WO2024084827A1 - 情報処理装置、情報処理方法、およびプログラム - Google Patents
情報処理装置、情報処理方法、およびプログラム Download PDFInfo
- Publication number
- WO2024084827A1 WO2024084827A1 PCT/JP2023/031384 JP2023031384W WO2024084827A1 WO 2024084827 A1 WO2024084827 A1 WO 2024084827A1 JP 2023031384 W JP2023031384 W JP 2023031384W WO 2024084827 A1 WO2024084827 A1 WO 2024084827A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- learning
- inference model
- information processing
- processing device
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/08—Learning methods
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/08—Learning methods
- G06N3/098—Distributed learning, e.g. federated learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
Definitions
- This disclosure relates to an information processing device, an information processing method, and a program.
- Patent Document 1 discloses a technique that uses a federated learning method in training a machine learning model for processing medical data.
- an information processing device includes a learning unit that learns an inference model by federated learning, and an acquisition unit that acquires, from a plurality of terminals, privacy protection data, which is data that has been subjected to privacy protection processing on local data obtained by each of the plurality of terminals, wherein the learning unit learns the inference model based on the privacy protection data and distributes to the plurality of terminals information related to the inference model, including hyperparameters of the inference model that are set based on the results of the learning, and the acquisition unit acquires from the plurality of terminals update information for the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters, using the local data as learning data, and the learning unit updates the inference model using the update information.
- the present disclosure also provides an information processing method executed by a computer, including: a processor learning an inference model by federated learning; acquiring from a plurality of terminals privacy protection data, which is data obtained by performing privacy protection processing on local data obtained by each of the plurality of terminals; learning the inference model based on the privacy protection data; distributing to the plurality of terminals information related to the inference model, including hyperparameters of the inference model that are set based on the results of the learning; acquiring from the plurality of terminals update information for the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters with the local data as learning data; and updating the inference model using the update information.
- a program for making a computer function as an information processing device that includes a learning unit that learns an inference model by federated learning, and an acquisition unit that acquires from a plurality of terminals privacy protection data, which is data that has been subjected to privacy protection processing on local data obtained by each of the plurality of terminals, the learning unit learns the inference model based on the privacy protection data, and distributes to the plurality of terminals information related to the inference model including hyperparameters of the inference model that are set based on the results of the learning, the acquisition unit acquires from the plurality of terminals update information for the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters with the local data as learning data, and the learning unit updates the inference model using the update information.
- FIG. 1 is an explanatory diagram illustrating a configuration example of an information processing system according to an embodiment of the present disclosure.
- 2 is a block diagram showing a configuration example of a terminal 10 according to the present embodiment.
- FIG. 2 is a block diagram showing an example of the configuration of an information processing device 20 according to the present embodiment.
- FIG. FIG. 2 is a sequence diagram for explaining a first operation example of the information processing system according to the present embodiment.
- FIG. 5 is a sequence diagram illustrating the flow of processing in a subroutine of S101 in the sequence diagram shown in FIG. 4.
- 11 is a diagram for explaining an example of a process for generating privacy protection data by a data processing unit 130.
- FIG. 5 is a sequence diagram illustrating the flow of processing in a subroutine of S107 in the sequence diagram shown in FIG. 4.
- FIG. 5 is a sequence diagram illustrating the flow of processing in a subroutine of S109 in the sequence diagram shown in FIG. 4.
- FIG. 11 is a sequence diagram for explaining a second operation example of the information processing system according to the embodiment.
- 10 is a sequence diagram for explaining the flow of subroutine processes of S131 and S105 in the sequence diagram shown in FIG. 9.
- 11 is a diagram for explaining another example of the process of generating privacy protection data by the data processing unit 130.
- FIG. FIG. 11 is a sequence diagram for explaining a third operation example of the information processing system according to the embodiment.
- FIG. 13 is a sequence diagram for explaining the flow of processes in the subroutines of S141 and S103 in the sequence diagram shown in FIG. 12.
- FIG. 11 is a diagram for explaining aggregation of privacy protection data in a third operation example of the information processing system according to the present embodiment.
- FIG. 13 is a conceptual diagram for explaining detection of an abnormal value in a third operation example of the present embodiment.
- FIG. 13 is another conceptual diagram for explaining the detection of an abnormal value in the third operation example of the present embodiment.
- FIG. 11 is a diagram for explaining generation of privacy protection data in a modified example of the present embodiment.
- FIG. 11 is a sequence diagram for explaining an example of operation of a modified example of the information processing system according to the present embodiment.
- 20 is a sequence diagram illustrating the flow of subroutine processes of S151 and S155 in the sequence diagram shown in FIG. 18.
- FIG. 9 is a block diagram showing a hardware configuration 90 according to an embodiment of the present disclosure.
- multiple components having substantially the same functional configuration may be distinguished by adding different numbers or letters after the same reference numeral. However, if there is no particular need to distinguish between multiple components having substantially the same functional configuration, each of the multiple components will be given only the same reference numeral.
- Inference models make it possible to perform various inferences with high accuracy based on unknown data. For this reason, the creation and use of inference models is widely used in a variety of fields.
- Federated Learning there is a method called Federated Learning that can protect the privacy of data when learning using data collected from multiple devices.
- a server issues instructions to each of multiple devices to learn a machine learning model, and the model is learned using data acquired by each of the multiple terminals.
- the learning results are collected on the server side, and the model held on the server side is updated based on the learning results.
- an inference model can be trained without exposing the data held by multiple devices and actually used for training to external devices.
- building a machine learning system can be broadly divided into three stages: preparing training data, optimizing the model (learning), and operating the model (inference).
- learning instructions using each setting value are sent from the server to multiple terminals.
- Each of the multiple terminals uses the received hyperparameter setting values to learn using data acquired at the multiple terminals as learning data.
- the server aggregates the learning results from the multiple terminals and updates the model parameters based on the aggregated results.
- the server redistributes the updated model parameters to the multiple terminals. The above process is repeated until learning converges.
- the above series of processes is repeated, and when learning using one trial pattern of hyperparameter setting values converges, the above series of processes is similarly repeated for the trial patterns of the remaining hyperparameter setting values.
- the more trial patterns are tried as the setting values of the hyperparameters the more optimal the hyperparameters can be set.
- an increase in the number of trial patterns for the hyperparameters has the disadvantage of increasing the time and processing load required for learning to select the hyperparameters.
- the inference accuracy of a machine learning model is greatly affected by the quality of the training data. For this reason, it is desirable to detect and correct outliers in the data and invalid data, such as incorrect values entered by malicious users, during the preparation stage of the training data.
- the tendency of the data acquired by the terminal may change due to factors such as changes in the social environment.
- a discrepancy may occur between the training data at the time the model was trained and the data acquired in the inference stage. In such a case, it is desirable to retrain the model.
- the information processing device 20 includes a communication unit 250 that acquires privacy protection data, which is data that has been subjected to privacy protection processing on local data, based on data that is actually used for learning and is acquired by each of the multiple terminals 10 (hereinafter, local data).
- the communication unit 250 is an example of an acquisition unit of the information processing device 20.
- the information processing device 20 includes a learning unit 230 that learns an inference model based on the collected privacy-preserving data. Furthermore, the learning unit 230 distributes the inference model, which includes information on hyperparameters set based on the results of the learning, to the terminal 10.
- the learning unit 230 of the information processing device 20 updates the inference model using model update information obtained by each of the multiple terminals learning the distributed inference model using local data as learning data.
- the information processing device 20 performs learning based on privacy-protected data, which is data that has been subjected to privacy protection processing on local data, which is data actually used for learning.
- privacy-protected data which is data that has been subjected to privacy protection processing
- local data which is data actually used for learning.
- learning for trialing hyperparameters is performed based on data that is closer to local data while protecting privacy information. Therefore, it is expected that the inference accuracy of the inference model will improve.
- each of the multiple terminals 10 learns an inference model using hyperparameters distributed from the information processing device 20. Therefore, the terminal 10 does not need to perform learning for trialing hyperparameters. Therefore, the processing load of learning on the terminal 10 side is reduced.
- the information processing device 20 performs a compilation process of the privacy protection data. Based on the result of the compilation process, the information processing device 20 distributes trend information indicating the statistical data trends of the local data to multiple terminals 10.
- the information processing device 20 monitors changes in the distribution trends of local data based on the results of compiling privacy-preserving data.
- the information processing device 20 detects that the distribution trend of the local data has changed, it re-learns the inference model.
- multiple terminals 10 can detect abnormal values contained in the learning data based on trend information that indicates the distribution trend of local data.
- the inference model can be retrained in response to changes in trends in the local data even after the inference model has moved to the operational stage.
- FIG. 1 is an explanatory diagram illustrating a configuration example of an information processing system according to an embodiment of the present disclosure.
- an information processing system includes multiple terminals 10 and an information processing device 20.
- Each terminal 10 and information processing device 20 are connected to each other via a network 30 so that they can communicate with each other.
- FIG. 1 illustrates an example in which the information processing system according to this embodiment includes three terminals 10, namely, terminal 10A, terminal 10B, and terminal 10C
- the number of terminals 10 according to this embodiment is not particularly limited.
- the information processing system according to this embodiment may include two terminals 10.
- the information processing system according to this embodiment may include three or more terminals 10.
- FIG. 1 illustrates an example in which the terminal 10 is realized by a smartphone
- the terminal 10 may be realized by other information processing terminals.
- the terminal 10 may be realized by a PC (Personal Computer), a tablet terminal, a game console, a wearable device, etc.
- PC Personal Computer
- FIG. 1 illustrates an example in which the terminal 10 is realized by a smartphone
- the terminal 10 may be realized by other information processing terminals.
- the terminal 10 may be realized by a PC (Personal Computer), a tablet terminal, a game console, a wearable device, etc.
- PC Personal Computer
- Terminal 10 uses the acquired local data as learning data to learn an inference model distributed from the information processing device 20.
- the terminal 10 transmits update information of the inference model to the information processing device 20 based on the learning results.
- the update information may be, for example, updated parameters obtained as a result of learning. Or, it may be difference information between parameters before and after the update.
- the terminal 10 performs privacy protection processing on the acquired local data to generate privacy protection data.
- the terminal 10 transmits the generated privacy protection data to the information processing device 20.
- the terminal 10 may receive trend information indicating data trends of local data distributed from the information processing device 20.
- the terminal 10 may detect fraudulent data, such as abnormal values or illegal values, from the newly acquired local data based on the trend information.
- the terminal 10 may generate privacy protection information based on the local data excluding the detected fraudulent data.
- the information processing device 20 of this embodiment distributes an inference model generated based on privacy protection information acquired from multiple terminals 10 to the terminals 10.
- the information processing device 20 receives update information for the inference model from multiple terminals 10, and updates the inference model based on the update information.
- the information processing device 20 distributes update information (updated model, hyperparameters, etc.) from the information processing device 20 relating to the updated inference model to multiple terminals 10.
- Network 30 The network 30 mediates communication between the terminal 10 and the information processing device 20 .
- Example of configuration of terminal 10>> Next, a configuration example of the terminal 10 according to the present embodiment will be described in detail.
- FIG. 2 is a block diagram showing an example of the configuration of a terminal 10 according to this embodiment.
- the terminal 10 may include an acquisition unit 110, a data processing unit 130, a learning unit 150, and a communication unit 170.
- the acquisition unit 110 collects various types of data.
- the data collected by the acquisition unit 110 may be used as learning data for an inference model on the terminal 10.
- the acquisition unit 110 may be equipped with various sensors for collecting sensor information that can be used as one element of the learning data for the inference model on the terminal 10.
- the acquisition unit 110 may acquire information such as the communication speed or bandwidth related to wireless communication between the communication unit 170 and another device.
- the acquisition unit 110 may acquire various data, such as sound data, character data, or image data, to be used as learning data, from an external device, such as an external storage device.
- the image data may be, for example, a medical image.
- the data collected by the acquisition unit 110 and used as learning data for the inference model on the terminal 10 is referred to as local data.
- the data processing unit 130 has a function of generating privacy protection data based on the local data acquired by the acquisition unit 110 .
- Privacy protection processing refers to processing that makes it difficult to identify and restore confidential elements, such as privacy information, contained in the local data.
- the privacy protection data generated by the data processing unit 130 can be in several data formats.
- the privacy-preserving data generated by the data processing unit 130 may be data obtained by performing a data conversion process on the local data that satisfies differential privacy.
- the data conversion process may be a process of assigning a random number of a predetermined strength to each of the elements included in the local data.
- a Laplace mechanism or a Gaussian mechanism may be used as a data conversion process that satisfies differential privacy.
- the privacy protection data may be data generated by performing a data conversion process on the local data by the data processing unit 130 to reduce the dimension of the data.
- the data processing unit 130 may use an auto-encoder algorithm to reduce the dimension of the local data.
- the privacy protection data generated by the data processing unit 130 may be data generated by performing an anonymization process on local data.
- statistics of local data may be calculated by the data processing unit 130.
- the privacy-preserving data may be statistical data generated by performing a data conversion process that satisfies differential privacy on the statistics of the local data calculated by the data processing unit 130.
- the privacy protection data may be data generated by the data processing unit 130 performing encryption processing on the calculated statistics of the local data that meets the requirements for secure computation.
- the data processing unit 130 may obtain trend information of data across multiple terminals 10 from the information processing device 20.
- the data processing unit 130 may detect abnormal values contained in the local data newly acquired by the acquisition unit 110 based on the trend information.
- the data processing unit 130 may perform a process of correcting or excluding data that is considered to be an abnormal value from the local data.
- the data processing unit 130 may generate privacy protection data based on the local data from which the data that is considered to be an abnormal value has been corrected or excluded.
- the learning unit 150 learns an inference model distributed from the information processing device 20 using the local data acquired by the acquisition unit 110 as learning data.
- the learning unit 150 outputs update information for the inference model based on the results of the above learning.
- the update information may be, for example, difference information for the parameters of the inference model before and after learning.
- the communication unit 170 communicates with the information processing device 20 via the network 30 .
- the communication unit 170 transmits the privacy protection data generated by the data processing unit 130 to the information processing device 20.
- the communication unit 170 also transmits update information for the inference model output as a result of learning by the learning unit 150 to the information processing device 20.
- the communication unit 170 also receives the inference model and update information about the inference model from the information processing device 20.
- the terminal 10 may further include, for example, an input unit that accepts information input by a user, a display unit that displays various types of information, etc.
- the configuration of the terminal 10 according to this embodiment can be flexibly modified according to the specifications and operation.
- FIG. 3 is a block diagram showing an example of the configuration of an information processing device 20 according to this embodiment.
- the information processing device 20 may include a generation unit 210, a learning unit 230, and a communication unit 250.
- the generation unit 210 performs a compilation process of the privacy protection data acquired from the terminal 10 .
- the generation unit 210 may use a secret calculation technique to perform the aggregation process on the privacy protection data in its encrypted state, without decrypting the privacy protection data.
- the generation unit 210 also estimates the distribution of local data based on the results of the privacy protection data aggregation process.
- the generating unit 210 may generate synthetic data based on the distribution of the estimated local data.
- synthetic data refers to data that is pseudo-sampled based on the distribution of the estimated local data.
- the generating unit 210 may also generate trend information indicating the estimated results of the statistical data trends of the local data based on the results of the aggregation process of the privacy protection data.
- the trend information may be generated by the generation unit 210 periodically collecting privacy protection data from multiple terminals 10 and calculating the difference in the results of the aggregation process of the privacy protection data for each fixed period of time.
- the data set of privacy protection data collected from the terminal 10 and stored in the information processing device 20 may itself be used as trend information.
- the generation unit 210 may cause the communication unit 250 to distribute the trend information to each of the multiple terminals 10.
- the generation unit 210 may monitor changes in the distribution trends of local data based on the results of the privacy protection data aggregation process.
- the learning unit 230 learns an inference model through associative learning.
- the inference model may be a Convolutional Neural Network (CNN).
- CNN Convolutional Neural Network
- the information processing device 20 may use the inference model to perform image recognition of a medical image acquired by the terminal 10, and diagnose a disease inferred from the medical image based on the recognition result.
- the inference model may be a Long Short-Term Memory (LSTM) model capable of handling time series data.
- the information processing device 20 may use the inference model to predict the quality of wireless communication based on wireless communication quality information, such as communication speed, acquired by the terminal 10.
- wireless communication quality information such as communication speed
- the learning unit 230 performs learning to select hyperparameters for the inference model using the privacy-protected data acquired from the terminal 10 or synthetic data generated based on the privacy-protected data as learning data.
- the learning unit 230 may learn the inference model according to the trial patterns of combinations of candidate values of hyperparameters determined by the administrator of the inference model, so that all trial patterns are covered.
- the hyperparameters of the inference model may be set by an administrator of the inference model based on the results of learning by the learning unit 230.
- the initial parameters of the inference model may be obtained based on the results of learning for hyperparameter trials by the learning unit 230.
- the learning unit 230 distributes the above inference model, including the set hyperparameter information, to multiple terminals 10.
- the learning unit 230 updates the inference model based on update information for the inference model acquired from the terminal 10.
- the learning unit 230 may use a federated learning technique to re-learn the inference model when the generating unit 210 detects that the distribution trend of the local data has changed.
- the communication unit 250 communicates with a plurality of terminals 10 via the network 30.
- the communication unit 250 is an example of an acquisition unit of the information processing device 20.
- the communication unit 250 transmits, for example, the inference model including the hyperparameter information set based on the results of learning by the learning unit 230, and update information for the inference model from the information processing device 20 to the terminal 10.
- the communication unit 250 also receives update information for the inference model from multiple terminals 10.
- the information processing device 20 may further include, for example, an input unit that accepts information input by a user, a display unit that displays various types of information, etc.
- the configuration of the information processing device 20 according to this embodiment can be flexibly modified according to the specifications and operation.
- first operation example an example will be described in which privacy-preserving data generated by the data processing unit 130 of the terminal 10 is data generated by performing a data conversion process that satisfies differential privacy and/or a process that reduces the dimension of the data on local data.
- FIG. 4 is a sequence diagram for explaining a first operation example of the information processing system according to this embodiment.
- the sequence diagram shown in FIG. 4 shows an overview of the processing flow in the first operation example.
- the information processing device 20 aggregates privacy protection data from each of the multiple terminals 10 (S101).
- the information processing device 20 learns the hyperparameters and an inference model according to the trial pattern for initial parameter search based on the aggregated privacy-preserving data (S107).
- the information processing device 20 and the multiple terminals 10 learn the inference model through federated learning (S109).
- the multiple terminals 10 perform inference using the inference model that has been learned on each terminal 10 (S111).
- the information processing device 20 and the terminal 10 monitor changes in the data distribution trend of the local data (S113).
- the information processing device 20 and the terminal 10 re-learn the model through associative learning (S115).
- FIG. 5 is a sequence diagram for explaining the flow of processing in the subroutine of S101 in the sequence diagram shown in FIG.
- the data processing unit 130 of each of the multiple terminals 10 performs privacy protection processing on local data by a method such as data conversion processing that satisfies differential privacy or dimensionality reduction processing (S201).
- FIG. 6 is a diagram for explaining an example of the process of generating privacy protection data by the data processing unit 130.
- the local data LD1 shown in FIG. 6 is an example of local data collected by the acquisition unit 110 of the terminal 10A.
- the post-conversion processing data DA1 shown in FIG. 6 is an example of privacy protection data generated by the data processing unit 130 performing the above-described data conversion processing based on the local data LD1.
- Privacy protection data is similarly generated based on local data in each of the other terminals 10 other than terminal 10A.
- the communication unit 170 of the terminal 10 transmits the data that has been subjected to the privacy protection process to the information processing device 20 as privacy protection data (S203).
- the information processing device 20 collects privacy protection data from the terminal 10 and performs a compilation process of the privacy protection data.
- the post-conversion process data DA2 shown in FIG. 6 shows an example of a data set of privacy protection data collected by the information processing device 20.
- model N1 shown in FIG. 6 indicates an inference model that the information processing device 20 learns. As shown in FIG. 6, the information processing device 20 can learn the model N1 based on the post-conversion processing data DA2.
- the information processing device 20 can aggregate statistics of privacy protection data across multiple terminals 10 based on the converted data DA2.
- the statistics S1 shown in FIG. 6 shows an example of the aggregated statistics of privacy protection data.
- the privacy protection information is data that has been subjected to privacy protection processing, such as data conversion processing that satisfies differential privacy, for local data. Therefore, the information processing device 20 can estimate the statistics of the original data, that is, the local data, based on the statistics S1, which are the results of the aggregation processing of the privacy protection information.
- the information processing device 20 may detect a change in the data trend acquired by the terminal 10 based on the statistic S1.
- the statistic S2 shows an example of trend information indicating a change in the data trend detected by the terminal 10.
- FIG. 7 is a sequence diagram for explaining the flow of processing in the subroutine of S107 in the sequence diagram shown in FIG.
- the information processing device 20 performs learning according to a trial pattern for selecting hyperparameters of an inference model based on the privacy protection data received from the terminal 10.
- one combination of hyperparameters is selected from the trial patterns of candidate values of hyperparameters determined by the administrator of the inference model (S205).
- the learning unit 230 sets the initial parameters of the inference model to learn the inference model using the selected hyperparameters (S207).
- the learning unit 230 uses the selected hyperparameters and initial parameters to train the inference model using the privacy-preserving data as learning data (S209).
- the information processing device 20 may repeat the processes of S205 to S209 until all trial patterns of the hyperparameters have been covered.
- the inference model administrator sets the hyperparameters of the inference model based on the results of the learning.
- the learning unit 230 sets the initial parameters of the inference model according to the set hyperparameters (S211).
- the terminal 10 may evaluate the selected hyperparameters and initial parameters based on the results of the learning.
- hyperparameters of the inference model may be set based on the evaluation results by the terminal 10.
- FIG. 8 is a sequence diagram illustrating the flow of the subroutine process of S109 in the sequence diagram shown in FIG.
- the learning unit 230 of the information processing device 20 causes the communication unit 250 to transmit an instruction to learn the inference model to the terminal 10 together with information on the hyperparameters and initial parameters set in S107 shown in FIG. 4 (S213).
- Each of the multiple terminals 10 uses the local data as training data to train the inference model (S215).
- Each of the multiple terminals 10 transmits update information for the inference model to the information processing device 20 based on the results of the learning (S217).
- the learning unit 230 of the information processing device 20 aggregates the update information received from the multiple terminals 10 and updates the inference model held by the information processing device 20 (S219).
- the privacy protection data is statistics of local data generated by each of the multiple terminals 10.
- the information processing device 20 estimates the distribution of the local data based on the statistics of the local data.
- the information processing device 20 can learn an inference model using synthetic data generated based on the estimated distribution of the local data as learning data.
- FIG. 9 is a sequence diagram for explaining a second operation example of the information processing system according to this embodiment. Note that S107, S109, S111, S113, and S115 shown in FIG. 9 are as explained above with reference to FIG. 4, and therefore redundant explanations will be omitted.
- the information processing device 20 aggregates the statistical data of the local data generated by the terminal 10 as privacy protection data (S131).
- the information processing device 20 generates composite data based on the aggregated privacy-preserving data (statistical data) (S105).
- the information processing device 203 of the information processing device 20 uses the synthetic data generated in S105 as learning data to learn an inference model. Next, the processes of S107 to S115 are performed.
- FIG. 10 is a sequence diagram for explaining the process flow of the subroutines S131 and S105 in the sequence diagram shown in FIG. 9.
- the data processing unit 130 of the terminal 10 performs a predetermined data conversion process on the statistical data, thereby performing privacy protection processing (S302).
- FIG. 11 is a diagram for explaining another example of the processing for generating privacy protection data by the data processing unit 130.
- the local data LD2 shown in FIG. 11 shows an example of local data collected by the acquisition unit 110 of the terminal 10A.
- the local data statistics LS1 shown in FIG. 11 indicate the statistics data of the local data calculated by the data processing unit 130 of the terminal 10A.
- the statistics DB1 after conversion processing indicates an example of privacy-preserving data generated by the data processing unit 130 performing a data conversion process that satisfies differential privacy on the local data statistics LS1.
- the data processing unit 130 of the terminal 10 causes the communication unit 170 to transmit the privacy-protected statistical data to the information processing device 20 as privacy-protected data (S303).
- the information processing device 20 estimates a distribution of local data across the multiple terminals 10, based on the privacy-protected statistical data received from each of the multiple terminals 10.
- the information processing device 20 generates composite data based on the estimated distribution of local data (S304).
- the statistics S3 show an example of statistics of local data collected by the information processing device 20.
- the information processing device 20 estimates the distribution of the local data based on the statistics S3, and generates synthetic data GD1 based on the estimated distribution.
- the information processing device 20 may use the synthetic data GD1 as training data to train the model N1.
- the information processing device 20 may calculate statistics S4 as trend information based on statistics S3.
- the local data acquired by multiple terminals 10 is statistics of features extracted from input data such as images, and label information associated with each feature.
- the feature may be a feature extracted from a medical image. Also, there may be multiple features.
- the label information included in the local data is two types of labels, correct and incorrect, in a binary classification problem in which input data is classified as correct or incorrect.
- FIG. 12 is a sequence diagram for explaining a third operation example of the information processing system according to this embodiment. Note that S105, S107, S109, S111, S113, and S115 shown in FIG. 12 are as explained above with reference to FIG. 4 and FIG. 9, so duplicate explanations will be omitted.
- the information processing device 20 and the terminal 10 perform a process of aggregating statistics of the local data that has been privacy-protected as privacy protection information.
- the information processing device 20 distributes trend information indicating the data trend of the local data estimated based on the results of the privacy protection data aggregation to each of the multiple terminals 10 (S141).
- Each of the multiple terminals 10 detects abnormal values contained in the local data based on the distributed trend information (S103).
- the terminal 10 learns the model in S109, it uses data obtained by removing outliers from the local data as learning data.
- FIG. 13 is a sequence diagram for explaining the process flow of the subroutines S141 and S103 in the sequence diagram shown in FIG. 12.
- FIG. 14 is a diagram for explaining the aggregation of privacy protection data in a third operation example of the information processing system according to this embodiment.
- terminal 10A, terminal 10B, and terminal 10C are illustrated as an example, but as described above, the number of terminals 10 according to this embodiment is not limited to this example.
- the information processing system according to this embodiment may include two or more terminals 10.
- the data processing unit 130 of the terminal 10A calculates feature statistics for each correct label for features included in the acquired local data (S401).
- the terminals 10B and 10C also calculate feature statistics for each correct label, respectively (S402, S403).
- Terminal 10A performs privacy protection processing on the calculated statistical data (S404).
- Terminals 10B and 10C also perform privacy protection processing (S405, S406).
- Local data statistics LS2, local data statistics LS3, and local data statistics LS4 shown in FIG. 14 indicate privacy-protected data, which is data in which privacy protection processing has been performed on the statistical data of the features for each correct label calculated by terminal 10A, terminal 10B, and terminal 10C, respectively.
- each of the multiple terminals 10 transmits the privacy-protected statistical data and label information to the information processing device 20 (S407, S408, S409).
- the generation unit 210 of the information processing device 20 performs a compilation process of the statistical data received from the multiple terminals 10 (S410).
- the generation unit 210 calculates the data trend (distribution of features) for each correct label based on the results of the aggregation process (S411).
- the counting result AS1 shown in FIG. 14 indicates the result of the counting process of the statistical data by the information processing device 20. Furthermore, the counting result AS2 indicates trend information, which is the trend of the data for each correct label, calculated based on the counting result AS1.
- the generation unit 210 of the information processing device 20 distributes the calculated data trend (trend information) to each of the multiple terminals 10 (S412, S413, S414).
- FIG. 15 is a conceptual diagram for explaining the detection of anomalies in the third operation example of this embodiment.
- the local data statistics LS5 shown in FIG. 15 indicate the distribution of features for each correct label included in the local data newly acquired by the acquisition unit 110 of the terminal 10.
- the points indicated by circles in the local data statistics LS5 indicate distribution points of features that have been labeled as correct. Furthermore, the distribution points indicated by Xs in the local data statistics LS5 indicate distribution points of features that have been labeled as incorrect.
- sample point P1 indicates the distribution point of the feature that is included in the local data statistics LS5 and has been assigned the correct label.
- Sample point P1 is a distribution point that has been assigned a correct label, but it can be seen that in the distribution of local data statistics LS5, it is distributed at a position away from the range indicated by the solid ellipse.
- sample point P1 can be considered an abnormal value based on the distribution trend of the local data acquired by terminal 10A.
- sample point P1 can be regarded as an abnormal value in the distribution trend of the local data when viewed across multiple terminals 10 as a whole.
- FIG. 16 is another conceptual diagram for explaining the detection of abnormal values in the third operation example of this embodiment.
- the distribution range AC1 and distribution range AC2 shown in FIG. 16 indicate the distribution tendency of data based on the trend information distributed from the information processing device 20 to the terminal 10A.
- sample point P1 is not included in the distribution range of distribution range AC1. Therefore, it can be understood that sample point P1 can be considered an abnormal value even in terms of the distribution trend of local data across multiple terminals 10.
- terminal 10A can detect abnormal values contained in the local data with a high degree of accuracy based on the distribution trend of the local data across multiple terminals 10 distributed from the information processing device 20.
- the trend information distributed from information processing device 20 is used to detect abnormal values contained in the local data (S416, S417).
- the information processing device 20 and the terminal 10 generate a generative model for generating synthetic data by associative learning.
- the information processing device 20 uses the synthetic data generated by the generative model as training data to train an inference model.
- FIG. 17 is a diagram for explaining the generation of privacy protection data in a modified example of this embodiment.
- Local data LD3 shown in FIG. 17 indicates local data acquired by terminal 10A.
- the generative model GN indicates a generative model generated by the associative learning of the information processing device 20 and the terminal 10.
- the generative model GN1 is a generative model that is learned by each of the multiple terminals 10, and is a model that corresponds to the generative model GN2.
- the generative model GN2 is a generative model that is updated by the information processing device 20 based on the learning results of each of the generative models GN1.
- Terminal 10A uses local data LD3 as training data to train generative model GN1.
- Generative model GN1 represents the generative model that has been trained by terminal 10A.
- Terminal 10A transmits update information of the generative model obtained as a result of learning generative model GN1 to the information processing device 20.
- the generative model is similarly learned on terminals 10 other than terminal 10A (terminal 10B, terminal 10C), and update information for the generative model is transmitted to the information processing device 20.
- the above-mentioned generative model may be, for example, a model generated by a DPGAN (Differential Privacy Generative Adversarial Network) algorithm, which is a generative model that prevents training data from being identified from a trained generative model by adding noise to the gradient and parameters of the loss function during the learning process.
- DPGAN Different Privacy Generative Adversarial Network
- the information processing device 20 may distribute the updated generative model GN2 or parameter information of the updated generative model GN2 to each of the terminals 10.
- the information processing device 20 generates synthetic data using the above-mentioned trained generative model.
- the synthetic data GD2 shown in FIG. 17 indicates synthetic data generated using the generative model GN2.
- the information processing device 20 may use the synthetic data GD2 as training data to train the model N1.
- the information processing device 20 may also monitor data trends of the local data based on the generated synthetic data.
- statistics S5 indicate trend information data generated from statistics of the local data estimated based on the synthetic data GD2.
- the inference model can be learned without the local data acquired by the terminal 10 itself being collected in the information processing device 20. Therefore, privacy protection in this information processing system can be guaranteed.
- synthetic data is generated using a generative model that is generated using local data as training data. This makes it possible to generate synthetic data that is closer to the data actually used for training. Therefore, the inference accuracy of the inference model by the information processing device 20 can be improved.
- FIG. 18 is a sequence diagram for explaining an example of operation in a modified example of the information processing system according to this embodiment.
- S103, S107, S109, S111, S113, and S115 shown in FIG. 18 are as explained above with reference to FIG. 4, FIG. 9, and FIG. 12, so duplicate explanations will be omitted.
- the information processing device 20 and the terminal 10 learn the generative model through associative learning (S151).
- the information processing device 20 then generates synthetic data based on the above generative model (S155).
- the learning unit 230 of the information processing device 20 learns the inference model using the synthetic data generated based on the generative model as learning data.
- FIG. 19 is a sequence diagram that explains the process flow of the subroutines S151 and S155 in the sequence diagram shown in FIG. 18.
- the learning unit 230 of the information processing device 20 transmits a learning instruction including setting values of the initial parameters and hyperparameters of the generative model to each of the multiple terminals 10 (S500).
- initial parameters and hyperparameters may be set to random values the first time the associative learning loop of S500 to S503 is performed.
- the learning units 150 of the multiple terminals 10 use the local data as learning data to learn the generative model for which they received a learning instruction from the information processing device 20 (S501).
- Each of the learning units 150 causes the communication unit 170 to transmit update information of the generative model to the information processing device 20 based on the learning results (S502).
- the information processing device 20 aggregates the update information of the generative model received from each of the multiple terminals 10, and updates the generative model based on the aggregated results (S503).
- the information processing device 20 and the terminal 10 repeat the processes of S500 to S502 until the learning of the generative model converges.
- FIG. 20 is a block diagram illustrating a hardware configuration 90 according to one embodiment of the present disclosure.
- the hardware configuration 90 can be applied to the terminal 10 and the information processing device 20.
- the hardware configuration 90 includes, for example, a processor 901, a ROM (Read Only Memory) 903, a RAM (Random Access Memory) 905, a host bus 907, a bridge 909, an external bus 911, an interface 913, an input device 915, an output device 917, a storage device 919, a drive 921, a connection port 923, and a communication device 925.
- a processor 901 a ROM (Read Only Memory) 903, a RAM (Random Access Memory) 905, a host bus 907, a bridge 909, an external bus 911, an interface 913, an input device 915, an output device 917, a storage device 919, a drive 921, a connection port 923, and a communication device 925.
- a processor 901 a ROM (Read Only Memory) 903
- RAM Random Access Memory
- a host bus 907 a bridge 909
- an external bus 911 an interface 913
- an input device 915 an output device 917
- the processor 901 functions, for example, as an arithmetic processing device and a control device, and controls the overall operation of each component or part of it based on various programs recorded in the ROM 903, the RAM 905, the storage device 919, or the removable recording medium 927.
- the ROM 903 is a means for storing the programs and/or data used in the calculations read by the processor 901.
- the RAM 905 temporarily or permanently stores, for example, the programs and/or various parameters that are appropriately changed when the programs are executed by the processor 901.
- the processor 901, ROM 903, and RAM 905 are connected to one another via, for example, a host bus 907 capable of high-speed data transmission.
- the host bus 907 is connected to an external bus 911, which has a relatively low data transmission speed, via, for example, a bridge 909.
- the external bus 911 is connected to various components via an interface 913.
- the input device 915 examples include a mouse, a keyboard, a touch panel, a button, a switch, and a lever. Furthermore, a remote controller capable of transmitting a control signal using infrared rays or other radio waves may be used as the input device 915.
- the input device 915 also includes an audio input device such as a microphone.
- the input device 915 may also include an imaging device and a sensor.
- the imaging device is a device that captures real space and generates a captured image using various components such as an imaging element, such as a CCD (Charge Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor), and a lens for controlling the formation of a subject image on the imaging element.
- the imaging device may capture still images or may capture moving images.
- the sensors are various types of sensors, such as distance sensors, acceleration sensors, gyro sensors, geomagnetic sensors, vibration sensors, light sensors, and sound sensors.
- the sensors obtain information about the state of the hardware configuration 90 itself, such as the attitude of the housing of the hardware configuration 90, or information about the surrounding environment of the hardware configuration 90, such as the brightness or noise around the hardware configuration 90.
- the sensors may also include a Global Positioning System (GPS) sensor that receives GPS signals to measure the latitude, longitude, and altitude of the device.
- GPS Global Positioning System
- the output device 917 includes various vibration devices capable of visually or audibly notifying the user of acquired information, such as display devices such as CRTs (Cathode Ray Tubes), LCDs (Liquid Crystal Displays), or organic EL (Electro-Luminescence), audio output devices such as speakers and headphones, printers, mobile phones, or facsimiles.
- display devices such as CRTs (Cathode Ray Tubes), LCDs (Liquid Crystal Displays), or organic EL (Electro-Luminescence)
- audio output devices such as speakers and headphones, printers, mobile phones, or facsimiles.
- the storage device 919 is a device for storing various types of data.
- a magnetic storage device such as a hard disk drive (HDD), a semiconductor storage device, an optical storage device, or a magneto-optical storage device is used.
- the drive 921 is a device that reads information recorded on a removable recording medium 927 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, or writes information to the removable recording medium 927 .
- a removable recording medium 927 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory
- the removable recording medium 927 is, for example, a DVD medium, a Blu-ray (registered trademark) medium, an HD DVD medium, various semiconductor storage media, etc.
- the removable recording medium 927 may be, for example, an IC card equipped with a non-contact type IC chip, an electronic device, etc.
- connection port 923 is a port for connecting an external device 929, such as a Universal Serial Bus (USB) port, an IEEE 1394 port, a Small Computer System Interface (SCSI) port, an RS-232C port, or an optical audio terminal.
- USB Universal Serial Bus
- SCSI Small Computer System Interface
- RS-232C Small Computer System Interface
- the external connection device 929 is, for example, a printer, a portable music player, a digital camera, a digital video camera, or an IC recorder.
- the communication device 925 is a communication device for connecting to a network, such as a communication card for a wired or wireless LAN (Local Area Network), Bluetooth (registered trademark), or WUSB (Wireless USB), a router for optical communications, a router for ADSL (Asymmetric Digital Subscriber Line), or a modem for various types of communications.
- a network such as a communication card for a wired or wireless LAN (Local Area Network), Bluetooth (registered trademark), or WUSB (Wireless USB), a router for optical communications, a router for ADSL (Asymmetric Digital Subscriber Line), or a modem for various types of communications.
- the steps in the processing of the operation of the terminal 10 and the information processing device 20 according to this embodiment do not necessarily have to be processed chronologically in the order described in the explanatory diagram.
- each step in the processing of the operation of the terminal 10 and the information processing device 20 may be processed in an order different from the order described in the explanatory diagram, or may be processed in parallel.
- a learning unit that learns an inference model by federated learning
- An acquisition unit that acquires privacy protection data from a plurality of terminals, the privacy protection data being data obtained by performing privacy protection processing on local data obtained by each of the plurality of terminals;
- the learning unit is training the inference model based on the privacy-preserving data; Distributing information about the inference model, including hyperparameters of the inference model that are set based on the results of the learning, to the multiple terminals;
- the acquisition unit is acquiring, from the plurality of terminals, update information of the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters with the local data as learning data;
- the learning unit is updating the inference model using the update information;
- Information processing device that acquires privacy protection data from a plurality of terminals, the privacy protection data being data obtained by performing privacy protection processing on local data obtained by each of the plurality of terminals;
- the learning unit is training the inference model based on the privacy-preserving data
- the learning unit learns the inference model using the privacy preserving data as learning data.
- the information processing device according to (1) (3) A generator for generating synthetic data based on the privacy preserving data, The learning unit learns the inference model using the synthetic data as learning data.
- the privacy-preserving data is data generated by performing a data conversion process that satisfies differential privacy on the local data.
- the data conversion process is a process of assigning a random number of a predetermined strength to each element included in the local data.
- (6) The information processing device according to (5), wherein the data conversion process is performed using a Laplace mechanism or a Gaussian mechanism.
- the privacy preserving data is generated by performing a data conversion process on the local data to reduce a dimension of the local data.
- the privacy-preserving data is statistical data generated by performing a data conversion process that satisfies differential privacy on statistical data of the local data,
- the generation unit estimates a distribution of the local data based on the privacy-preserving data; generating the synthetic data based on the estimated distribution;
- the privacy protection data is statistical data generated by performing an encryption process that satisfies requirements for secure computation on the statistical data of the local data,
- the generating unit performs a compilation process of the privacy protection data while the privacy protection data remains encrypted, estimating a distribution of the local data based on a result of the aggregation process; generating the synthetic data based on the estimated distribution;
- the information processing device according to (3).
- the generating unit performs a compilation process of the privacy protection data, Distributing trend information indicating a statistical data trend of the local data to each of the plurality of terminals based on a result of the aggregation process;
- the acquisition unit acquires, from each of the plurality of terminals, the privacy protection data generated by each of the plurality of terminals by correcting or excluding local data that is deemed to be an abnormal value based on the trend information.
- the information processing device according to (3).
- the privacy preserving data includes feature amounts of elements included in the local data and label information associated with each feature amount;
- the generation unit distributes a distribution of the feature amounts for each of the label information as the trend information.
- the information processing device according to (10).
- the generating unit monitors changes in distribution trends of the local data based on a result of the aggregation process of the privacy protection data, the learning unit, when the generation unit detects that the distribution trend has changed, re-learns the inference model using associative learning;
- the information processing device according to (3).
- the generation unit generates the synthetic data based on a generative model generated based on distribution information of the local data estimated from the privacy-preserving data or the local data.
- the information processing device according to (3).
- the acquisition unit acquires, from the multiple terminals, difference information indicating a difference between parameters of the inference model before and after the update as update information of the inference model, and
- the learning unit updates the inference model based on the difference information.
- the information processing device according to any one of (1) to (14).
- the processor Learning an inference model through federated learning; Obtaining privacy protection data from a plurality of terminals, the data being data obtained by performing privacy protection processing on local data obtained by each of the plurality of terminals; training the inference model based on the privacy-preserving data; Distributing information about the inference model, including hyperparameters of the inference model that are set based on the results of the learning, to the multiple terminals; acquiring, from the plurality of terminals, update information of the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters with the local data as learning data; updating the inference model using the update information; and 2.
- An information processing method implemented by a computer comprising: (16) Computer, a learning unit that learns an inference model by federated learning; An acquisition unit that acquires privacy protection data from a plurality of terminals, the privacy protection data being data obtained by performing privacy protection processing on local data obtained by each of the plurality of terminals; The learning unit is training the inference model based on the privacy-preserving data; Distributing information about the inference model, including hyperparameters of the inference model that are set based on the results of the learning, to the multiple terminals; The acquisition unit is acquire, from the plurality of terminals, update information of the inference model obtained by each of the plurality of terminals learning the inference model using the distributed hyperparameters with the local data as learning data; The learning unit is updating the inference model using the update information; A program that functions as an information processing device.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioethics (AREA)
- Databases & Information Systems (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Probability & Statistics with Applications (AREA)
- Algebra (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
1.概要
2.システム構成例
3.機能構成例
3-1.端末10
3-2.情報処理装置20
4.動作例
4-1.第1の動作例
4-2.第2の動作例
4-3.第3の動作例
5.変形例
6.ハードウェア構成例
7.まとめ
まず、本開示の一実施形態の概要について述べる。
図1は、本開示の一実施形態に係る情報処理システムの構成例を示す説明図である。
本実施形態に係る端末10は、取得したローカルデータを学習データとして、情報処理装置20から配布される推論モデルの学習を行う。
本実施形態に係る情報処理装置20は、複数の端末10から取得したプライバシー保護情報に基づき生成した推論モデルを、端末10に配布する。
本実施形態に係るネットワーク30は、端末10と情報処理装置20との間における通信を仲介する。
<<3-1.端末10の構成例>>
続いて、本実施形態に係る端末10の構成例について詳細に説明する。
本実施形態に係る取得部110は、各種のデータを収集する。
本実施形態に係るデータ処理部130は、取得部110により取得されたローカルデータに基づいて、プライバシー保護データを生成する機能を有する。
本実施形態に係る学習部150は、取得部110により取得されるローカルデータを学習データとして、情報処理装置20から配布される推論モデルの学習を行う。
本実施形態に係る通信部170は、ネットワーク30を介して情報処理装置20との通信を行う。
次に、本実施形態に係る情報処理装置20の構成例について詳細に説明する。
本実施形態に係る生成部210は、端末10から取得されたプライバシー保護データの集計処理を行う。
本実施形態に係る学習部230は、連合学習により推論モデルの学習を行う。
本実施形態に係る通信部250は、ネットワーク30を介して複数の端末10と通信を行う。通信部250は、情報処理装置20の取得部の一例である。
続いて、図4~図16を参照して、本開示の一実施形態に係る情報処理システムの動作例を説明する。
まず、図4~8を参照し、本実施形態に係る情報処理システムの第1の動作例を説明する。第1の動作例では、端末10のデータ処理部130により生成されるプライバシー保護データが、ローカルデータに対して差分プライバシーを満たすデータ変換処理または/およびデータの次元を削減する処理が行われることにより生成されたデータである例を説明する。
図5は、図4に示したシーケンス図におけるS101のサブルーチンの処理の流れを説明するシーケンス図である。
図7は、図4に示したシーケンス図におけるS107のサブルーチンの処理の流れを説明するシーケンス図である。
次に、図8は、図4に示したシーケンス図におけるS109のサブルーチンの処理の流れを説明するシーケンス図である。
次に、図9~11を参照して、本実施形態に係る情報処理システムの第2の動作例を説明する。
図10に示したように、複数の端末10のデータ処理部130は、ローカルデータの統計量を計算する(S301)。
情報処理装置20は、複数の端末10の各々から受信したプライバシー保護処理済みの統計量データに基づき、複数の端末10全体でのローカルデータの分布を推定する。情報処理装置20は、推定したローカルデータの分布に基づき、合成データを生成する(S304)。
次に、図12~図16を参照して、本実施形態に係る情報処理システムの第3の動作例を説明する。
図13に示したように、端末10Aのデータ処理部130は、取得したローカルデータに含まれる特徴量について、正解ラベル毎に特徴量の統計量を計算する(S401)。端末10Bおよび端末10Cにおいても、それぞれ、正解ラベル毎に特徴量の統計量が計算される(S402、S403)。
端末10Aは、情報処理装置20から配布された傾向情報に基づいて、取得部110が取得したローカルデータに含まれる異常値の検知を行う(S415)。
次に、図17~図19を用いて、上記で説明した本実施形態に係る情報処理システムの変形例を説明する。
図19に示したように、情報処理装置20の学習部230は、複数の端末10の各々に、生成モデルの初期パラメータおよびハイパーパラメータの設定値を含む学習指示を送信する。(S500)。
次いで、情報処理装置20の生成部210は、学習済みの生成モデルを用いて合成データを生成する(S504)。
以上、本開示の実施形態を説明した。次に、本開示の一実施形態に係る端末10および情報処理装置20に共通するハードウェア構成例について説明する。
プロセッサ901は、例えば、演算処理装置および制御装置として機能し、ROM903、RAM905、ストレージ装置919、またはリムーバブル記録媒体927に記録された各種プログラムに基づいて各構成要素の動作全般またはその一部を制御する。
ROM903は、プロセッサ901に読み込まれるプログラムまたは/および演算に用いられるデータ等を格納する手段である。RAM905には、例えば、プロセッサ901に読み込まれるプログラムまたは/および、当該プログラムを実行する際に適宜変化する各種パラメータ等が一時的または永続的に格納される。
プロセッサ901、ROM903、およびRAM905は、例えば、高速なデータ伝送が可能なホストバス907を介して相互に接続される。一方、ホストバス907は、例えば、ブリッジ909を介して比較的データ伝送速度が低速な外部バス911に接続される。また、外部バス911は、インターフェース913を介して種々の構成要素と接続される。
入力装置915には、例えば、マウス、キーボード、タッチパネル、ボタン、スイッチおよびレバーなどが用いられる。さらに、入力装置915としては、赤外線またはその他の電波を利用して制御信号を送信することが可能なリモートコントローラ(以下、リモコン)が用いられることもある。また、入力装置915には、マイクロフォンなどの音声入力装置が含まれる。
出力装置917は、例えば、CRT(Cathode Ray Tube)、LCD(Liquid Crystal Display)、または有機EL(Electro-Luminescence)などのディスプレイ装置、スピーカおよびヘッドホンなどのオーディオ出力装置、プリンタ、携帯電話、またはファクシミリ等、取得した情報を利用者に対して視覚的または聴覚的に通知することが可能な種々の振動デバイスを含む。
ストレージ装置919は、各種のデータを格納するための装置である。ストレージ装置919としては、例えば、ハードディスクドライブ(HDD)などの磁気記憶デバイス、半導体記憶デバイス、光記憶デバイス、または光磁気記憶デバイスなどが用いられる。
ドライブ921は、例えば、磁気ディスク、光ディスク、光磁気ディスク、または半導体メモリなどのリムーバブル記録媒体927に記録された情報を読み出し、またはリムーバブル記録媒体927に情報を書き込む装置である。
リムーバブル記録媒体927は、例えば、DVDメディア、Blu-ray(登録商標)メディア、HD DVDメディア、各種の半導体記憶メディア等である。もちろん、リムーバブル記録媒体927は、例えば、非接触型ICチップを搭載したICカード、または電子機器等であってもよい。
接続ポート923は、例えば、USB(Universal Serial Bus)ポート、IEEE1394ポート、SCSI(Small Computer System Interface)ポート、RS-232Cポート、または光オーディオ端子等のような外部接続機器929を接続するためのポートである。
外部接続機器929は、例えば、プリンタ、携帯音楽プレーヤ、デジタルカメラ、デジタルビデオカメラ、またはICレコーダ等である。
通信装置925は、ネットワークに接続するための通信デバイスであり、例えば、有線または無線LAN(Local Area Network)、Bluetooth(登録商標)、またはWUSB(Wireless USB)用の通信カード、光通信用のルータ、ADSL(Asymmetric Digital Subscriber Line)用のルータ、または、各種通信用のモデムなどである。
以上、添付図面を参照しながら本開示の好適な実施形態について詳細に説明したが、本開示の技術的範囲はかかる例に限定されない。本開示の技術分野における通常の知識を有する者であれば、請求の範囲に記載された技術的思想の範疇内において、各種の変更例または修正例に想到し得ることは明らかであり、これらについても、当然に本開示の技術的範囲に属するものと了解される。
(1)
連合学習により推論モデルの学習を行う学習部と、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得する取得部と、を備え、
前記学習部は、
前記プライバシー保護データに基づいて前記推論モデルの学習を行い、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布し、
前記取得部は、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得し、
前記学習部は、
前記更新情報を用いて前記推論モデルを更新する、
情報処理装置。
(2)
前記学習部は、前記プライバシー保護データを学習データとして前記推論モデルの学習を行う、
前記(1)に記載の情報処理装置。
(3)
前記プライバシー保護データに基づいて合成データを生成する生成部をさらに備え、
前記学習部は、前記合成データを学習データとして前記推論モデルの学習を行う、
前記(1)に記載の情報処理装置。
(4)
前記プライバシー保護データは、前記ローカルデータに対して、差分プライバシーを満たすデータ変換処理が行われることにより生成されたデータである、
前記(2)または(3)に記載の情報処理装置。
(5)
前記データ変換処理は、前記ローカルデータに含まれる要素の各々に対して、予め定められた強度の乱数を付与する処理である、
前記(4)に記載の情報処理装置。
(6)
前記データ変換処理は、ラプラスメカニズム、または、ガウシアンメカニズムを用いて実施される、前記(5)に記載の情報処理装置。
(7)
前記プライバシー保護データは、前記ローカルデータに対して、前記ローカルデータの次元を削減するデータ変換処理が行われることにより生成される、
前記(2)に記載の情報処理装置。
(8)
前記プライバシー保護データは、前記ローカルデータの統計量データに対し差分プライバシーを満たすデータ変換処理が行われることにより生成された統計量データであり、
前記生成部は、前記プライバシー保護データに基づき前記ローカルデータの分布を推定し、
推定された前記分布に基づいて前記合成データを生成する、
前記(3)に記載の情報処理装置。
(9)
前記プライバシー保護データは、前記ローカルデータの統計量データに対し、秘密計算における要件を満たす暗号化処理が行われることにより生成された統計量データであり、
前記生成部は、前記プライバシー保護データが暗号化されたままの状態で、前記プライバシー保護データの集計処理を行い、
前記集計処理の結果に基づいて前記ローカルデータの分布を推定し、
推定された前記分布に基づいて前記合成データを生成する、
前記(3)に記載の情報処理装置。
(10)
前記生成部は、前記プライバシー保護データの集計処理を行い、
当該集計処理の結果に基づき、前記ローカルデータの統計的なデータ傾向を示す傾向情報を前記複数の端末の各々に配布し、
前記取得部は、前記複数の端末の各々により前記傾向情報に基づき異常値と見做されたローカルデータを修正または除外して生成された前記プライバシー保護データを、前記複数の端末の各々から取得する、
前記(3)に記載の情報処理装置。
(11)
前記プライバシー保護データは、前記ローカルデータに含まれる要素の特徴量および各特徴量に関連付けられたラベル情報を含み、
前記生成部は、前記ラベル情報ごとの前記特徴量の分布を、前記傾向情報として配布する、
前記(10)に記載の情報処理装置。
(12)
前記生成部は、前記プライバシー保護データの集計処理結果に基づき前記ローカルデータの分布傾向の変化を監視し、
前記学習部は、前記生成部により前記分布傾向が変化したことが検知されると、連合学習を用いて前記推論モデルの再学習を行う、
前記(3)に記載の情報処理装置。
(13)
前記生成部は、前記プライバシー保護データまたは前記ローカルデータから推定される前記ローカルデータの分布情報に基づいて生成される生成モデルに基づき、前記合成データを生成する、
前記(3)に記載の情報処理装置。
(14)
前記取得部は、前記複数の端末から、前記推論モデルの更新情報として、更新前および更新後の推論モデルのパラメータの差分を示す差分情報を取得し、
前記学習部は、当該差分情報に基づいて前記推論モデルを更新する、
前記(1)~(14)のいずれか一項に記載の情報処理装置。
(15)
プロセッサが、
連合学習により推論モデルの学習を行うことと、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得することと、
前記プライバシー保護データに基づいて前記推論モデルの学習を行うことと、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布することと、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得することと、
前記更新情報を用いて前記推論モデルを更新することと、
を含む、コンピュータにより実行される情報処理方法。
(16)
コンピュータを、
連合学習により推論モデルの学習を行う学習部と、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得する取得部と、を備え、
前記学習部は、
前記プライバシー保護データに基づいて前記推論モデルの学習を行い、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布し、
前記取得部は、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得し、
前記学習部は、
前記更新情報を用いて前記推論モデルを更新する、
情報処理装置、として機能させるための、プログラム。
110 取得部
130 データ処理部
150 学習部
170 通信部
20 情報処理装置
210 生成部
230 学習部
250 通信部
30 ネットワーク
Claims (16)
- 連合学習により推論モデルの学習を行う学習部と、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得する取得部と、を備え、
前記学習部は、
前記プライバシー保護データに基づいて前記推論モデルの学習を行い、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布し、
前記取得部は、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得し、
前記学習部は、
前記更新情報を用いて前記推論モデルを更新する、
情報処理装置。 - 前記学習部は、前記プライバシー保護データを学習データとして前記推論モデルの学習を行う、
請求項1に記載の情報処理装置。 - 前記プライバシー保護データに基づいて合成データを生成する生成部をさらに備え、
前記学習部は、前記合成データを学習データとして前記推論モデルの学習を行う、
請求項1に記載の情報処理装置。 - 前記プライバシー保護データは、前記ローカルデータに対して、差分プライバシーを満たすデータ変換処理が行われることにより生成されたデータである、
請求項2に記載の情報処理装置。 - 前記データ変換処理は、前記ローカルデータに含まれる要素の各々に対して、予め定められた強度の乱数を付与する処理である、
請求項4に記載の情報処理装置。 - 前記データ変換処理は、ラプラスメカニズム、または、ガウシアンメカニズムを用いて実施される、
請求項5に記載の情報処理装置。 - 前記プライバシー保護データは、前記ローカルデータに対して、前記ローカルデータの次元を削減するデータ変換処理が行われることにより生成される、
請求項2に記載の情報処理装置。 - 前記プライバシー保護データは、前記ローカルデータの統計量データに対し差分プライバシーを満たすデータ変換処理が行われることにより生成された統計量データであり、
前記生成部は、前記プライバシー保護データに基づき前記ローカルデータの分布を推定し、
推定された前記分布に基づいて前記合成データを生成する、
請求項3に記載の情報処理装置。 - 前記プライバシー保護データは、前記ローカルデータの統計量データに対し、秘密計算における要件を満たす暗号化処理が行われることにより生成された統計量データであり、
前記生成部は、前記プライバシー保護データが暗号化されたままの状態で、前記プライバシー保護データの集計処理を行い、
前記集計処理の結果に基づいて前記ローカルデータの分布を推定し、
推定された前記分布に基づいて前記合成データを生成する、
請求項3に記載の情報処理装置。 - 前記生成部は、前記プライバシー保護データの集計処理を行い、
当該集計処理の結果に基づき、前記ローカルデータの統計的なデータ傾向を示す傾向情報を前記複数の端末の各々に配布し、
前記取得部は、前記複数の端末の各々により前記傾向情報に基づき異常値と見做されたローカルデータを修正または除外して生成された前記プライバシー保護データを、前記複数の端末の各々から取得する、
請求項3に記載の情報処理装置。 - 前記プライバシー保護データは、前記ローカルデータに含まれる要素の特徴量および各特徴量に関連付けられたラベル情報を含み、
前記生成部は、前記ラベル情報ごとの前記特徴量の分布を、前記傾向情報として配布する、
請求項10に記載の情報処理装置。 - 前記生成部は、前記プライバシー保護データの集計処理結果に基づき前記ローカルデータの分布傾向の変化を監視し、
前記学習部は、前記生成部により前記分布傾向が変化したことが検知されると、連合学習を用いて前記推論モデルの再学習を行う、
請求項3に記載の情報処理装置。 - 前記生成部は、前記プライバシー保護データまたは前記ローカルデータから推定される前記ローカルデータの分布情報に基づいて生成される生成モデルに基づき、前記合成データを生成する、
請求項3に記載の情報処理装置。 - 前記取得部は、前記複数の端末から、前記推論モデルの更新情報として、更新前および更新後の推論モデルのパラメータの差分を示す差分情報を取得し、
前記学習部は、当該差分情報に基づいて前記推論モデルを更新する、
請求項1に記載の情報処理装置。 - プロセッサが、
連合学習により推論モデルの学習を行うことと、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得することと、
前記プライバシー保護データに基づいて前記推論モデルの学習を行うことと、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布することと、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得することと、
前記更新情報を用いて前記推論モデルを更新することと、
を含む、コンピュータにより実行される情報処理方法。 - コンピュータを、
連合学習により推論モデルの学習を行う学習部と、
複数の端末の各々により得られるローカルデータに対してプライバシー保護処理が行われたデータであるプライバシー保護データを、前記複数の端末から取得する取得部と、を備え、
前記学習部は、
前記プライバシー保護データに基づいて前記推論モデルの学習を行い、
前記学習の結果に基づいて設定される前記推論モデルのハイパーパラメータを含む前記推論モデルに関する情報を前記複数の端末に配布し、
前記取得部は、
前記複数の端末から、前記複数の端末の各々により、前記ローカルデータを学習データとして、配布された前記ハイパーパラメータを用いた前記推論モデルの学習が行われることにより得られる前記推論モデルの更新情報を取得し、
前記学習部は、
前記更新情報を用いて前記推論モデルを更新する、
情報処理装置、として機能させるための、プログラム。
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2024551285A JPWO2024084827A1 (ja) | 2022-10-20 | 2023-08-30 | |
| EP23879479.6A EP4571601A4 (en) | 2022-10-20 | 2023-08-30 | INFORMATION PROCESSING DEVICE, INFORMATION PROCESSING METHOD, AND PROGRAM |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2022-168484 | 2022-10-20 | ||
| JP2022168484 | 2022-10-20 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2024084827A1 true WO2024084827A1 (ja) | 2024-04-25 |
Family
ID=90737571
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2023/031384 Ceased WO2024084827A1 (ja) | 2022-10-20 | 2023-08-30 | 情報処理装置、情報処理方法、およびプログラム |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP4571601A4 (ja) |
| JP (1) | JPWO2024084827A1 (ja) |
| WO (1) | WO2024084827A1 (ja) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112580826A (zh) * | 2021-02-05 | 2021-03-30 | 支付宝(杭州)信息技术有限公司 | 业务模型训练方法、装置及系统 |
| CN113033824A (zh) * | 2021-04-21 | 2021-06-25 | 支付宝(杭州)信息技术有限公司 | 模型超参数确定方法、模型训练方法及系统 |
| JP2021117964A (ja) | 2020-01-22 | 2021-08-10 | キヤノンメディカルシステムズ株式会社 | 医用システム及び医用情報処理方法 |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11443240B2 (en) * | 2019-09-06 | 2022-09-13 | Oracle International Corporation | Privacy preserving collaborative learning with domain adaptation |
| US11941520B2 (en) * | 2020-01-09 | 2024-03-26 | International Business Machines Corporation | Hyperparameter determination for a differentially private federated learning process |
-
2023
- 2023-08-30 WO PCT/JP2023/031384 patent/WO2024084827A1/ja not_active Ceased
- 2023-08-30 JP JP2024551285A patent/JPWO2024084827A1/ja active Pending
- 2023-08-30 EP EP23879479.6A patent/EP4571601A4/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2021117964A (ja) | 2020-01-22 | 2021-08-10 | キヤノンメディカルシステムズ株式会社 | 医用システム及び医用情報処理方法 |
| CN112580826A (zh) * | 2021-02-05 | 2021-03-30 | 支付宝(杭州)信息技术有限公司 | 业务模型训练方法、装置及系统 |
| CN113033824A (zh) * | 2021-04-21 | 2021-06-25 | 支付宝(杭州)信息技术有限公司 | 模型超参数确定方法、模型训练方法及系统 |
Non-Patent Citations (1)
| Title |
|---|
| See also references of EP4571601A4 |
Also Published As
| Publication number | Publication date |
|---|---|
| JPWO2024084827A1 (ja) | 2024-04-25 |
| EP4571601A4 (en) | 2025-10-08 |
| EP4571601A1 (en) | 2025-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US12417401B2 (en) | Systems, devices, and/or processes for behavioral content processing | |
| EP3500985B1 (en) | Systems and methods for distributed training of deep learning models | |
| CN115210752B (zh) | 缺陷检测方法及装置、模型训练方法及装置及电子设备 | |
| US20230115987A1 (en) | Data adjustment system, data adjustment device, data adjustment method, terminal device, and information processing apparatus | |
| WO2020027454A1 (en) | Multi-layered machine learning system to support ensemble learning | |
| CN107591211A (zh) | 基于移动终端控制的健康监护方法和系统 | |
| CN112183627B (zh) | 生成预测密度图网络的方法和车辆年检标数量检测方法 | |
| US20230237774A1 (en) | Data collection system, sensor device, data collection device, and data collection method | |
| CN113674856A (zh) | 基于人工智能的医学数据处理方法、装置、设备及介质 | |
| KR20230138335A (ko) | 제품 이미지를 기반으로 불량 제품을 감지하는 인공 지능 장치 및 그 방법 | |
| WO2019022052A1 (ja) | 情報処理装置、システムおよび情報処理方法 | |
| WO2023191374A1 (ko) | 구조식 이미지를 인식하는 인공 지능 장치 및 그 방법 | |
| CN111797302A (zh) | 模型处理方法、装置、存储介质及电子设备 | |
| JP7552684B2 (ja) | 情報処理装置、情報処理端末、方法、プログラム、およびモデル | |
| WO2024084827A1 (ja) | 情報処理装置、情報処理方法、およびプログラム | |
| WO2019198900A1 (en) | Electronic apparatus and control method thereof | |
| CN113709383B (zh) | 配置视频特效方法、装置、设备及存储介质 | |
| JP7192854B2 (ja) | 演算装置、演算方法、プログラムおよび判別システム | |
| Li et al. | A comprehensive review of machine learning advances on data change: A cross-field perspective | |
| CN118043828A (zh) | 学习装置、学习方法、感测装置及数据收集方法 | |
| JP7673799B2 (ja) | 情報処理装置、情報処理サーバ、情報処理方法、およびコンピュータにより読み取り可能な非一時的な記憶媒体 | |
| WO2020129412A1 (ja) | 学習装置、識別装置およびプログラム | |
| US20250272611A1 (en) | Information processing device, information processing method, and program | |
| WO2021214861A1 (ja) | 学習装置、学習済みモデル生成方法、分類装置、分類方法、及びコンピュータ読み取り可能な記録媒体 | |
| US20260079698A1 (en) | Ai gatekeeper for shared code repositories |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23879479 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2024551285 Country of ref document: JP |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2023879479 Country of ref document: EP |
|
| ENP | Entry into the national phase |
Ref document number: 2023879479 Country of ref document: EP Effective date: 20250311 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| WWP | Wipo information: published in national office |
Ref document number: 2023879479 Country of ref document: EP |