Learning data processing device, method, and program

Abstract

A learning data processing device 2 comprises: a composite data generation unit 22 for generating composite data, which is pseudo learning data of learning raw data, on the basis of at least one piece of learning raw data; and an outlier composite data removal unit 27 for comparing the distribution of the learning raw data and the distribution of the composite data, and excluding, from the composite data, outlier composite data that deviates from the distribution of the learning raw data.

Description

Learning data processing device, method, and program

[0001] The disclosed technology relates to a technology for processing data to train a model.

[0002] Machine learning, especially supervised learning, is becoming widespread in various application fields. As the scope of applications expands, the data used for learning is becoming more sensitive and is typically continuously updated along with the model.

[0003] For widespread use in fields that handle sensitive data, there is a constraint that the data, labels, and models used in machine learning cannot be disclosed to anyone other than the provider, necessitating a secure environment and processing method. Furthermore, when building and updating training datasets and models, an interactive processing method is required in which label providers and model providers can observe the operating status of the model, such as the characteristics of the actual data and the model output results, to achieve sufficient inference performance.

[0004] Data sandboxing is known as a method for processing sensitive data in a secure environment without disclosing it to anyone other than the data provider. Data sandboxing is a technology that ensures a secure environment on hardware by utilizing hardware encryption technology (TEE), and allows both learning and inference to be performed in plaintext within that environment.

[0005] Furthermore, a conventional technique is known in which synthetic data, which is pseudo-learning data, is generated from raw training data in a data sandbox and the synthetic data is presented to a generator of labels, models, etc. (see, for example, Non-Patent Document 1). This allows for the generation of models, etc., without presenting the raw training data to anyone other than the provider of the raw training data.

[0006] Osamu Zeisho, Takayuki Miura, Kazuki Iwahana, Masanori Kii, "Active Synthetic Data Generation: Utilizing Privacy-Preserving Synthetic Data for Machine Learning with Annotation," Research Report Mobile Computing and New Social Systems (MBL), 2023-MBL-107(9), 1-8 (2023-05-11), 2188-8817

[0007] However, the accuracy of the models generated by the technique of Non-Patent Document 1 was sometimes low.

[0008] The disclosed technology aims to provide a training data processing device, method, and program for generating training data that can generate a more accurate model.

[0009] A training data processing device, which is one aspect of the disclosed technology, includes a synthetic data generation unit that generates synthetic data, which is pseudo-training data of the training raw data, based on at least one training raw data, and an outlier synthetic data removal unit that compares the distribution of the training raw data with the distribution of the synthetic data and removes outlier synthetic data, which is synthetic data that deviates from the distribution of the training raw data, from the synthetic data.

[0010] According to the disclosed technology, it is possible to generate training data that can generate a more accurate model.

[0011] Fig. 1 is a diagram showing an example of the functional configuration of a training data processing device. Fig. 2 is a diagram showing an example of a processing procedure of a training method. Fig. 3 is a diagram showing an example of the functional configuration of a training data processing device. Fig. 4 is a diagram showing an example of the functional configuration of a computer.

[0012] Hereinafter, embodiments of the disclosed technology will be described with reference to the drawings. Note that components having the same functions in the drawings are given the same reference numerals, and redundant description will be omitted.

[0013] 1, the training data processing device 2 includes, for example, a training raw data storage unit 21, a composite data generation unit 22, a composite data storage unit 23, and a deviating composite data removal unit 27. As will be described later, the training data processing device 2 may further include a presentation data extraction unit 25 shown by a dashed line in FIG. 1, a model learning management unit 26 shown by a dashed line in FIG. 1, and a model storage unit 24 shown by a dotted line in FIG. 1.

[0014] The training data processing device 2 is connected to the training raw data providing device 1, the model providing device 3, and the user device 4 so as to be able to send and receive data thereto. The training data processing device 2 may also be connected to a label providing device 5 shown by a dashed line in Fig. 1 so as to be able to send and receive data thereto.

[0015] A data sandbox is created in the training data processing device 2. The data sandbox is a secure environment on hardware that utilizes hardware encryption technology (TEE) and is a processing execution environment isolated from the outside (see, for example, Reference 1). The processing of each unit of the training data processing device 2 (the training raw data storage unit 21, the composite data generation unit 22, the composite data storage unit 23, the model storage unit 24, the presentation data extraction unit 25, the model learning management unit 26, and the outlier composite data removal unit 27) is performed in the data sandbox. The processing of at least one of these units may be performed in the data sandbox.

[0016] [Reference 1] Seiichiro Mochida, Takahiko Nagata, and Atsushi Mihara, "Next-generation data hub technology that realizes a data-centric society through high-quality, highly reliable data distribution," NTT Technical Journal, December 2020 issue, special feature "Efforts toward realizing future information processing infrastructure," pp. 23-28. In the data sandbox, data cannot be viewed from any device other than the device that provided the data without the permission of the device that provided the data. For example, raw training data cannot be viewed from any device other than raw training data providing device 1 without the permission of raw training data providing device 1. Similarly, models cannot be viewed from any device other than model providing device 3 without the permission of model providing device 3.

[0017] The learning method is realized, for example, by each component of the learning data processing device 2 performing the processes of steps S22 and S27 shown in FIG. 2, which will be described below.

[0018] Each component of the learning data processing device 2 will now be described.

[0019] <Learning Raw Data Storage Unit 21> The learning raw data providing device 1 transmits learning raw data acquired in its own local environment or the like to the learning data processing device 2.

[0020] As a result, the learning raw data storage unit 21 stores the learning raw data provided by the learning raw data providing device 1.

[0021] The learning raw data providing device 1 may transmit the learning raw data using an encrypted communication method. This can prevent leakage of the learning raw data. Of course, the communication method used by the learning raw data providing device 1 when transmitting the learning raw data is not limited to the encrypted communication method.

[0022] The number of learning raw data providing devices 1 may be one or more. That is, the learning raw data storage unit 21 may store at least one piece of learning raw data provided from at least one learning raw data providing device 1.

[0023] <Synthetic Data Generator 22> The synthetic data generator 22 generates synthetic data, which is pseudo-learning data of the learning raw data, based on at least one learning raw data (step S22).

[0024] In this example, at least one piece of learning raw data is stored in the learning raw data storage unit 21, so the synthetic data generation unit 22 generates synthetic data, which is pseudo-learning data of the learning raw data, based on the learning raw data read from the learning raw data storage unit 21 (step S22).

[0025] The composite data generating unit 22 generates composite data, for example, by specifying a privacy jet in advance and taking differential privacy into consideration.

[0026] The generated composite data is output to the outlier composite data removal unit 27 .

[0027] Examples of the timing at which the synthetic data generation unit 22 generates synthetic data include when raw data for learning is stored in the raw data for learning storage unit 21, when the raw data for learning stored in the raw data for learning storage unit 21 is updated, and at any predetermined time interval. An example of the predetermined time interval is T hours, where T is an integer.

[0028] The timing at which the synthetic data generation unit 22 generates synthetic data may be every time the learning raw data stored in the learning raw data storage unit 21 is updated, or may be once every several updates of the learning raw data stored in the learning raw data storage unit 21. In this way, the timing at which the synthetic data generation unit 22 generates synthetic data may be determined taking into consideration the timing at which the learning raw data stored in the learning raw data storage unit 21 is updated.

[0029] The predetermined time interval may also be determined taking into consideration the timing at which the learning raw data stored in the learning raw data storage unit 21 is updated.

[0030] <Defective Combined Data Removal Unit 27> The defective combined data removal unit 27 receives the composite data generated by the composite data generation unit 22. The defective combined data removal unit 27 also reads the learning raw data from the learning raw data storage unit 21.

[0031] The outlier composite data removal unit 27 compares the distribution of the learning raw data with the distribution of the composite data, and removes outlier composite data, which is composite data that deviates from the distribution of the learning raw data, from the composite data (step S27).

[0032] The composite data from which the outlier composite data has been removed is stored in the composite data storage unit 23 .

[0033] The outlier composite data removal unit 27, for example, divides the learning raw data and composite data into multiple clusters by clustering them, and removes composite data included in clusters where the proportion of composite data is equal to or greater than a predetermined proportion from the composite data as outlier composite data.

[0034] An example of the predetermined ratio is a real number c between 0 and 1 inclusive.

[0035] For example, suppose that c = 1, the number of synthetic data included in a certain cluster is 100, and the number of raw learning data included in the certain cluster is 0. In this case, the proportion of synthetic data included in the certain cluster is 1, which is greater than or equal to c (= 1), and therefore the synthetic data included in the certain cluster is determined to be outlier synthetic data.

[0036] The outlier combined data removal unit 27 performs this process for each cluster to determine outlier combined data, and then removes the determined outlier combined data from the combined data.

[0037] In addition, the outlier composite data removal unit 27 may, for example, divide the learning raw data and the synthetic data into multiple clusters by clustering them, and remove synthetic data included in a cluster in which the number of learning raw data is below a predetermined threshold from the synthetic data as outlier composite data.

[0038] For example, suppose the number of synthetic data included in a certain cluster is 100, the number of raw learning data included in the certain cluster is 1, and the predetermined threshold is 5. In this case, the number of raw learning data included in the certain cluster is 1, which is less than the predetermined threshold (= 5), so the synthetic data included in the certain cluster is determined to be outlier synthetic data.

[0039] The outlier combined data removal unit 27 performs this process for each cluster to determine outlier combined data, and then removes the determined outlier combined data from the combined data.

[0040] <Composite Data Storage Unit 23> The composite data storage unit 23 stores composite data from which the outlier composite data has been removed by the outlier composite data removal unit 27.

[0041] In this way, by removing the outlier synthetic data, the accuracy of the model generated from the synthetic data excluding the outlier synthetic data becomes higher because the distribution of the synthetic data excluding the outlier synthetic data becomes closer to the distribution of the raw training data compared to the distribution of the synthetic data before the outlier synthetic data is removed.

[0042] Therefore, the above embodiment makes it possible to generate training data (synthetic data excluding outlier synthetic data) that can generate a model with higher accuracy.

[0043] In this example, the synthetic data from which the outlier synthetic data has been removed is used as training data. For example, the model providing device 3 performs training based on the synthetic data from which the outlier synthetic data has been removed to generate a model (step S3). In this example, the model providing device 3 receives the synthetic data stored in the synthetic data storage unit 23, performs training based on the received synthetic data, and generates a model.

[0044] The model providing device 3 may receive the composite data using an encrypted communication method. This can prevent leakage of the composite data. Of course, the communication method used by the model providing device 3 when receiving the composite data is not limited to an encrypted communication method.

[0045] If the training data processing device 2 is equipped with a model storage unit 24 indicated by a dotted line in FIG. 1, the generated model is transmitted to the training data processing device 2 and stored in the model storage unit 24.

[0046] The model providing device 3 may transmit the model using an encrypted communication method. This can prevent the model from being leaked. Of course, the communication method used by the model providing device 3 when transmitting the model is not limited to an encrypted communication method.

[0047] In addition, reception of the synthetic data by the model providing device 3 may be realized by allowing the operator of the model providing device 3 to view the synthetic data stored in the synthetic data storage unit 23 via a web browser.

[0048] A user device 4 that wishes to use a model transmits data to be input into the model to the training data processing device 2. The training data processing device 2 obtains an output result when the received data is input into the model read from the model storage unit 24. The obtained output result is transmitted to the user device 4. In this way, the user device 4 can use the model stored in the model storage unit 24.

[0049] [First Modification] A label providing device 5 shown by a broken line in FIG. 1 may be provided.

[0050] In this case, the composite data read from the composite data storage unit 23 is sent to the label providing device 5 .

[0051] The composite data may be transmitted to the label provision device 5 using an encrypted communication method. This can prevent leakage of the composite data. Of course, the communication method used when transmitting the composite data to the label provision device 5 is not limited to the encrypted communication method.

[0052] The label providing device 5 assigns labels to the synthetic data received from the training data processing device 2 (step S5). The labeled synthetic data is transmitted to the model providing device 3. The labeled synthetic data is sometimes called training data.

[0053] The labeled composite data may be transmitted to the training data processing device 2 and stored in the composite data storage unit 23. This allows the labeled composite data to be managed by the training data processing device 2. Of course, the labeled composite data may be managed independently by the label providing device 5.

[0054] The model providing device 3 performs learning using the labeled synthetic data to generate a model. The subsequent processing is the same as in the above embodiment.

[0055] In this way, the label providing device 5 assigns labels to the composite data, making it possible to perform supervised learning or semi-supervised learning.

[0056] [Second Modification] The learning data processing device 2 may include a presentation data extraction unit 25 indicated by the dashed line in FIG.

[0057] The presentation data extraction unit 25 extracts a portion of the composite data from which the defective composite data has been removed (step S25). To this end, the presentation data extraction unit 25 extracts a portion of the composite data stored in the composite data storage unit 23, for example.

[0058] The presentation data extraction unit 25 may extract a portion of the composite data randomly, or may extract a portion of the composite data in consideration of at least one of the distribution of past composite data stored in the composite data storage unit 23 and the distribution of the current composite data. For example, the presentation data extraction unit 25 may extract a portion of the composite data in consideration of a comparison between the distribution of past composite data stored in the composite data storage unit 23 and the distribution of the current composite data.

[0059] Furthermore, the number of pieces of composite data extracted by the presentation data extraction unit 25 may be a constant, or may be dynamically determined, for example, taking into consideration the privacy budget and / or the current distribution of composite data.

[0060] By extracting a portion of the synthetic data by the presentation data extraction unit 25, it is possible to increase the uncertainty of the distribution of the synthetic data, making it more difficult to estimate the raw data for training. In other words, by not providing all of the synthetic data to the model providing device 3 or the label providing device 5, it is possible to reduce the possibility of the raw data for training being leaked.

[0061] Furthermore, by having the presentation data extraction unit 25 extract a portion of the synthetic data so that the change in data distribution before and after the update is sufficiently small, differential privacy can be ensured even between datasets of synthetic data without adding any or only a small amount of noise for differential privacy when generating the synthetic data.

[0062] In addition, the presentation data extraction unit 25 may extract a portion of the synthetic data from which outlier synthetic data has been removed, so as to improve the accuracy of the model generated by learning using the extracted portion of synthetic data (step S25).

[0063] For example, the presentation data extraction unit 25 extracts a portion of the synthetic data from which the outlier synthetic data has been removed, which improves the accuracy of the model, based on a score calculated based on the distribution of the synthetic data from which the outlier synthetic data has been removed.

[0064] In addition, the presentation data extraction unit 25 may extract a portion of the synthetic data from which the outlier synthetic data has been removed, which will improve the accuracy of the model, based on a score calculated based on the output when the portion of the synthetic data from which the outlier synthetic data has been removed is input into a model currently being trained.

[0065] Scores can be, for example, traditional ones based on the uncertainty of inference results using entropy, etc., ones that express uncertainty statistically in a Bayesian manner using mutual information, etc., ones that focus on data distribution, ones based on kernel density estimation or clustering, or combinations of these.

[0066] The presentation data extraction unit 25 may include, in the extracted portion of synthetic data, both synthetic data extracted randomly and synthetic data that is expected to improve the accuracy of the model based on the score.

[0067] The extracted composite data is transmitted to, for example, the model providing device 3. In this case, the model providing device 3 performs learning based on the composite data extracted by the presentation data extraction unit 25 to generate a model. The subsequent processing is the same as in the above embodiment.

[0068] Note that a model may be generated by repeating the extraction of a portion of the synthetic data by the presentation data extraction unit 25 and the generation of a model by the model providing device 3 using the extracted portion of the synthetic data. This repetitive process is executed until a preset condition is satisfied. Examples of the preset condition include a condition that the number of repetitions reaches a predetermined number, and a condition that the output of the model reaches a predetermined accuracy. Note that a portion of the synthetic data extracted from the second time onward may include synthetic data extracted previously.

[0069] 1 is provided, the composite data extracted by the presentation data extraction unit 25 may be transmitted to the label provision device 5. In this case, the label provision device 5 assigns a label to the composite data extracted by the presentation data extraction unit 25. The subsequent processing is the same as in the first modified example. In this way, the model providing device 3 may perform learning using the composite data to which the label has been assigned by the label provision device 5, which assigns a label to the composite data extracted by the presentation data extraction unit 25, and generate a model.

[0070] In addition, the label providing device 5 may assign a label to the composite data generated by the composite data generation unit 22, and the labeled composite data may be stored in the composite data storage unit 23, and the presentation data extraction unit 25 may extract a portion of the labeled composite data stored in the composite data storage unit 23.

[0071] The extracted synthetic data is transmitted to, for example, the model providing device 3. In this case, the model providing device 3 performs learning and generates a model based on the labeled synthetic data extracted by the presentation data extraction unit 25. The subsequent processing is the same as in the above embodiment.

[0072] The learning raw data providing device 1, the model providing device 3, and the label providing device 5 are separate devices.

[0073] [Third Modification] The learning data processing device 2 may include a model learning management unit 26 indicated by a two-dot chain line in FIG.

[0074] The model learning management unit 26 uses the raw learning data and / or the synthetic data to perform additional learning of the model generated by the model providing device 3 (step S26). To this end, the model learning management unit 26 performs additional learning of the model stored in the model storage unit 24, for example, using the raw learning data stored in the raw learning data storage unit 21 and / or the synthetic data stored in the synthetic data storage unit 23.

[0075] The additional learning using raw training data is additional learning using all or part of the raw training data. Similarly, the additional learning using synthetic data is additional learning using all or part of the synthetic data.

[0076] In this way, by having the model learning management unit 26 perform additional learning, it is possible to secure a sufficient amount of learning data while confining most of the data within the data sandbox, thereby improving the performance of the model.

[0077] [Other Modifications] The specific configurations of the embodiments of the disclosed technology are not limited to the configurations described above. The specific configurations of the embodiments of the disclosed technology can be appropriately modified in design, etc., within the scope of the spirit of the embodiments of the disclosed technology.

[0078] The various processes described in the embodiments of the disclosed technology may not only be performed chronologically in the order described, but may also be performed in parallel or individually depending on the processing capacity of the device performing the processes or as needed.

[0079] For example, a data sandbox does not need to be created in the training data processing device 2. That is, the processing of each unit of the training data processing device 2 (the training raw data storage unit 21, the composite data generation unit 22, the composite data storage unit 23, the model storage unit 24, the presentation data extraction unit 25, the model learning management unit 26, and the outlier composite data removal unit 27) does not need to be performed in the data sandbox.

[0080] The learning raw data storage unit 21 may be provided outside the learning raw data providing device 1, as exemplified in Fig. 3. In other words, the learning raw data providing device 1 does not need to be provided with the learning raw data storage unit 21. In this case, the learning raw data is transmitted from the learning raw data storage unit 21 to the learning data processing device 2, for example, using an encrypted communication method. Of course, the communication method used when transmitting the learning raw data from the learning raw data storage unit 21 to the learning data processing device 2 is not limited to an encrypted communication method. In the example of Fig. 3 as well, there may be multiple learning raw data providing devices 1.

[0081] The model providing device 3 may be made up of a plurality of devices. Similarly, the label providing device 5 may be made up of a plurality of devices.

[0082] For example, data may be exchanged directly between the components of the learning data processing device, or may be exchanged via a storage unit (not shown).

[0083] The label providing device 5 and the model providing device 3 may be the same device. In this case, for example, the labeled synthetic data and the model are transmitted to the training data processing device 2 at the same time.

[0084] The user device 4 may be a device different from the learning raw data providing device 1, the model providing device 3, and the label providing device 5, or may be the same device.

[0085] Each of the learning raw data providing device 1, the model providing device 3, the label providing device 5, and the user device 4 may be configured from multiple devices.

[0086] At least one of the label providing device 5 and the model providing device 3 may be able to access the raw data for learning.

[0087] [Program, Recording Medium] The processing of each unit of each of the above-mentioned devices may be realized by a computer, in which case the processing content of the functions to be possessed by each device is described by a program. Then, by loading this program into storage unit 1020 of computer 1000 shown in Figure 7 and operating arithmetic processing unit 1010, input unit 1030, output unit 1040, display unit 1060, etc., various processing functions of each of the above-mentioned devices are realized on the computer.

[0088] The program describing the processing contents can be recorded on a computer-readable recording medium, such as a non-transitory recording medium, specifically a magnetic recording device, an optical disk, or the like.

[0089] The program may be distributed, for example, by selling, transferring, lending, etc. a portable recording medium such as a DVD or CD-ROM on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to another computer via a network, thereby distributing the program.

[0090] A computer executing such a program, for example, first stores the program recorded on a portable recording medium or transferred from a server computer in its own non-transitory storage device, auxiliary storage unit 1050. Then, when executing a process, the computer loads the program stored in auxiliary storage unit 1050, its own non-transitory storage device, into storage unit 1020 and executes processing in accordance with the loaded program. Alternatively, as another form of execution of this program, the computer may load the program directly from a portable recording medium into storage unit 1020 and execute processing in accordance with the program. Furthermore, the computer may execute processing in accordance with the received program each time a program is transferred from a server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the processing function by issuing an execution instruction and obtaining the results, thereby executing the above-described processing through a so-called ASP (Application Service Provider) type service. Note that, in this embodiment, the program includes information used for processing by a computer that is equivalent to a program (e.g., data that is not a direct instruction to the computer but has properties that define computer processing).

[0091] In this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing may be realized by hardware. For example, each of the learning raw data storage unit 21, the composite data generation unit 22, the composite data storage unit 23, the model storage unit 24, the presentation data extraction unit 25, and the model learning management unit 26 may be configured by a processing circuit.

[0092] It goes without saying that other modifications are possible without departing from the spirit of the present invention.

[0093] All publications, patent applications, and technical standards mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, or technical standard was specifically and individually indicated to be incorporated by reference.

Claims

1. A training data processing device comprising: a synthetic data generation unit that generates synthetic data, which is pseudo-training data of at least one raw training data, based on the raw training data; and an outlier synthetic data removal unit that compares a distribution of the raw training data with a distribution of the synthetic data, and removes outlier synthetic data, which is synthetic data that deviates from the distribution of the raw training data, from the synthetic data.

2. A training data processing device according to claim 1, wherein the outlier synthetic data removal unit divides the raw training data and the synthetic data into a plurality of clusters by clustering, and removes synthetic data contained in a cluster in which the proportion of synthetic data is equal to or greater than a predetermined proportion from the synthetic data as the outlier synthetic data.

3. A training data processing device according to claim 1, wherein the training raw data cannot be viewed from any device other than the training raw data providing device without permission from the training raw data providing device that provides the at least one training raw data.

4. A learning data processing device according to claim 1 or 3, further comprising a presentation data extraction unit that extracts a portion of the synthetic data from which the outlier synthetic data has been removed, and the portion of the synthetic data extracted by the presentation data extraction unit is used for learning.

5. A training data processing device according to claim 4, wherein the presentation data extraction unit extracts a portion of the synthetic data from which the outlier synthetic data has been removed so as to improve the accuracy of a model generated by training using the extracted portion of the synthetic data.

6. A learning data processing method including: a synthetic data generation step in which a synthetic data generation unit generates synthetic data, which is pseudo-learning data of at least one raw learning data, based on the raw learning data; and an outlier synthetic data removal step in which an outlier synthetic data removal unit compares the distribution of the raw learning data with the distribution of the synthetic data, and removes outlier synthetic data, which is synthetic data that deviates from the distribution of the raw learning data, from the synthetic data.

7. A program for causing a computer to function as each part of the learning data processing device of claim 1.

Images