm so that the vein shape and palm print shape appear reliably. Generally, with cameras in the visible light range, it is easy to acquire the palm print shape on the surface of the palm, but it is difficult to acquire the clear shape of the veins inside the body. However, based on the knowledge of the main inventor, with the improvement of the biometric authentication matching method described in detail later, differences in density do not affect authentication as long as there is no data loss. Therefore, the simplest and most stable processing is to generate an 8-bit (256-level) grayscale image from the RGB color image. The calculation formula can be obtained by multiplying the RGB values ​​of each pixel by a coefficient, as follows: Grayscale value = 0.21 * R + 0.72 * G + 0.07 * B Alternatively, you can create a grayscale image using (R + G + B) / 3 or 0.2989 * R + 0.5870 * G + 0.1140 * B.

[0023] Next, by applying a line segment extraction filter to the grayscale image, an image 23 is obtained in which the vein shape and palm print shape are extracted, as shown in Figure 4. Furthermore, a circular region is set in the center of the palm, and a circular image 24 is obtained that includes the vein shape and palm print shape within that region. To obtain linear transition, rotation, and scaling-invariant features from the resulting image, the following processing is performed as shown in Figure 5.

[0024] The Radon transform 26 is used as a method for extracting features. This method projects the two-dimensional palm print and vein shape images 24 onto an axis in the θ direction (θ = 0 to 180°), and expresses the features using the position ρ on the projection axis and θ.

[0025] Next, a Fourier transform 27 in the ρ direction is applied to the Radon-transformed feature data, and only the amplitude component is extracted. Specifically, the amplitude component 28 is obtained by taking the square root of the sum of the squares of the real and imaginary parts after the Fourier transform. By extracting only the amplitude component, the model becomes linearly shift-invariant in the ρ direction.

[0026] Next, a logarithmic coordinate transformation 29 is performed with respect to the ρ direction. Specifically, ρ is transformed into log(ρ), and the feature data is in a logarithmic polar coordinate system. The feature extraction data 30 obtained in this way becomes biological data that represents the characteristics of vein shape and palm print shape.

[0027] First, to facilitate the subsequent similarity calculation process, the feature extraction data in logarithmic polar coordinates is converted into a phase-only image, in which only the phase is extracted. Specifically, a Fourier transform 31 is performed on the feature extraction data 30 in logarithmic polar coordinates to set the amplitude component to 1. Since the phase-only image transformation is also well known, a detailed explanation is omitted. In this embodiment, the biological data 33 obtained by the Fourier transform becomes the biological template data.

[0028] (Biometric data storage unit 12) The biological data 33 obtained through the process described above is stored in the biological data storage unit 12.

[0029] (Biometric data presence confirmation unit 41) The biometric data existence verification unit 41 of the central server 4 sends a request to the management unit 13 of the multiple edge terminals 1 to check whether biometric data of a predetermined data length is stored in the biometric data storage unit 12 of the edge terminal 1.

[0030] (Management Department 13) The management unit 13 of the edge terminal 1 includes a biometric data existence verification unit 131, which, in accordance with the biometric data existence verification request, checks whether biometric data of a predetermined data length is stored in the biometric data storage unit 12 within the edge terminal 1.

[0031] In this embodiment, we confirm that the biometric data 33 is a 256x128=32,768-dimensional complex vector, and that 256x128×8=262,144 bits of binary data are stored.

[0032] The management unit 13 transmits the biometric data existence confirmation result to the central server 4, and the biometric data existence confirmation unit 41 of the central server 4 receives the biometric data existence confirmation result. If the existence of biometric data can be confirmed based on the existence confirmation result, the process proceeds to the federated learning implementation flow described in the following sections for the edge terminal 1, and terminates if the existence cannot be confirmed.

[0033] (A modified example of the biometric authentication unit 11) In this embodiment, palm recognition using a smartphone camera was used, but facial recognition and iris recognition using a smartphone camera can also be considered biometric authentication units 11. Furthermore, biometric authentication devices mounted on a smartphone, such as a fingerprint authentication device, can also be used. In addition, in automobiles, facial recognition, palm recognition, and behavioral authentication using an in-vehicle camera can also be used.

[0034] Furthermore, in small edge devices such as IoT devices, it is also possible to use fingerprint authentication devices or vein authentication devices connected to the device via USB or wireless connection.

[0035] (Associative Learning Architecture) In this embodiment, since associative learning is performed on images, an initial model that serves as the starting point for learning is generated using a CNN (Convolutional Neural Network), a deep learning model widely used in the field of image recognition, as its architecture. CNNs mimic the workings of the human visual cortex and automatically extract features from images, making them useful for various tasks such as image classification, object detection, and image generation. Therefore, a detailed explanation will be omitted here. Of course, it is also possible to choose other architectures.

[0036] (Update section 44) The initial global learning model distributed by the update unit 44 has its parameters set with optimal initial values ​​for efficient learning, rather than random initial values. The generated model is converted into an optimized format so that it can run efficiently on the edge device 1. For example, an open-source framework specifically for mobile devices can be used. Alternatively, various frameworks may be mixed together.

[0037] (Transmission of global learning model) The distribution unit 43 distributes the global learning model to multiple smartphones participating in federated learning via a secure communication channel that incorporates measures such as encryption to ensure the security of the model. This allows each edge terminal 1 to proceed with learning using its individual data based on the downloaded model.

[0038] (Learning Section 14) The learning unit 14 saves the received model within the terminal.

[0039] (Learning processing execution unit 135) The learning process execution unit 135 has a function to check whether predetermined conditions for performing the learning process, which will be described in detail later, have been met after the model has been stored in the terminal, and if the conditions are met, it instructs the learning unit 14 to perform the learning process.

[0040] (Data for reinforcement learning) The learning unit 14 preprocesses the image data stored in the terminal. 1. Adjust images to a consistent size. For example, resize images to 224x224 pixels. 2. Keep the pixel values ​​within a specific range. 3. Increase data diversity by rotating, flipping, cropping, etc., images.

[0041] (Local reinforcement learning) The learning unit 14 inputs preprocessed image data into a locally stored model and updates the model parameters using backpropagation. More specifically, the learning unit 14 creates a local model by copying the global model or performing random initialization, performs forward propagation calculations on the input image, calculates the error between the output and the correct label, and updates the model parameters based on the gradient obtained by backpropagation to train the local model. Hyperparameters such as the learning rate and number of epochs are either set from the central server 4 or automatically adjusted on the smartphone to optimize the learning process.

[0042] (Sending local reinforcement learning parameters) The learning unit 14 reduces the amount of data transmitted by sending only the differences in the model parameters after local reinforcement learning, and sends them to the central server 4 via a secure communication channel. It is also possible to transmit parameters while protecting privacy using techniques such as differential privacy.

[0043] (Collection of results from collaborative reinforcement learning) The data collection unit 42 aggregates the parameters transmitted from each smartphone, generates a new global model using methods such as federated averaging, and stores the aggregated parameters for use in the next training round.

[0044] The update unit 44 generates a new global model based on the aggregated parameters. Furthermore, it converts the model generated here into an optimized format for distribution to each smartphone and distributes the model to each smartphone via a secure communication channel.

[0045] The update unit 44 repeats the above steps and continues learning until the accuracy of the model improves.

[0046] The associative learning flow described above is just one example; various methods are possible. For example, a series of processes in a processing flow can also be represented by the following formula. 1. Model initialization and distribution Model W0, initialized on central server 4, is distributed to each edge terminal 1. 2. Local learning Each client k updates the model using local data Dk with the following formula.

[0047]

number

[0048] 3. Aggregation of model parameters Using the model parameters Wk sent from each client, the central server 4 generates a new global model Wglobal. In the federated averaging method, the calculation is performed using the following formula:

[0049]

number

[0050] 4. Distribution of the global model The updated global model Wglobal has been distributed to each client. 5. Repetition Repeat the above process and continue learning until convergence occurs.

[0051] Up to this point, the federated learning system includes multiple edge terminals 1 and a central server 4, where the edge terminal 1 is a biometric authentication unit 11 that acquires user-specific biometric data from the user's body parts. The system includes a biometric data storage unit 1212 for storing biometric data acquired by the biometric authentication unit 11, a learning unit 14 for executing AI model learning processing, and a management unit 13 for causing the learning unit 14 to execute learning processing if user biometric data is registered on the edge terminal 1. The federated learning system described above includes a central server 4 comprising a biometric data presence confirmation unit 41 that checks whether biometric data is registered on the edge terminal 1, a distribution unit 43 that distributes the AI ​​model to the multiple edge terminals 1, a collection unit 42 that collects learning results from the multiple edge terminals 1, and an update unit 44 that updates the AI ​​model based on the learning results collected from the multiple edge terminals 1. The edge terminals 1 and the central server 4 communicate via a network. This system configuration is merely an example, and other configurations are possible. For example, it is possible to interpose an edge computer in the network and share the processing of the central server 4. It is also possible to divide the edge terminals 1 into small groups and perform distributed processing among the edge terminals 1 within each group.

[0052] <Second Embodiment> Federated learning is a technology that leverages the advantages of both centralized management and distributed processing, enabling the construction of large-scale machine learning models while protecting privacy. It reflects a new concept of a more decentralized, individual-driven internet. Therefore, in response to requests for participation in federated learning, a mechanism is provided that allows individual users to examine the content of the task and accept or reject it.

[0053] (System Configuration) As shown in Figure 6, The central server 4 further includes a participation request distribution unit 45 that calls upon the multiple edge terminals 1 to participate in federated learning. The management unit 13 is a collaborative learning system as described in the first embodiment, characterized in that it has a task content review function that determines whether to accept or reject a request to participate in collaborative learning based on the content of the task.

[0054] (Participation Request Distribution Section 45) The biometric data existence verification unit 131 receives the biometric data existence verification result and, after confirming that biometric data is registered, makes a request to each terminal to disclose the task details and invite them to participate in federated learning.

[0055] (Task content review function) The management unit 13 of the edge terminal 1 is equipped with a task content review function, and it determines whether to accept or reject the request content notified from the participation request distribution unit 45 based on the criteria described in the section.

[0056] (1) General-purpose tasks General-purpose tasks are tasks that do not require users to perform any special actions or behaviors. Users can learn while respecting privacy by simply using their smartphones in their daily lives, for example, by using messages on social media, images they have taken, and location-based activity data. These are tasks that place little burden on the user. (2) Dedicated tasks A dedicated task is one that requires the user to perform a specific action or behavior, or that imposes time constraints or information constraints. For example, a task that specifies, such as "Take 50 photos of people wearing your preferred fashion in the city," would be considered a dedicated task. (3) Multitasking Check if the system is already participating in other associative learning activities. For example, if the tasks are similar, the workload is less, multitasking is possible, and transfer learning is easier. Therefore, in the case of multitasking, the content of the tasks needs to be further examined.

[0057] The management unit 13 sends a response to the request to the central server 4, and the central server 4 receives the participation request. Based on the confirmation result, the process proceeds to the federated learning implementation flow described above with edge terminals 1 that have been confirmed to accept the request, and if the request is rejected, the process with that edge terminal 1 is terminated. The configuration other than that described above is basically the same as that of the first embodiment, so further detailed explanation of the second embodiment is omitted.

[0058] <Third Embodiment> Federative learning using smartphones can be challenging due to multiple constraints on the edge device 1, such as limited computing resources, communication environment, and battery capacity, which can make training large-scale models and their efficient execution difficult. Specifically, these constraints include delays in exchanging model parameters due to bandwidth limitations, slower training speeds due to insufficient processing power, limitations in data processing capabilities due to memory capacity constraints, and a decline in user experience due to increased battery consumption.

[0059] As shown in Figure 7, the management unit 13 is characterized by having a resource management unit 132 that manages the timing at which the learning unit 14 executes the learning process based on the call and communication status of the edge terminal 1 and the status of resources 15 such as the CPU, memory, and battery. Since the configuration other than that described above is basically the same as that of any one of the first to second embodiments, a further detailed explanation of the third embodiment will be omitted.

[0060] <Fourth Embodiment> (System Configuration) The edge terminal 1 further comprises a generation unit that generates an identifier based on the biometric data, and the collection unit 42 is configured to further collect the identifier from the plurality of edge terminals 1.

[0061] As shown in Figure 8, the management unit 13 of the edge terminal 1 has an identifier generation unit 133, which instructs the identifier generation unit 16 to generate an identifier after biometric data is registered. The identifier generation unit 16 generates an identifier based on the biometric data. The generated identifier is stored in an arbitrary location, such as the biometric data storage unit 12.

[0062] As mentioned above, the learning unit 14 is configured to send the model parameters to the central server 4 after local reinforcement learning is completed. The identifier is sent to the central server 4 at the same time as the parameter transmission. This makes it possible to associate the transmitted learning parameters with the users participating in federated learning.

[0063] (Identifier generation method) The following describes how to generate identifiers from biometric data.

[0064] As mentioned above, biological data can be represented as an N-dimensional vector. Specifically, the biological data in this embodiment is a complex number sequence with 256 x 128 elements, which becomes a complex number vector with 256 x 128 = 32,768 dimensions.

[0065] A 256x128=32,768-dimensional complex vector has 8 bits in both its real and imaginary parts, so it can be represented as 256x128×8=262,144 bits of binary data.

[0066] It is well known that binary data consists of the numbers "0" and "1" in binary, and is a data format that computers can read. Considering the combinations of two binary digits, there are four possibilities: "00" and "01", and "10" and "11". If we connect these 262,144 binary digits as a single unit, we get a chain-like structure with 131,072 sequences. This data sequence becomes the identifier.

[0067] For example, if we replace the two binary digits "00" with A, "01" with T, "10" with G, and "11" with C, we get the same structure as biological DNA. As is widely known, biological DNA is a base sequence combining four substances: adenine (A), thymine (T), guanine (G), and cytosine (C), and is unique to each individual. Therefore, identifiers with the same sequence structure as human DNA can be made unique. Since the configuration other than that described above is basically the same as that of any one of the first to third embodiments, a further detailed explanation of the fourth embodiment will be omitted.

[0068] <Fifth Embodiment> (System Configuration) The configuration of the token reward provision system will be explained with reference to Figure 9. The central server 4 further includes a token provision unit 46 that provides token-type rewards to the edge terminals 1 that have participated in federated learning. The token receiving unit 134 of the edge terminal 1 is configured to accept the reward if the central server 4 provides it and the user's identity can be verified through biometric authentication, and to refuse to accept the reward if the user's identity cannot be verified.

[0069] More specifically, the update unit 44 generates various models, performs simulations, and updates the models. In this system, it is possible to visualize the providers of data and parameters for each updated model using identifiers. Blockchain is a common method for this visualization. Furthermore, a method that does not change the data length of the identifier can be considered, mimicking the meiotic division of living organisms. Using these visualization methods, it is possible, for example, to aggregate the time spent on learning and pay rewards as compensation for participating in federative learning. It is also possible to pay token rewards to users who provide parameters to high-performing models according to their contribution.

[0070] Token rewards can be paid, for example, by a central server 4 equipped with a wallet 17, which then pays the rewards to the wallet 17 on each edge terminal 1. Alternatively, token rewards can be paid using smart contracts with NFTs (Non-Fungible Tokens) in response to the completion of predetermined tasks. Since the methods of paying tokens are widely known, we will omit their explanation here.

[0071] (Identity verification using biometric authentication) To prevent fraudulent receipt of these tokens through impersonation, biometric authentication data can be used. As mentioned above, the biometric data 33 is stored in the biometric data storage unit 12. This is template data for biometric authentication. Before receiving the token, identity verification When authentication is required, the biometric authentication unit 11 acquires an image for authentication, extracts biometric features as described above, and generates authentication data. Furthermore, the biometric authentication unit 11 performs individual authentication by comparing the authentication data with the template data.

[0072] Furthermore, federated learning platform operators can also issue their own tokens. These tokens function as currency on the platform and are used as rewards for users or as payment for services. For example, it is possible to purchase a smartphone battery using these tokens and send it to the user. Alternatively, communication data can be purchased with these tokens and used by the user. Therefore, in this embodiment, the scope of the present invention extends not only to digital tokens but also to the provision of goods and services. Since the configuration other than that described above is basically the same as that of any one of the first to fourth embodiments, a further detailed explanation of the fifth embodiment will be omitted.

[0073] <Sixth Embodiment> The AI ​​agent residing within the edge device 1, such as a smartphone, is a personal assistant that responds to various user needs. It collects surrounding data, learns the optimal patterns for the user, and makes proactive suggestions. The AI ​​agent evolves the edge device 1 from a mere communication device into an intelligent partner that supports our lives. In this embodiment, the AI ​​agent 18 autonomously handles all the functions of the management unit 13 on the edge device 1 of the federated learning system described in detail above.

[0074] (System Configuration) As shown in Figure 10, the management unit 13 is an AI agent 18, and the AI ​​agent 18 autonomously determines whether to accept or reject a request to participate in the federative learning based on the content of the task, the amount of the reward, and the resource usage status of the terminal. Since the other basic configurations are essentially the same as any one of the first to fifth embodiments described above, a further detailed explanation of the sixth embodiment will be omitted.

[0075] The AI ​​agent 18 in this system not only autonomously decides whether to participate in federative learning, but also achieves more advanced efficiency in terminal resources. Specifically, it learns the user's app usage patterns and can optimize which apps to launch and when, and which apps to run in the background. This reduces battery consumption and allows for longer smartphone use. It also enables maximum utilization of storage capacity by automatically deleting unnecessary data and moving frequently used data to high-speed storage. Depending on the network conditions, it adjusts the amount and quality of data transfer, and if the communication environment is poor, it compresses data or reduces resolution to ensure smooth communication.

[0076] The AI ​​agent 18 in this system analyzes the user's behavior history in conjunction with time of day and location to allocate resources optimally to that situation. For example, it can prioritize the navigation app during commutes and the music streaming app at home, providing services tailored to the user's behavior patterns.

[0077] The AI ​​agent 18 in this system builds a learning model that takes into account context such as the user's emotional state and surrounding environment. For example, it can have the function to provide appropriate services according to the situation, such as reducing notifications when the user wants to concentrate or suggesting calming music when the user wants to relax.

[0078] Furthermore, the AI ​​agent 18 in this system learns the user's past behavior patterns and predicts future behavior. Based on this prediction, it secures the necessary resources in advance, enabling smoother operation. For example, it can automatically launch necessary apps before a meeting or automatically start charging when the battery level is low, enabling predictive resource management.

[0079] <Note> This embodiment includes the following disclosures.

[0080] (Note 1) A federated learning system including multiple edge terminals and a central server, The aforementioned edge terminal is A biometric authentication unit that acquires user-specific biometric data from the user's body parts, A biometric data storage unit for storing biometric data acquired by the biometric authentication unit, A learning unit that performs the training process for the AI ​​model, If the user's biometric data is registered on the edge terminal, the system includes a management unit that causes the learning unit to execute a learning process. The aforementioned central server, A biometric data existence confirmation unit that confirms whether biometric data is registered in the edge terminal, A distribution unit that distributes the AI ​​model to the multiple edge terminals, A collection unit that collects learning results from the aforementioned multiple edge terminals, An update unit updates the AI ​​model based on the learning results collected from the aforementioned multiple edge terminals, A federative learning system equipped with the following features.

[0081] (Note 2) The central server further includes a participation request distribution unit that calls upon the multiple edge terminals to participate in federated learning. The management unit is characterized by having a task content review function that determines whether to accept or reject a request to participate in the federated learning based on the content of the task. The associative learning system described in Appendix 1.

[0082] (Note 3) The management unit is characterized by having a resource management function that manages the timing at which the learning unit executes the learning process based on the resource status of the edge terminal. The associative learning system described in Appendix 1 or 2.

[0083] (Note 4) The aforementioned resources include memory, processor, and battery. The associative learning system described in Appendix 3.

[0084] (Note 5) The learning results include the parameters of the AI ​​model. The associative learning system described in Appendix 1 or 2.

[0085] (Note 6) The aforementioned management department is an AI agent. The associative learning system described in Appendix 1 or 2.

[0086] (Note 7) The biometric authentication unit is characterized by being a biometric authentication device connected to the edge terminal. The associative learning system described in Appendix 1 or 2.

[0087] (Note 8) The edge terminal further comprises an identifier generation unit that generates an identifier based on the biometric data. The associative learning system described in Appendix 1 or 2.

[0088] (Note 9) The collection unit further collects the identifiers from the plurality of edge terminals. The associative learning system described in Appendix 8.

[0089] (Note 10) The central server further comprises a provisioning unit that provides token-based rewards to the edge terminals participating in federative learning. The associative learning system described in Appendix 1 or 2.

[0090] (Note 11) If the edge terminal receives the reward from the central server, it will accept the reward if the user's identity can be verified through biometric authentication, and will refuse to accept the reward if the user's identity cannot be verified. The associative learning system described in Appendix 10.

[0091] (Note 12) The distribution unit distributes the updated AI model to the multiple edge terminals. The associative learning system described in Appendix 1 or 2.

[0092] (Note 13) The update unit generates multiple AI models based on the learning results collected from the multiple edge terminals, and selects one AI model as the updated AI model based on the evaluation results of each AI model. The associative learning system described in Appendix 1 or 2.

[0093] (Note 14) The management unit is characterized by being an AI agent that autonomously determines whether to accept or reject a request to participate in the federative learning based on the content of the task, the amount of the reward, and the resource usage status of the terminal. The associative learning system described in Appendix 2.

[0094] (Note 15) A federated learning method performed by multiple edge terminals and a central server, The aforementioned edge terminal is By obtaining user-specific biometric data from the user's body parts, The biometric data acquired by the biometric authentication unit is stored, If the user's biometric data is registered on the edge device, the AI ​​model's training process is executed. The aforementioned central server, Check if biometric data is registered on the aforementioned edge terminal. The AI ​​model is distributed to the multiple edge terminals. Collect learning results from the aforementioned multiple edge devices, The AI ​​model is updated based on the learning results collected from the aforementioned multiple edge devices. Associative learning methods.

[0095] (Note 16) Multiple edge terminals and a central server, The aforementioned edge terminal is By obtaining user-specific biometric data from the user's body parts, The biometric data acquired by the biometric authentication unit is stored, If the user's biometric data is registered on the edge device, the AI ​​model's training process is executed. The aforementioned central server, Check if biometric data is registered on the aforementioned edge terminal. The AI ​​model is distributed to the multiple edge terminals. Collect learning results from the aforementioned multiple edge devices, The AI ​​model is updated based on the learning results collected from the aforementioned multiple edge devices. A program that implements associative learning methods.

[0096] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims, not in the sense described above, and is intended to include all modifications in the sense and scope equivalent to the claims. Furthermore, the present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims, and embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of Symbols]

[0097] 1: Edge devices 4: Central Server

Claims

1. A federated learning system including multiple edge terminals and a central server, The aforementioned edge terminal is A biometric authentication unit that acquires user-specific biometric data from the user's body parts, A biometric data storage unit for storing biometric data acquired by the biometric authentication unit, A learning unit that performs the AI ​​model training process, If the user's biometric data is registered on the edge terminal, the system includes a management unit that causes the learning unit to execute a learning process. The aforementioned central server is A biometric data existence confirmation unit that confirms whether biometric data is registered in the edge terminal, A distribution unit that distributes the AI ​​model to the multiple edge terminals, A collection unit that collects learning results from the aforementioned multiple edge terminals, An update unit updates the AI ​​model based on the learning results collected from the aforementioned multiple edge terminals, A federative learning system equipped with the following features.

2. The central server further includes a participation request distribution unit that calls upon the multiple edge terminals to participate in federated learning. The management unit is characterized by having a task content review function that determines whether to accept or reject a request to participate in the collaborative learning based on the content of the task. The associative learning system according to claim 1.

3. The management unit is characterized by having a resource management function that manages the timing at which the learning unit executes the learning process based on the resource status of the edge terminal. The associative learning system according to claim 1 or 2.

4. The aforementioned resources include memory, processor, and battery. The associative learning system according to claim 3.

5. The learning results include the parameters of the AI ​​model. The associative learning system according to claim 1 or 2.

6. The aforementioned management department is an AI agent. The associative learning system according to claim 1 or 2.

7. The biometric authentication unit is characterized by being a biometric authentication device connected to the edge terminal. The associative learning system according to claim 1 or 2.

8. The edge terminal further comprises an identifier generation unit that generates an identifier based on the biometric data. The associative learning system according to claim 1 or 2.

9. The collection unit further collects the identifiers from the plurality of edge terminals. The associative learning system according to claim 8.

10. The central server further comprises a provisioning unit that provides token-based rewards to the edge terminals participating in federative learning. The associative learning system according to claim 1 or 2.

11. If the edge terminal receives the reward from the central server, it will accept the reward if the user's identity can be verified through biometric authentication, and will refuse to accept the reward if the user's identity cannot be verified. The associative learning system according to claim 10.

12. The distribution unit distributes the updated AI model to the multiple edge terminals. The associative learning system according to claim 1 or 2.

13. The update unit generates multiple AI models based on the learning results collected from the multiple edge terminals, and selects one AI model as the updated AI model based on the evaluation results of each AI model. The associative learning system according to claim 1 or 2.

14. The management unit is characterized by being an AI agent that autonomously determines whether to accept or reject a request to participate in the federative learning based on the content of the task, the amount of the reward, and the resource usage status of the terminal. The associative learning system according to claim 2.

15. A federated learning method performed by multiple edge terminals and a central server, The aforementioned edge terminal is By obtaining user-specific biometric data from the user's body parts, The biometric data acquired by the biometric authentication unit is stored, If the user's biometric data is registered on the edge terminal, the AI ​​model's training process is executed. The aforementioned central server is Check if biometric data is registered on the aforementioned edge terminal. The AI ​​model is distributed to the multiple edge terminals. Collect learning results from the aforementioned multiple edge devices, The AI ​​model is updated based on the learning results collected from the aforementioned multiple edge terminals. Associative learning methods.

16. Multiple edge terminals and a central server, The aforementioned edge terminal is By obtaining user-specific biometric data from the user's body parts, The biometric data acquired by the biometric authentication unit is stored, If the user's biometric data is registered on the edge terminal, the AI ​​model's training process is executed. The aforementioned central server is Check if biometric data is registered on the aforementioned edge terminal. The AI ​​model is distributed to the multiple edge terminals. Collect learning results from the aforementioned multiple edge devices, The AI ​​model is updated based on the learning results collected from the aforementioned multiple edge terminals. A program that implements associative learning methods.

Images