Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
  • G06F21/60—Protecting data
  • G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
  • G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
  • G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/0475—Generative networks

Definitions

• the technology disclosed in this specification (hereinafter referred to as the "present disclosure") relates to an information processing device and an information processing method, and a computer program that perform processing to provide data used, for example, in deep learning.
• the collected learning data may contain personal information that should be protected.
• the acquisition of personal information requires the consent of the person concerned.
• the training data required for deep learning is enormous, it is not easy to obtain the consent of the person concerned.
• the purpose of the present disclosure is to provide an information processing device, an information processing method, and a computer program that can be used for deep learning, for example, and perform processing to provide data that protects personal information and ensures fairness. be.
• the present disclosure has been made in consideration of the above problems, and the first aspect thereof is: a data acquisition unit that acquires data; a bias determination unit that detects bias regarding the data set acquired by the data acquisition unit; a learned base model that outputs an instruction regarding the bias or an instruction to generate data that alleviates the bias based on a determination result of the bias determination unit;
• This is an information processing device comprising:
• the instruction regarding the bias is a verbal instruction. Further, the base model outputs an instruction to acquire data for alleviating the bias or to generate data for alleviating the bias, based on a response from a user to an instruction regarding the bias. The base model may acquire data that alleviates the bias or change instructions for generating data related to the bias based on interaction with the user.
• the base model has already learned a predetermined law, and acquires data that alleviates the bias.
• an instruction to generate data that alleviates the bias is output based on the learned law.
• the predetermined law mentioned here is the Personal Information Protection Law.
• a second aspect of the present disclosure is: a data acquisition step of acquiring data; a bias determination step of detecting bias regarding the data set acquired in the data acquisition step; Based on the determination result in the bias determination step, outputting an instruction regarding the bias or outputting an instruction to generate data that alleviates the bias using the learned base model;
• This is an information processing method having the following.
• a third aspect of the present disclosure is: a data acquisition unit that acquires data; a bias determination unit that detects bias regarding the data set acquired by the data acquisition unit; an output unit that outputs an instruction regarding the bias or an instruction to generate data that alleviates the bias, using a learned base model, based on a determination result of the bias determination unit; A computer program written in a computer-readable format to cause a computer to function as a computer program.
• a computer program according to the third aspect of the present disclosure defines a computer program written in a computer readable format so as to implement predetermined processing on a computer.
• a cooperative effect is exerted on the computer, and the same effect as that of the information processing device according to the first aspect of the present disclosure is achieved. effect can be obtained.
• an information processing device an information processing method, and a computer program that perform processing to provide data that can be used for deep learning, protects personal information, and ensures fairness. can.
• FIG. 1 is a diagram showing the functional configuration of an information processing apparatus 100 according to the first embodiment.
• FIG. 2 is a diagram showing an example of a functional configuration for performing processing to protect personal information included in image data.
• FIG. 3 is a diagram showing another functional configuration example for performing processing to protect personal information included in image data.
• FIG. 4 is a flowchart showing a processing procedure for the information processing device 100 to transmit data with protected personal information to a third party data server.
• FIG. 5 is a flowchart showing another processing procedure for the information processing device 100 to transmit data with protected personal information to a third party data server.
• FIG. 6 is a diagram showing the functional configuration of an information processing device 600 according to the second embodiment.
• FIG. 7 is a diagram showing a first operation example of the image generation unit 602.
• FIG. 8 is a diagram showing a second operation example of the image generation unit 602.
• FIG. 9 is a diagram showing a second operation example of the image generation unit 602 (an example using a classifier for each element of personal information that requires consideration).
• FIG. 10 is a diagram showing a second operation example of the image generation unit 602 (an example using a classifier for each element of personal information that requires consideration).
• FIG. 11 is a flowchart showing a processing procedure for the information processing device 600 to generate image data using a base model and transmit it to a third party data server.
• FIG. 12 is a flowchart showing another processing procedure for the information processing device 600 to generate image data using the base model and transmit it to a third party data server.
• FIG. 13 is a diagram showing a configuration example of the information processing device 2000.
• FIG. 14 is a diagram showing a typical infrastructure model after GPT-3.
• FIG. 15 is a diagram showing the interface of each basic model that handles visual and language.
• FIG. 16 is a diagram showing the functional configuration of an information processing device according to the third example.
• FIG. 17 is a diagram showing an example of the operation of the information processing device 1600.
• FIG. 18 is a flowchart showing a processing procedure for collecting data in the information processing apparatus 1600.
• FIG. 19 is a diagram showing another functional configuration of the information processing device according to the third example.
• FIG. 20 is a diagram showing an example of the operation of the information processing device 1900.
• FIG. 21 is a flowchart showing a processing procedure for collecting data in the information processing apparatus 1900.
• this disclosure proposes a technology that can be used for deep learning and provides data that protects personal information and ensures fairness.
• the present disclosure provides for converting data obtained or collected by various means into data with protected personal information and transmitting the data to a third party data server (or converting the data into data in other forms). Proposal regarding technology (distributing).
• the present disclosure proposes a technique for automatically generating unbiased data as a second embodiment and a third embodiment.
• a technique for automatically generating unbiased data By performing learning using large amounts of unbiased data generated using this technology, it is possible to ensure the fairness of machine learning models. Since automatically generated data does not contain personal information, there is no risk of violating the Personal Information Protection Act even if it is used for various purposes such as deep learning.
• the learning target is a machine learning model for performing tasks related to image processing such as image classification
• the learning data for this is basically image data.
• data obtained or collected by various means is converted into data with protected personal information and transmitted to a third party data server. (or distribute data in other formats).
• FIG. 1 shows the functional configuration of an information processing apparatus 100 according to this embodiment.
• the illustrated information processing device 100 includes a data acquisition section 101, an image processing section 102, and a data output section 103.
• the data acquisition unit 101 uses various means to acquire image data for use in deep learning.
• the data acquisition unit 101 (or information processing device 100) is installed on the cloud, and the data acquisition unit 101 collects image data taken by a large number of digital cameras (not shown) as edge devices. You can.
• the data acquisition unit 101 may search or patrol a wide area network such as the Internet to collect image data that can be used for deep learning.
• the image processing unit 102 performs image processing on each image data acquired by the data acquisition unit 101.
• the image processing unit 102 mainly converts the face image included in the image data into a fictitious face by face image conversion, without missing information. Performs processing to convert personal information into image data that is protected.
• the image processing unit 102 may perform other image processing (such as image quality adjustment) on the image data.
• the data output unit 103 outputs the image data after the image processing by the image processing unit 102, that is, the image data with protected personal information, to a device external to the information processing device 100.
• the data output unit 103 outputs image data to a data server 110 that is managed or operated by a third party and stores learning data for deep learning.
• the data output unit 103 may output to other data archives.
• the image processing unit 102 performs processing to protect personal information on the image data acquired by the data acquisition unit 101. Therefore, the data output unit 103 can transmit the processed image data to a third party without worrying about personal information. Furthermore, the third party's data server 110 also transmits the image data provided from the information processing device 100 to a third party (for example, a learning device that performs deep learning) without worrying about personal information. be able to.
• a third party for example, a learning device that performs deep learning
• FIG. 2 shows an example of a functional configuration for performing processing to protect personal information included in image data in the image processing unit 102.
• the image processing unit 102 is configured so that the personal information detection unit 201 and the protection processing unit 202 perform processing to convert personal information in image data into data that does not require protection. There is.
• the personal information detection unit 201 When the personal information detection unit 201 captures the image data acquired by the data acquisition unit 101, it detects personal information included in the image data. Specifically, the personal information detection unit 201 detects a face image included in the image data by face detection processing. The personal information detection unit 201 may use a trained model to detect personal information such as a face image. The protection processing unit 202 then processes the personal information included in the original image data so that the personal information cannot be identified, and converts the personal information into image data that does not require any concern.
• the protection processing unit 202 may perform processing to blindfold, mosaic, or blur the face image in the image data as processing to protect personal information included in the image data.
• attribute information such as the race, gender, and age of the original person is missing, resulting in a decrease in data quality.
• the protection processing unit 202 performs face conversion processing to replace a person image included in the image data with another person's image having the same attribute information as that person.
• the information processing device 100 can supply data that protects personal information while maintaining quality without omitting attribute information, etc., so it can be provided to a third party as good learning data for deep learning. It will be possible to provide the data or even distribute the data via a third party.
• the protection processing unit 202 uses skin color conversion processing to quantify the skin color of the facial area, so that it is no longer necessary to handle racial information, which is sensitive personal information. You can do it like this.
• FIG. 3 shows another functional configuration example for performing processing to protect personal information included in image data in the image processing unit 102.
• the image processing unit 102 uses a personal information detection unit 301, an attribute information detection unit 302, another person image generation unit 303, and a face replacement processing unit 304 to convert the person image in the image data to an appropriate other person. Processing is performed to replace it with an image.
• the personal information detection section 301 is similar to the personal information detection section 201 in FIG.
• the attribute information detection section 302, the different person image generation section 303, and the face replacement processing section 304 correspond to the protection processing section 202 in FIG.
• the personal information detection unit 301 When the personal information detection unit 301 captures the image data acquired by the data acquisition unit 101, it detects personal information included in the image data. Specifically, the personal information detection unit 301 detects a face image included in the image data by face detection processing (same as above).
• the attribute information detection unit 302 detects attribute information of the person image detected by the personal information detection unit 301.
• the attribute information referred to here includes race, gender, age, etc. If necessary, various information such as occupation and place of birth may be included.
• the other person image generation unit 303 Based on the attribute information detected by the attribute information detection unit 302, the other person image generation unit 303 generates another person image having the same attribute information as the person image detected from the original image data by the personal information detection unit 301. Then, the face replacement processing unit 304 replaces the personal information included in the original image data with the image of another person generated by the other person image generation unit 303, thereby converting the personal information into image data that does not require concern about the personal information.
• a process for protecting personal information included in the image data a process of blindfolding, mosaicing, or blurring the face image in the image data may be performed.
• attribute information such as the race, gender, and age of the original person is missing, making the image unsuitable as training data for deep learning.
• the image processing unit 102 shown in FIG. It is configured to perform a process of replacing a human image in image data.
• the information processing device 100 can supply data that protects personal information while maintaining quality without omitting original attribute information, etc., so it can be used as good learning data for deep learning.
• data can be provided to other parties, and the data can be distributed via third parties.
• the different person image generation unit 303 needs to generate a different person image whose authenticity cannot be determined from the original person image. Therefore, in this embodiment, the other person image generation unit 303 uses a generative adversarial network (GAN) to generate another person image.
• GAN is an unsupervised learning method that deepens the learning of input data by competing with a generator and a discriminator, each consisting of a neural network, to generate data that does not exist or to analyze the characteristics of existing data.
• GAN generative adversarial network
• the different person image generation unit 303 uses StyleGAN2, which is a further improvement of StyleGAN that realizes high-resolution image generation using Progressive Growing (see, for example, Non-Patent Document 1), to generate a person image and Images of different people having the same attribute information may be artificially generated.
• the other person image generation unit 303 may use a basic model (described later) to artificially generate another person image having the same attribute information as the person image.
• FIG. 4 shows a processing procedure for converting acquired or collected data into data with protected personal information in the information processing device 100 shown in FIG. 1, and transmitting the data to a third party data server. It is shown in the form of a flowchart. However, in the processing procedure shown in FIG. 4, it is assumed that the image processing unit 102 has the functional configuration shown in FIG.
• the data acquisition unit 101 uses various means to acquire image data for use in deep learning (step S401).
• the data acquisition unit 101 may collect image data taken by digital cameras of many users, or may search or patrol a wide area network such as the Internet to collect image data that can be used for deep learning. You may.
• the personal information detection unit 201 detects a person image as personal information included in the image data (step S402). Specifically, the personal information detection unit 201 detects a face image included in the image data by face detection processing.
• the protection processing unit 202 performs processing to protect the personal information included in the image data (step S403).
• face images in image data may be blindfolded, mosaiced, or blurred, but due to information deterioration, they are no longer suitable as training data for deep learning. Problems such as this arise. Therefore, the protection processing unit 202 performs face conversion processing to replace the person image included in the image data with another person's image having the same attribute information as that person.
• the protection processing unit 202 may use a skin color conversion process that quantifies the skin color of the face, so that it is not necessary to handle racial information that is sensitive personal information.
• the data output unit 103 outputs the image data after the image processing by the image processing unit 102, that is, the image data with protected personal information, to a device external to the information processing device 100, such as a third party data server. Step S404), this process ends.
• FIG. 5 shows other processing steps for converting acquired or collected data into data with protected personal information and transmitting it to a third party data server in the information processing apparatus 100 shown in FIG. It is shown in the form of a flowchart. However, in the processing procedure shown in FIG. 5, it is assumed that the image processing unit 102 has the functional configuration shown in FIG.
• the data acquisition unit 101 uses various means to acquire image data for use in deep learning (step S501).
• the data acquisition unit 101 may collect image data taken by digital cameras of many users, or may search or patrol a wide area network such as the Internet to collect image data that can be used for deep learning. You may.
• the personal information detection unit 301 detects a person image as personal information included in the image data (step S502). Specifically, the personal information detection unit 301 detects a face image included in the image data by face detection processing.
• the attribute information detection unit 302 detects attribute information of the person image detected by the personal information detection unit 301 (step S503).
• the attribute information detection unit 302 detects, for example, race, gender, age, etc. of a person image. If necessary, detect occupation, place of birth, etc.
• the other person image generation unit 303 generates another person image having the same attribute information as the person image detected from the original image data by the personal information detection unit 301 (step S504).
• the other person image generation unit 303 generates another person image having the same attributes using a GAN or a base model.
• the face replacement processing unit 404 performs protection processing by replacing the person image included in the original image data with the other person image generated by the other person image generation unit 303 (step S505).
• the data output unit 103 outputs the image data after the image processing by the image processing unit 102, that is, the image data with protected personal information, to a device external to the information processing device 100, such as a third party data server. Step S506), this process ends.
• Foundation models are machine learning models that are trained at scale (usually through self-supervised learning) on vast amounts of unlabeled data and can adapt to a wide range of downstream tasks at scale.
• Early examples of foundational models were large-scale pre-trained language models such as BERT and GPT-3, and subsequently multiple multimodal foundational models were developed such as DALL-E, Flamingo, and Florence.
• the basic model can learn various data such as text, images, audio, structured data, and 3D signals, and apply it to various tasks such as question and answer, sentiment analysis, information extraction, image capture, object recognition, and command following. Can be done.
• the base model can be constructed using the Transformer architecture, for example.
• the foundation model is state-of-the-art at the time of this filing, and is not limited to the Transformer architecture, but can also be configured using other architectures that already exist or that will be developed in the future.
• FIG. 14 shows a typical infrastructure model after GPT-3 on the time axis.
• the upper half of FIG. 14 is a language model
• the lower half is a visual-language model.
• FIG. 15 summarizes the interfaces of each basic model that handles visual and language.
• "visual vision" here includes both still images and moving images.
• an application is assumed in which an unbiased image is generated as a data set by giving a verbal instruction.
• DALL-E is a 12 billion parameter version of GPT-3 that is trained to generate images from text descriptions using a dataset of text and image pairs. It can create anthropomorphic versions, combine unrelated concepts in plausible ways, render text, apply transformations to existing images, and much more.
• FIG. 6 shows the functional configuration of an information processing device 600 according to this embodiment.
• the illustrated information processing device 600 includes an input section 601, an image generation section 602, and a data output section 603.
• the input unit 601 inputs a language that instructs the generation of image data that will become a dataset for deep learning.
• the input unit 601 may input language data as text via a keyboard, for example, or may input language data via a GUI (Graphical User Interface). Further, the input unit 601 may input voice data uttered by a user and convert the data into text using voice recognition.
• the image generation unit 602 automatically generates a large amount of image data that will become a dataset for deep learning, as instructed in the language input via the input unit 601. As described above, the image generation unit 602 generates image data from the language by applying one of the base models DALL-E, Imagen, or Parti. It is desirable that the base model applied to the image generation unit 602 has already learned predetermined laws, such as the Personal Information Protection Act. This allows the image generation unit 602 to remove bias while automatically generating a large amount of image data in which personal information is protected.
• the data output unit 603 outputs the image data generated by the image generation unit 602 to a device external to the information processing device 600.
• the data output unit 603 outputs image data to a data server 610 that is managed or operated by a third party and stores learning data for deep learning.
• the data output unit 603 may output to other data archives.
• the information processing device 600 by using the basic model, it is possible to generate an image by giving a verbal instruction and construct a fair dataset. Since the data automatically generated by the basic model does not contain any personal information, it can be said that there is no risk of violating the Personal Information Protection Act even if it is used for various purposes such as deep learning.
• FIG. 7 shows a first operation example of the image generation unit 602.
• a language instructing to generate male and female facial images for each family lineage (race) is inputted via the input unit 601.
• each combination of family lineage and gender is treated as one category.
• the image generation unit 602 can construct an unbiased data set by generating the same number of image data in each category using the base model 700.
• FIG. 8 shows a second operation example of the image generation unit 602. In this operation example, it is assumed that a language instructing to generate facial images for each family lineage (race) is input via the input unit 601.
• the image generation unit 602 uses the base model 801 to generate image data for each category. Then, the determination unit 802 at the subsequent stage determines whether the image data output from the base model 801 has a bias between categories. Then, when a bias is detected, the determination unit 802 instructs the base model 801 to supplement the number of data for categories with insufficient data to eliminate the bias. The determination unit 802 may verbally instruct the base model 801 to eliminate the bias.
• the basic model 801 can eliminate the bias by expanding the data of categories for which the number of data is insufficient.
• the determination unit 802 does not directly feed back the bias determination result to the base model 801, but rather displays the bias determination result (for example, "The ratio of African men and women is unbalanced") to the user via a GUI screen or the like. ) may be notified.
• the user inputs the language for expanding the missing category data (for example, "Expand image data of African women") through the input unit 601, and in response, An unbiased data set can be expanded by the model 801 expanding the data as directed.
• the determination unit 802 is realized, for example, by preparing a classifier that classifies the image data generated by the base model 801 for each element (sensitive attribute) of personal information that requires consideration.
• FIG. 9 shows a second example of the operation of the image generation unit 602 when the only element of personal information that requires consideration is "gender."
• the determination unit 802 includes a classifier 901 that classifies image data into men and women.
• the image generation unit 602 generates image data for each category using the base model 801 (same as above).
• the classifier 901 classifies the image data output from the base model 801 into males and females for each family lineage, and determines whether the ratio of males and females is balanced in each family lineage.
• the classifier 901 then instructs the base model 801 to expand the image data of the missing gender for the family in which an unbalanced gender ratio has been detected.
• the basic model 801 can eliminate data bias by expanding the gender data of families for which insufficient data has been indicated.
• FIG. 10 shows a second example of the operation of the image generation unit 602 when three factors, gender, age, and hair color, require consideration.
• the determination unit 802 includes a classifier 1001 that classifies image data as male or female, a classifier 1002 that classifies image data as age, and a classifier 1003 that classifies image data as paper color.
• the image generation unit 602 generates image data for each category using the base model 801 (same as above).
• the classifier 1001 classifies the image data output from the base model 801 into males and females for each family lineage, and determines whether the ratio of males and females is balanced in each family lineage. Then, the classifier 1001 instructs the base model 801 to expand the image data of the missing gender for a family in which an unbalanced gender ratio has been detected.
• the classifier 1002 classifies the image data output from the base model 801 by age for each family lineage, and determines whether the ratios for each age group are balanced in each family lineage. Then, the classifier 1002 instructs the base model 801 to expand the missing age image data for the family in which an unbalanced age ratio has been detected.
• the classifier 1003 classifies the image data output from the base model 801 by hair color for each family lineage, and determines whether the proportions of each hair color are balanced in each family lineage. The classifier 1003 then instructs the base model 801 to expand the image data of the missing hair color for the family in which an unbalanced hair color ratio has been detected.
• the basic model can eliminate data bias by expanding data for elements for which data shortage is indicated by any of the classifiers 1001 to 1003.
• Figure 10 shows a second example of the operation of the image generation unit 602 when three elements, gender, age, and hair color, need to be taken into consideration. However, if other elements, such as eye color, scene, lighting, and angle, need to be taken into consideration, a data set that eliminates bias for all elements can be obtained by adding a classifier for each element.
• FIG. 11 shows a processing procedure for generating an image by giving a verbal instruction and transmitting it to a third party data server by using the basic model in the information processing apparatus 600 shown in FIG. is shown in the form of a flowchart. However, it is assumed that the image processing unit 602 has the functional configuration shown in FIG.
• the input unit 601 inputs a language that instructs the generation of image data that will become a dataset for deep learning (step S1101).
• the input unit 601 may input language data as text via a keyboard, for example, or may input language data via a GUI. Further, the input unit 601 may input voice data uttered by a user and convert the data into text using voice recognition.
• the image generation unit 602 automatically generates a large amount of image data that becomes a dataset for deep learning according to the language input in step S1101 (step S1102).
• the data output unit 603 outputs the image data after the image processing by the image processing unit 102, that is, the image data with protected personal information, to a device external to the information processing device 100, such as a third party data server. Step S1103), this process ends.
• FIG. 12 shows another processing procedure for generating an image in accordance with a verbal instruction and transmitting it to a third party data server by using the basic model in the information processing apparatus 600 shown in FIG. is shown in the form of a flowchart. However, it is assumed that the image processing unit 602 has the functional configuration shown in FIGS. 8 to 10.
• the input unit 601 inputs a language that instructs the generation of image data that will become a dataset for deep learning (step S1201).
• the input unit 601 may input language data as text via a keyboard, for example, or may input language data via a GUI. Further, the input unit 601 may input voice data uttered by a user and convert the data into text using voice recognition.
• the image generation unit 602 automatically generates a large amount of image data that becomes a dataset for deep learning according to the language input in step S1101 (step S1202).
• the determining unit 802 determines whether or not the image data generated in step S1202 is biased for each element of personal information that requires consideration (step S1203). This bias determination process is performed using a classifier prepared for each element of personal information that requires consideration. Each classifier classifies the generated image data into the elements of interest (gender, age, hair color, etc.), and checks whether the ratio of the image data is equal for each element (step S1204).
• step S1204 if it is determined that there is a bias in the image data for any element (Yes in step S1204), the determination unit 802 changes the base model to the base model 801 so as to supplement the missing elements.
• An instruction is output and image data expansion processing is performed (step S1205). After that, the process returns to step S1203, and the presence or absence of bias in the image data is determined again.
• the data output unit 603 outputs the data after the image processing by the image processing unit 102, that is, The image data with protected personal information is output to a device external to the information processing apparatus 100, such as a third party's data server 610 (step S1206), and the process ends.
• the third embodiment of the present disclosure also proposes a technique for constructing a fair data set by data expansion.
• FIG. 16 shows an example of a functional configuration of an information processing apparatus 1600 according to this embodiment.
• the illustrated information processing apparatus 1600 includes a data acquisition section 1601, a bias determination section 1602, and an instruction output section 1603.
• the data acquisition unit 1601 acquires data that satisfies instructions input in a format such as a language.
• the data acquisition unit 1601 mainly acquires image data that becomes a dataset for deep learning.
• the data acquisition unit 1601 outputs the acquired data set to a device external to the information processing device 1600.
• the data acquisition unit 1601 outputs image data to a data server 1610 that is managed or operated by a third party and stores learning data for deep learning.
• the data acquisition unit 1601 may output to other data archives.
• the data acquisition unit 1601 generates image data from the language by applying one of the basic models of DALL-E, Imagen, and Parti. It is desirable that the basic model applied to the data acquisition unit 1601 has learned predetermined laws such as the Personal Information Protection Act. As a result, the data acquisition unit 1601 can be expected to remove bias as much as possible and at the same time automatically generate a large amount of image data in which personal information is protected. However, the data acquisition unit 1601 is not necessarily limited to acquiring data by generating data itself using a base model or the like.
• the data acquisition unit 1601 may be installed on the cloud and collect image data taken by a large number of digital cameras (not shown) as edge devices, or may search a wide area network such as the Internet or It is also possible to collect image data according to instructions while visiting the site.
• the bias determination unit 1602 determines the bias regarding the data set acquired by the data acquisition unit 1601. For example, the bias determination unit 1602 determines whether the image data acquired by the data acquisition unit 1601 has a bias between categories based on a predetermined definition of fairness. The bias determination unit 1602 prepares a classifier that classifies image data for each category to be determined (for example, each element of personal information that requires consideration), and calculates the ratio of the number of image data classified by each classifier. The bias may be determined based on the imbalance. Specifically, the bias determination unit 1602, like the determination unit 802 shown in FIG. A plurality of classifiers may be provided, such as a classifier that classifies by color.
• the bias determination unit 1602 calculates the ratio of the number of pixel data for each category, such as the male-female ratio, age ratio, hair color ratio, and skin color ratio, for the image data acquired by the data acquisition unit 1601, and calculates the ratio of the number of pixel data for each category. Quantitatively evaluate based on the definition of fairness to determine bias.
• the definitions of fairness include Demographic parity, Equalized odds, Equal opportunity, etc.
• Demographic parity is a definition of fairness that the prediction distribution should be the same regardless of the category. For example, when hiring human resources for a company, the definition of fairness is that the proportion of people hired and rejected is the same, regardless of gender or race.
• the definition of fairness in Equalized odds is that the population ratio of True Positive Rate (TPR) and False Positive Rate (FPR) is equal regardless of the sensitive attribute.
• the population ratio of the false positive rate (FPR), which falsely detects that they will recidivate when they do not recidivate, and the true positive rate (TPR), which predicts that they will recidivate but actually recidivate is The standard of fairness is to be the same regardless of "race.”
• Equal Opportunity defines fairness as having the same population ratio of True Positive Rate (TPR) regardless of the sensitive attribute (however, unlike Equalized odds, the ratio of FPR is not considered).
• the ratio of the true positive rate (TPR) of predicting that a loan debtor will not default and actually not defaulting is the same regardless of gender. Become a standard of fairness.
• the instruction output unit 1603 Based on the determination result of the bias determination unit 1602, the instruction output unit 1603 outputs at least one or both of an instruction regarding bias and an instruction to generate data to alleviate the bias.
• the instruction output unit 1603 basically outputs instructions in language.
• the instruction output unit 1603 outputs such an instruction to the data acquisition unit 1601.
• the data acquisition unit 1601 acquires data based on instructions received from the instruction output unit 1603 so as to alleviate data bias. Further, the instruction output unit 1603 may output instructions to a device external to the information processing device 1600.
• the instruction output unit 1603 outputs an instruction regarding bias and an instruction to generate data to alleviate the bias based on the bias determination result using the learned basic model.
• the trained basic model applied to the instruction output unit 1603 uses a language as an output interface, and therefore outputs instructions in language.
• the instruction output unit 1603 can output at least one or both of an instruction regarding bias and an instruction to generate data that alleviates bias in accordance with the learned law.
• the learned basic model applied to the instruction output unit 1603 has learned the Personal Information Protection Act.
• the instruction output unit 1603 provides an instruction regarding bias and an instruction to generate data that alleviates bias so as not to violate the Personal Information Protection Act (or to remove bias regarding elements of personal information that require consideration). At least one or both of these can be output.
• a base model when a base model is also applied to the data acquisition unit 1601, it may be an integrated model with the trained base model applied to the instruction output unit 1603, or may be a mutually independent model.
• FIG. 17 shows an example of the operation of the information processing device 1600.
• an instruction to generate facial images for each family (race) is input to the information processing apparatus 1600 in a language or the like, and unbiased facial images are collected.
• the data acquisition unit 1601 When the data acquisition unit 1601 receives an instruction to acquire facial images for each family lineage (race), it acquires image data for each category generated by the base model based on this instruction.
• the bias determination unit 1602 determines whether the image data acquired by the data acquisition unit 1601 has a bias between categories. If there are multiple elements of personal information that require consideration, such as gender, age, hair color, etc., the bias determination unit 1602 uses a classifier for each element to check the balance of the ratio of each element in the acquired image data. to determine the presence or absence of bias.
• the instruction output unit 1603 outputs an instruction regarding the bias and an instruction to generate data that alleviates the bias based on the bias determination result.
• the instruction output unit 1603 By applying the learned basic model to the instruction output unit 1603, it is possible to output an optimal prompt as an instruction regarding bias or an instruction to generate data that alleviates bias.
• the data acquisition unit 1601 acquires the image data generated by the base model again based on the prompt output from the instruction output unit 1603. In this way, the information processing apparatus 1600 can collect image data in which bias between categories is alleviated for each element of personal information that requires consideration.
• FIG. 18 shows a processing procedure for collecting data in the information processing device 1600 in the form of a flowchart.
• the data acquisition unit 1601 uses the base model to acquire a large amount of image data according to the instruction (step S1802).
• the bias determining unit 1602 determines whether or not there is a bias in the image data acquired in step S1802 for each element of personal information that requires consideration (step S1803).
• This bias determination process is performed using a classifier prepared for each element of personal information that requires consideration.
• Each classifier classifies the generated image data into the respective elements of interest (gender, age, hair color, etc.), and checks whether the ratio of the image data is equal for each element (step S1804).
• the bias determination unit 1602 outputs the determination result to the instruction output unit 1603. Then, the instruction output unit 1603 outputs an instruction regarding bias and an instruction to generate data for alleviating the bias based on the bias determination result (step S1805).
• the learned basic model By applying the learned basic model to the instruction output unit 1603, it is possible to output an optimal prompt as an instruction regarding bias or an instruction to generate data that alleviates bias.
• the data acquisition unit 1601 acquires the image data generated by the base model again based on the prompt output from the instruction output unit 1603. In this way, the information processing apparatus 1600 can collect image data in which bias between categories is alleviated for each element of personal information that requires consideration.
• the information processing device 1600 transfers the collected image data to, for example, a third party.
• the data is output to a device external to the information processing apparatus 1600, such as the user's data server 1610 (step S1806), and this processing ends.
• FIG. 19 shows another functional configuration example of the information processing apparatus 1900 according to this embodiment.
• the illustrated information 1900 includes a data acquisition section 1901, a bias determination section 1902, a presentation section 1903, and an instruction output section 1904.
• the data acquisition unit 1901 acquires image data that satisfies instructions input in a format such as language and serves as a dataset mainly for deep learning.
• the data acquisition unit 1901 generates image data from the language by applying one of the basic models of DALL-E, Imagen, and Parti. It is desirable that the basic model applied to the data acquisition unit 1901 has learned predetermined laws such as the Personal Information Protection Act (same as above).
• the data acquisition unit 1901 outputs the acquired data set to a device external to the information processing apparatus 1900, such as a data server 1910.
• the bias determination unit 1902 determines the bias regarding the dataset acquired by the data acquisition unit 1901. For example, the bias determination unit 1902 determines whether the image data acquired by the data acquisition unit 1901 has a bias between categories based on a predetermined fairness definition (described above). For example, the bias determination unit 1902 prepares a classifier that classifies image data for each category to be determined (for example, each element of personal information that requires consideration), and calculates the number of image data classified by each classifier. Bias is determined by quantitatively evaluating the calculation results of the ratios for each category (male/female ratio, age ratio, hair color ratio, skin color ratio, etc.) based on the definition of fairness (same as above).
• the presentation unit 1903 presents the determination result by the bias determination unit 1902 for the image acquired by the data acquisition unit 1901 to the user.
• the presentation section 1903 includes a data display section 1903A and an interaction processing section 1903B.
• the data display unit 1903A presents data regarding the bias determination results on a screen or the like.
• the interaction processing unit 1903B processes interaction with the user regarding the bias determination results.
• the data display unit 1903A presents the statistical and quantitative bias determination results calculated by the bias determination unit 1902 and the qualitative bias determination results to the user via, for example, a GUI screen.
• Statistical and quantitative bias determination results are based on the ratio of the number of data between categories for each element of personal information that requires consideration (in the case of race, ITA (Individual Typology Angle), which is an evaluation index for skin color, etc.) etc.
• the data display unit 1903A classifies and displays the images acquired by the data acquisition unit 1901 according to a prompt input by the user via the interaction processing unit 1903B as a qualitative bias determination result.
• the data display unit 1903A displays the results of classification based on facial feature amounts specified by the user via the interaction processing unit 1903B.
• the instruction output unit 1904 acquires data for alleviating the bias, or generates at least one or both of an instruction regarding the bias and an instruction to generate data for alleviating the bias, based on a response from the user to the bias determination result. The data is then output to the data acquisition unit 1901. Further, the instruction output unit 1904 may output instructions to a device external to the information processing device 1900.
• the instruction output unit 1904 uses the trained foundation model to output instructions regarding bias and instructions for acquiring data that mitigates bias, based on the user's response.
• the instruction output unit 1904 may also generate instructions for acquiring data using a generative AI such as ChatGPT. It is expected that the user may input a long instruction sentence that is difficult to understand (or difficult to quantify) to the dialogue processing unit 1903B, but the instruction output unit 1904 can replace it with an optimal prompt suitable for input to a computer by using a generative AI.
• the trained foundation model applied to the instruction output unit 1904 is assumed to have learned a specific law (same as above), and outputs instructions for acquiring data that conforms to the specific law.
• a foundation model is also applied to the data acquisition unit 1901, it may be a model integrated with the trained foundation model applied to the instruction output unit 1904, or they may be models independent of each other.
• this information processing device 1900 is equipped with a presentation unit 1903 that presents the bias determination result to the user and receives a response from the user. It is.
• the presentation unit 1903 will be explained in more detail.
• the presentation unit 1903 displays a quantitative score (in the case of race, ITA, which is an evaluation index for skin color, etc.) and a qualitative display for the part where the image acquired by the data acquisition unit 1901 is presented to a human. (The images themselves are classified and displayed according to the prompts) in the data display section 1903A. Such a display method makes it easier for humans to intervene in the bias determination results through the interaction processing unit 1903B.
• the information processing device 1900 determines the racial ratio of the face image acquired by the data acquisition unit 1901 in the bias determination unit 1902, displays the result on the data display unit 1903A, and further determines the racial ratio through the user or the interaction processing unit 1903B. It is configured so that it can give instructions for adjustment. Not only racial ratios, but also ratios related to other sensitive attributes such as male-female ratios and age ratios, it is not necessarily a good idea to have the same number as the actual world ratio, and this may lead to bias. In such a complex social structure, the information processing device 1900 can eliminate distortions in social factors through human adjustment on a case-by-case basis.
• the information processing device 1900 allows humans to ask and answer questions to a basic model, which is a large-scale language model, in an interactive format, so that the information processing device 1900 can analyze information for each sensitive attribute such as race, gender, age, hair color, etc. This is a system that allows you to know what ratio of acquired image data will create a good AI.
• FIG. 20 shows an example of the operation of the information processing device 1900.
• an instruction to generate facial images for each family lineage (race) is input to the information processing apparatus 1900 in a language or the like, and unbiased facial images are collected.
• the data acquisition unit 1901 When the data acquisition unit 1901 receives an instruction to acquire facial images for each family lineage (race), it acquires image data for each category generated by the base model based on this instruction.
• the bias determination unit 1902 determines whether the image data acquired by the data acquisition unit 1901 has a bias between categories. If there are multiple elements of personal information that require consideration, such as gender, age, hair color, etc., the bias determination unit 1902 uses a classifier for each element to check the balance of the ratio of each element in the acquired image data. to determine the presence or absence of bias.
• the data display unit 1903A presents the statistical and quantitative bias determination results by the bias determination unit 1902 and the qualitative bias determination results to the user.
• the interaction processing unit 1903B can respond to and interact with the bias determination results through interaction with the user.
• the instruction output unit 1904 uses the learned base model to output an instruction regarding bias and an instruction to generate data to alleviate the bias based on the response from the user.
• the learned basic model By applying the learned basic model to the instruction output unit 1904, it is possible to output an optimal prompt as an instruction regarding bias or an instruction to generate data that alleviates bias.
• the data acquisition unit 1901 acquires the image data generated by the base model again based on the prompt output from the instruction output unit 1904.
• the information processing apparatus 1900 can collect image data in which bias between categories is alleviated for each element of personal information that requires consideration.
• the ratio of the number of data classified by each sensitive attribute such as race, gender, age, hair color, etc. from the huge amount of image data acquired by the data acquisition unit 1901 is compared with the actual world ratio and statistical/mathematical ratio.
• the general idea of fairness is to align them so that they are equal.
• Such simple bias mitigation methods may lead to bias and may fail to resolve ethical and social issues. Therefore, humans should choose how to adjust the ratio of the number of data depending on the purpose and application.
• a human interacts with the statistical and quantitative bias determination results and qualitative bias determination results presented by the data display unit 1903A to display the basic model in an interactive manner through the interaction processing unit 1903B.
• the instruction output unit 1904 generates an optimal prompt for the base model for correcting bias from the interaction with a human through the interaction processing unit 1903B, and provides the generated prompt to the data acquisition unit 1901.
• the human may not be able to determine what kind of bias exists in a huge facial image data set, such as hundreds of millions of images. In such a case, the human repeatedly asks the base model questions regarding the bias via the interaction processing unit 1903B, and the base model summarizes or corrects the data set so that the human can understand the bias.
• the data acquisition unit 1901 can acquire, for example, an image whose background is not linked to a race or an image with diverse facial features in response to an instruction to acquire a facial image of a specific race. become.
• the information processing device 1900 has a function of collecting data sets so that the data sets are equivalent to the ratios of the real world and statistical/mathematical ratios, and the ratio of the number of data through interaction with humans via a basic model. It is equipped with the ability to adjust the system, and can remove residual bias caused by complex social structures depending on the purpose.
• FIG. 21 shows a processing procedure for collecting data in the information processing device 1900 in the form of a flowchart.
• the data acquisition unit 1601 uses the base model to acquire a large amount of image data according to the instruction (step S2102).
• the bias determination unit 1902 determines the presence or absence of bias in the image data acquired in step S2102 for each element of personal information that requires consideration (step S2103). This bias determination process is performed using a classifier prepared for each element of personal information that requires consideration. Each classifier classifies the generated image data into the respective elements of interest (gender, age, hair color, etc.), and checks whether the ratio of the image data is equal for each element (step S2104).
• the bias determination unit 1602 outputs the determination result to the presentation unit 2103. Then, in the presentation unit 1903, the data display unit 1903A presents the statistical and quantitative bias determination results and the qualitative bias determination results calculated by the bias determination unit 1902 to the user (step S2105). Furthermore, the interaction processing unit 1903B processes interaction with the user regarding the bias determination results.
• the instruction output unit 1904 outputs an instruction regarding bias and an instruction to generate data that mitigates bias based on the response from the user (step S2106).
• the instruction output unit 1904 By applying the trained base model to the instruction output unit 1904, it is possible to output an optimal prompt as an instruction regarding bias or an instruction to generate data that mitigates bias.
• the data acquisition unit 1901 acquires the image data generated by the base model again based on the prompt output from the instruction output unit 1603. In this way, the information processing apparatus 1900 can collect image data in which bias between categories is appropriately alleviated for each element of personal information that requires consideration.
• 1900 transfers the collected image data to, for example, a third party.
• the information is output to a device external to the information processing device 1900, such as the data server 1610 (step S2107), and this processing ends.
• FIG. 13 shows a specific example of the hardware configuration of the information processing device 2000.
• the information processing device 2000 shown in FIG. 10 is composed of, for example, a PC.
• the information processing device 2000 can operate as the information processing device shown in FIGS. 1, 6, and 16.
• the information processing device 2000 shown in FIG. 13 includes a CPU 2001, a ROM (Read Only Memory) 2002, a RAM (Random Access Memory) 2003, a host bus 2004, a bridge 2005, an expansion bus 2006, an interface unit 2007, It includes an input section 2008, an output section 2009, a storage section 2010, a drive 2011, and a communication section 2013.
• the CPU 2001 functions as an arithmetic processing device and a control device, and controls the overall operation of the information processing device 2000 according to various programs.
• the ROM 2002 non-volatilely stores programs used by the CPU 2001 (such as a basic input/output system) and calculation parameters.
• the RAM 2003 is used to load programs used in the execution of the CPU 2001, and to temporarily store parameters such as work data that change as appropriate during program execution. Programs loaded into the RAM 2003 and executed by the CPU 2001 include, for example, various application programs and an operating system (OS).
• OS operating system
• the CPU 2001, ROM 2002, and RAM 2003 are interconnected by a host bus 2004 composed of a CPU bus and the like. Through the cooperative operation of the ROM 2002 and the RAM 2003, the CPU 2001 can execute various application programs in an execution environment provided by the OS to realize various functions and services.
• the OS is, for example, Microsoft Windows or Unix.
• application programs include applications that perform processing to protect personal information included in image data, applications that perform processing to generate images of another person to protect personal information, and applications that perform processing to protect personal information contained in image data. It is assumed that the application includes an application that performs processing to automatically generate image data based on.
• the host bus 2004 is connected to an expansion bus 2006 via a bridge 2005.
• the expansion bus 2006 is, for example, a PCI (Peripheral Component Interconnect) bus or PCI Express, and the bridge 2005 is based on the PCI standard.
• PCI Peripheral Component Interconnect
• the bridge 2005 is based on the PCI standard.
• the interface unit 2007 connects peripheral devices such as an input unit 2008, an output unit 2009, a storage unit 2010, a drive 2011, and a communication unit 2013 in accordance with the standard of the expansion bus 2006.
• peripheral devices such as an input unit 2008, an output unit 2009, a storage unit 2010, a drive 2011, and a communication unit 2013 in accordance with the standard of the expansion bus 2006.
• the information processing apparatus 2000 may further include peripheral devices that are not shown.
• the peripheral devices may be built into the main body of the information processing device 2000, or some peripheral devices may be externally connected to the main body of the information processing device 2000.
• the input unit 2008 includes an input control circuit that generates an input signal based on input from the user and outputs it to the CPU 2001.
• the input unit 2008 can input language data using a keyboard, mouse, touch panel, microphone, etc. may include other devices.
• the output unit 2009 includes, for example, a display device such as a liquid crystal display (LCD) device, an organic EL (electro-luminescence) display device, and an LED (light emitting diode).
• the storage unit 2010 stores files such as programs (applications, OS, etc.) executed by the CPU 2001 and various data.
• the storage unit 2010 may function, for example, as the data accumulation unit 801 and accumulate a large amount of data to be subjected to multivariate analysis.
• the storage unit 2010 is configured with a large-capacity storage device such as an SSD (Solid State Drive) or an HDD (Hard Disk Drive), but may also include an external storage device.
• the removable storage medium 2012 is a cartridge-type storage medium such as a microSD card, for example.
• the drive 2011 performs read and write operations on the loaded removable storage medium 2013.
• the drive 2011 outputs data read from the removable recording medium 2012 to the RAM 2003 or the storage unit 2010, or writes data on the RAM 2003 or the storage unit 2010 to the removable recording medium 2012.
• the communication unit 2013 is a device that performs wireless communication such as Wi-Fi (registered trademark), Bluetooth (registered trademark), and cellular communication networks such as 4G and 5G.
• the communication unit 2013 also includes terminals such as USB (Universal Serial Bus) and HDMI (registered trademark) (High-Definition Multimedia Interface), and has the function of performing data communication with USB devices such as scanners and printers, displays, etc. You may also have more.
• the present disclosure it is possible to provide data that can be distributed without worrying about personal information.
• the present disclosure can be applied primarily to collecting data used for deep learning, the gist of the present disclosure is not limited thereto.
• the data provided by this disclosure can be used for various applications other than deep learning.
• a data acquisition unit that acquires data
• a bias determination unit that detects bias regarding the data set acquired by the data acquisition unit
• a learned base model that outputs an instruction regarding the bias or an instruction to generate data that alleviates the bias based on a determination result of the bias determination unit
• the base model acquires data that alleviates the bias, or outputs an instruction to generate data that alleviates the bias, based on a response from the user to an instruction regarding the bias.
• the base model acquires data that alleviates the bias or changes instructions for generating data related to the bias based on interaction with the user;
• the information processing device according to (3) above.
• the base model has already learned a predetermined law, and the instruction to acquire data that alleviates the bias or generate data that alleviates the bias is output based on the learned law. be done, The information processing device according to (1) above.
• the data acquisition unit acquires image data
• the base model outputs an instruction regarding the skin color of the face image included in the image data, or outputs an instruction to generate image data in which the skin color of the face image included in the face image data is converted.
• the data acquisition unit acquires the language;
• the base model generates image data specified in language, or outputs an instruction to generate image data specified in language,
• the information processing device according to (1) above.
• the bias determination unit detects a bias in the image data generated by the base model.
• the bias determination unit detects bias for each element of personal information that requires consideration; The information processing device according to (9) above.
• the base model further performs a process of supplementing image data based on the bias determination result by the bias determination unit.
• the information processing device according to (9) above.
• a data acquisition unit that acquires data
• a bias determination unit that detects bias regarding the data set acquired by the data acquisition unit
• an output unit that outputs an instruction regarding the bias or an instruction to generate data that alleviates the bias, using a learned base model, based on a determination result of the bias determination unit
• a computer program written in computer-readable form to cause a computer to function as a computer program.
• An information processing device comprising:
• the input unit inputs image data
• the processing unit executes processing to protect personal information included in the image data.
• the information processing device according to (21) above.
• the processing unit executes a process of replacing a face image included in the image data with an image of another person;
• the information processing device according to (22) above.
• the processing unit generates an image of another person using the learned model.
• the information processing device according to (23) above.
• the processing unit executes a process of converting the skin color of the face image included in the image data.
• the information processing device according to any one of (22) to (24) above.
• the input unit inputs a language, the processing unit generates image data instructed in language;
• the information processing device according to (21) above.
• the processing unit generates image data using the base model;
• the information processing device according to (26) above.
• the processing unit detects a bias in the image data generated by the base model.
• the processing unit detects bias for each element of personal information that requires consideration;
• the information processing device according to (28) above.
• the processing unit further performs a process of supplementing the image data based on the bias detection result.
• the information processing device according to any one of (28) or (29) above.
• an input unit a processing unit that performs processing to generate an image in which personal information is protected based on the data input by the input unit;
• DESCRIPTION OF SYMBOLS 100... Information processing device, 101... Data acquisition part, 102... Image processing part 103... Data output part, 110... Data server 201... Personal information detection part, 202... Protection processing part 301... Personal information detection part, 302... Attribute information Detection unit 303...Another person image generation unit, 304...Face replacement processing unit 600...Information processing device, 601...Input unit, 602...Image generation unit 603...Data output unit, 610...Data server 700...Foundation model, 801...Foundation model , 802... Judgment section 901... Classifier, 1001-1003... Classifier 1600... Information processing device, 1601... Data acquisition section 1602... Bias judgment section, 1603...
• Instruction output section 1900 ... Information processing device, 1901... Data acquisition section 1902 ...Bias determination section, 1903... Presentation section 1903A... Data display section, 1903B... Dialogue processing section 1904... Instruction output section 2000... Information processing device, 2001... CPU, 2002... ROM 2003...RAM, 2004...Host bus, 2005...Bridge 2006...Expansion bus, 2007...Interface section 2008...Input section, 2009...Output section, 2010...Storage section 2011...Drive, 2012...Removable recording medium 2013...Communication section

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Health & Medical Sciences (AREA)
• General Health & Medical Sciences (AREA)
• Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Software Systems (AREA)
• Computational Linguistics (AREA)
• Bioethics (AREA)
• Artificial Intelligence (AREA)
• Computing Systems (AREA)
• Life Sciences & Earth Sciences (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• Molecular Biology (AREA)
• Biomedical Technology (AREA)
• Mathematical Physics (AREA)
• Biophysics (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Medical Informatics (AREA)
• Databases & Information Systems (AREA)
• Computer Hardware Design (AREA)
• Computer Security & Cryptography (AREA)
• Image Analysis (AREA)

Priority Applications (2)

Applications Claiming Priority (2)

Publications (1)

Family

ID=90275223

Family Applications (1)

Country Status (3)

Citations (4)

• 2023
  • 2023-09-13 WO PCT/JP2023/033316 patent/WO2024058202A1/ja not_active Ceased
  • 2023-09-13 EP EP23865551.8A patent/EP4589460A4/en active Pending
  • 2023-09-13 JP JP2024547337A patent/JPWO2024058202A1/ja active Pending

Patent Citations (4)

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events