Learning device, data generation device, learning method, and learning program

Abstract

To provide a learning device, a data generating device, a learning method and a learning program capable of efficiently analyzing a tradeoff between factors relating to the reliability of a generator that has performed a learning using a hostile generating network.SOLUTION: In a data generating device 1, a hostile generating network is constructed so as to include: a generator G that generates processed data similar to true data on the basis of noises; and a plurality of identifiers D that is caused to compete with the generator G so as to satisfy each of a plurality of factors relating to the reliability of the processed data. A data generating device 1 includes a learning unit 11 which learns each identifier D so as to maximize the loss function on the basis of the identification precision of the identifier D and which learns the generator G so as to minimize the loss function using training data and the processed data. The loss function contains a weighting coefficient representing the importance of each of the plurality of factors. The generator G has the noises and the weighting coefficient as inputs. The learning unit 11 repeatedly learns the identifier D and the generator G while selecting the weighting coefficient in accordance with a predetermined distribution.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method for generating processed data using a generative adversarial network. [Background technology]

[0002] It has been known that a generator for generating processed data similar to true data can be constructed by learning the distribution of training data using a generative adversarial network (see, for example, Non-Patent Document 1). The generative adversarial network includes a generator that generates processed data from noise, and a discriminator that discriminates between true data and processed data. The generator and discriminator are each composed of a neural network, and the generator and discriminator are trained by competing with each other using true data, so that the generator can generate processed data similar to true data.

[0003] In addition, in recent years, there has been growing interest in the reliability of machine learning. For this reason, it is necessary to construct generative adversarial networks that also take into consideration requirements related to reliability, such as fairness and privacy. There is a method called FairGAN as a generative adversarial network that takes fairness into consideration (for example, see Non-Patent Document 2). In FairGAN, a generator is trained so that the distribution of processed data is similar between sensitive attributes and non-sensitive attributes.

[0004] An example of a generative adversarial network that takes privacy into consideration is privGAN (see, for example, Non-Patent Document 3). When a generative adversarial network is used, the processed data generated by a trained generator is more likely to be similar to the training data than the test data, and this property can be exploited to launch a membership inference attack to infer the training data. In privGAN, membership inference attacks are suppressed by training the generator so that the processed data generalizes to multiple training data sets, not just one training data set.

[0005] In both FairGAN and privGAN, in addition to the generator and the aforementioned classifier (classifier A), a second classifier (classifier B) is used to take into account various requirements regarding reliability, and the generator, classifier A, and classifier B are trained by competing with each other in the same way as in a normal generative adversarial network. However, a parameter for determining the priority of classifier A and classifier B is assigned as a weighting coefficient for classifier B. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] IJ Goodfellow et al., "Generative Adversarial Networks," NIPS 2014. [Non-Patent Document 2] X. Depeng et al., "FairGAN: Fairness-aware Generative Adversarial Networks," IEEE BigData 2018. [Non-Patent Document 3] S. Mukherjee et al., "privGAN: Protecting GANs from membership inference attacks at low cost to utility," PETS 2021. [Non-Patent Document 4] H. Chang and R. Shokri, "On the Privacy Risks of Algorithmic Fairness," CoRR abs / 2011.03731, 2020. Summary of the Invention [Problem to be solved by the invention]

[0007] However, there has been no discussion of a generative adversarial network that can simultaneously consider multiple requirements for the reliability of machine learning. It is also known that there are trade-offs between these requirements. For example, it is known that models constructed with consideration for fairness are vulnerable to membership inference attacks (see, for example, Non-Patent Document 4). For this reason, when considering multiple requirements simultaneously, it is necessary to consider not only a simple combination of methods that consider each requirement, but also the trade-off between requirements.

[0008] One way to solve this problem is to combine the various techniques by introducing parameters to set the importance of each requirement. However, simply giving the parameters as weighting coefficients requires retraining the generator every time the parameter value is changed. In this case, training a generative adversarial network is computationally expensive, making it difficult to perform trade-off analysis.

[0009] The present invention aims to provide a learning device, a data generation device, a learning method, and a learning program that can efficiently perform trade-off analysis between requirements regarding the reliability of a generator trained using a generative adversarial network. [Means for solving the problem]

[0010] The learning device of the present invention is configured to have a generative adversarial network including a generator that generates processed data similar to true data based on noise, and a plurality of classifiers that compete with the generator to satisfy each of a plurality of requirements regarding the reliability of the processed data, and is equipped with a learning unit that uses training data and the processed data to train each classifier to maximize a loss function based on the classification accuracy of the classifier, and trains the generator to minimize the loss function, the loss function including a weighting coefficient indicating the importance of each of the plurality of requirements, the generator receives the noise and the weighting coefficient as input, and the learning unit repeats learning of the classifiers and the generator while selecting the weighting coefficient according to a predetermined distribution.

[0011] One of the plurality of classifiers may be for satisfying a requirement regarding accuracy of the processed data.

[0012] The predetermined distribution may have a peak in an analysis range for each of the predefined weighting factors.

[0013] The data generating device according to the present invention inputs noise and a specified weighting coefficient to the generator trained by the learning device, and outputs the processed data.

[0014] In a learning method according to the present invention, a computer that configures a generative adversarial network including a generator that generates processed data similar to true data based on noise, and a plurality of classifiers that compete with the generator to satisfy each of a plurality of requirements regarding the reliability of the processed data executes a learning step in which, using training data and the processed data, each classifier is trained to maximize a loss function based on the classification accuracy of the classifier, and the generator is trained to minimize the loss function, the loss function including a weighting factor indicating the importance of each of the plurality of requirements, and the generator receives the noise and the weighting factor as input, and in the learning step, repeats learning of the classifiers and the generator while selecting the weighting factor according to a predetermined distribution.

[0015] A learning program according to the present invention causes a computer to function as the learning device. Effect of the Invention

[0016] According to the present invention, it is possible to efficiently perform trade-off analysis between requirements regarding the reliability of a generator trained using a generative adversarial network. [Brief description of the drawings]

[0017] [Figure 1] FIG. 2 is a diagram illustrating a functional configuration of a data generating device according to an embodiment. [Diagram 2] 10 is a flowchart showing a processing algorithm in a learning phase of the data generating device according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] An example of an embodiment of the present invention will now be described. The data generation device of this embodiment controls multiple requirements related to the reliability of a generator trained by a generative adversarial network using a loss function that uses a weighting factor that indicates the importance of each requirement. By configuring the generative adversarial network so that these weighting factors are input to the generator, it is possible to generate processed data with specified weighting factors using the trained generator.

[0019] FIG. 1 is a diagram showing the functional configuration of a data generating device 1 (learning device) in this embodiment. The data generating device 1 is an information processing device (computer) including a control unit 10, a storage unit 20, and various input / output interfaces.

[0020] The control unit 10 is a part that controls the entire data generating device 1, and realizes each function in this embodiment by appropriately reading and executing various programs stored in the storage unit 20. The control unit 10 may be a CPU.

[0021] The storage unit 20 is a storage area for various programs for causing the hardware group to function as the data generating device 1, various data, and the like, and may be a ROM, a RAM, a flash memory, a hard disk drive (HDD), or the like. Specifically, the memory unit 20 stores a program (data generation program) for causing the control unit 10 to execute each function of this embodiment, as well as parameters of a learning model that constitutes a generative adversarial network, learning data, etc.

[0022] The control unit 10 generates a generator G and a discriminator D as a learning model based on various parameters stored in the storage unit 20. i (1≦i≦n), a learning unit 11 for learning these learning models, and an output unit 12 for outputting processed data based on noise. By operating these functional units, the data generating device 1 outputs processed data that is similar to true data and satisfies multiple reliability requirements.

[0023] The learning unit 11 uses the training data and the processed data to generate a classifier D i Each classifier D is assigned to maximize the loss function based on the classification accuracy of i We train a generator G to minimize this loss function. That is, the data generating device 1 includes a generator G that generates processed data similar to true data based on noise, and a plurality of discriminators D that compete with the generator G to satisfy each of a plurality of requirements regarding the reliability of the processed data. i A generative adversarial network including the above is constructed, and in a learning phase using the learning unit 11, a generator G is learned.

[0024] Weighting coefficients indicating the importance of each of the multiple factors constituting the loss function in machine learning are input to the generator G together with noise, and the learning unit 11 selects the weighting coefficients according to a predetermined distribution while training the discriminator D i Then, the training of the generator G is repeated.

[0025] The output unit 12 inputs noise and a weighting coefficient designated by a user to the generator G trained by the training unit 11, and outputs processed data. That is, in the generation phase using the output unit 12, the data generating device 1 generates and outputs processed data according to the weighting coefficients designated by the user using a trained generator G.

[0026] Here, the details of the generative adversarial network configured by the data generator 1 will be described in comparison with the conventional method. First, a conventional generative adversarial network is constructed as follows:

[0027] x is distributed as P data Let z be the true data that follows the distribution P z Let the noise follow the formula: G is a generator for generating the processed data x', and D is a discriminator for distinguishing between the real data and the processed data. The generator G and the discriminator D are constructed using a neural network.

[0028] By training generator G to satisfy the following optimization problem, generator G can generate processed data that is similar to the true data.

number

[0029] Here, the more similar the processed data generated using the generator G is to the true data, the more the processed data can be used in the same way as the true data, and the higher the accuracy (the higher the convenience). Therefore, the classifier D becomes a classifier for satisfying the reliability requirement regarding accuracy.

[0030] The data generating device 1 of this embodiment extends a conventional generative adversarial network, and generates processed data in consideration of n requirements (for example, fairness, privacy, etc.) related to the reliability of machine learning.

[0031] V1,…,V nAs constraints to realize each requirement, we set the generator G and the discriminators D1, ..., D n Here, V1 and D1 are the evaluation function and the discriminator related to the accuracy of the processed data. Also, r2,…,r n Let V2,…,V be the evaluation function that expresses the importance of each requirement other than accuracy. n The weighting coefficients are r2,…,r n are the distributions P r2 ,…,P rn shall be in accordance with the following.

[0032] In addition, the distribution of weighting coefficients P r2 ,…,P rn is preferably a probability distribution having a peak in the range of each predefined weighting coefficient, that is, in the range in which a trade-off analysis of the importance of each requirement is performed.

[0033] The generator G has a distribution P z In addition to the noise z that follows r2,…,r n In the learning phase, the data generation device 1 trains the generator G to satisfy the conditions of the optimization problem based on the following loss function, so that the generator G can generate processed data taking into account the importance of each requirement.

number

[0034] That is, for a trained generator G, the weighting coefficients r2,…,r n By specifying the value of , it is possible to generate processed data according to the importance of each requirement. To change the importance of each requirement, the weight coefficients r2, ..., r n By simply changing the value of , it is possible to generate processed data for different importance without retraining the generator G with new weighting coefficients.

[0035] FIG. 2 is a flowchart showing a processing algorithm in the learning phase of the data generating device 1 in this embodiment. Let X be a set of training data, and the data generating device 1 uses this algorithm to train a generator G so as to satisfy equation (1).

[0036] In step S1, the learning unit 11 generates a generator G and discriminators D1, . . . , D n Initialize.

[0037] In step S2, the learning unit 11 initializes the variable i to 1 and sets the weighting coefficients r2, ..., r n Each of these is distributed as P r2 ,…,P rn Select according to.

[0038] In step S3, the learning unit 11 selects a plurality of training data from the set of training data X. Furthermore, the learning unit 11 divides a plurality of noises z into a distribution P z and obtains the corresponding processed data from the generator G. Then, the learning unit 11 updates the discriminator D1 and the generator G using the obtained training data and processed data.

[0039] In step S4, the learning unit 11 adds 1 to the variable i, and then the multiple noises z are distributed as a distribution P z Select according to the corresponding D i The processing data required to update (i>1) is obtained from the generator G. The learning unit 11 uses the obtained processed data to train a classifier D i and update the generator G.

[0040] In step S5, the learning unit 11 checks whether i=n, ​​i.e., whether all D i If the determination is YES, the process proceeds to step S6, and if the determination is NO, the process returns to step S4.

[0041] In step S6, the learning unit 11 determines whether the process from step S2 to step S5 has been repeated a certain number of times. If the determination is YES, the process ends, and if the determination is NO, the process returns to step S2.

[0042] As described above, according to this embodiment, the data generating device 1 (learning device) constructs a generative adversarial network using a generator G for generating processed data and n classifiers D for satisfying reliability requirements, and trains the generator G and the n classifiers D by having them compete against each other using real data. At this time, the data generating device 1 assigns a parameter for specifying the importance of each requirement as a weighting coefficient to the loss function in machine learning, which makes it possible to generate processed data according to the importance of each requirement.

[0043] Furthermore, the generator G is configured to take weighting coefficients as input, and the data generating device 1 trains each classifier D to maximize the expected value of the loss function while selecting the weighting coefficients according to a predetermined distribution, and trains the generator G to minimize the expected value of the loss function. This allows the data generating device 1 to generate processed data according to different importance by reusing the trained generator G without retraining the generator G in response to changes in the weighting coefficients. As a result, it is possible to efficiently perform a trade-off analysis between requirements regarding the reliability of the generator G trained using the generative adversarial network.

[0044] In the generative adversarial network, one of the n classifiers (D1) is related to the accuracy of the processed data, and the other n-1 classifiers D i These concerns include requirements other than accuracy, such as fairness and privacy. This allows the data generating device 1 to generate highly accurate processed data similar to true data while simultaneously considering various requirements regarding reliability.

[0045] The data generating device 1 employs a continuous distribution for the probability distribution of the weighting coefficients in the loss function, the distribution having a peak in the range of values ​​for performing a trade-off analysis that is expected in advance, thereby enabling the data generating device 1 to appropriately train the generator G that uses the weighting coefficients as input.

[0046] This will, for example, improve the reliability of machine learning, making it possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0047] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Furthermore, the effects described in the above-described embodiments are merely a list of the most preferable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the embodiments.

[0048] The learning method and data generation method by the data generation device 1 are realized by software. When realized by software, a program constituting this software is installed in an information processing device (computer). These programs may be recorded on a removable medium such as a CD-ROM and distributed to users, or may be distributed by being downloaded to the user's computer via a network. Furthermore, these programs may be provided to the user's computer as a Web service via a network without being downloaded. [Explanation of symbols]

[0049] D Discriminator G generator 1 Data generation device (learning device) 10 Control section 11 Learning Department 12 Output section 20 Memory section

Claims

1. A generator that generates processed data similar to the true data based on noise; A generative adversarial network is constructed, the generative adversarial network including a plurality of classifiers that compete with the generator in order to satisfy each of a plurality of requirements regarding the reliability of the processed data; a learning unit that uses training data and the processed data to train each classifier so as to maximize a loss function based on the classification accuracy of the classifier, and to train the generator so as to minimize the loss function; the loss function includes a weighting coefficient indicating the importance of each of the plurality of requirements; The generator receives the noise and the weighting coefficients as input, The learning unit is a learning device that repeats learning of the classifier and the generator while selecting the weighting coefficients according to a predetermined distribution.

2. The learning device according to claim 1 , wherein one of the plurality of classifiers is adapted to satisfy a requirement regarding accuracy of the processed data.

3. 3. The learning device according to claim 1, wherein the predetermined distribution has a peak in an analysis range of each of the weighting coefficients defined in advance.

4. 4. A data generating device that inputs noise and a specified weighting coefficient to the generator trained by the learning device according to claim 1, and outputs the processed data.

5. A generator that generates processed data similar to the true data based on noise; A computer that configures an adversarial generative network including a plurality of classifiers that compete with the generator to satisfy each of a plurality of requirements regarding the reliability of the processed data, executing a learning step of learning each classifier so as to maximize a loss function based on the classification accuracy of the classifier, and learning the generator so as to minimize the loss function, using training data and the processed data; the loss function includes a weighting coefficient indicating the importance of each of the plurality of requirements; The generator receives the noise and the weighting coefficients as input, A learning method in which, in the learning step, learning of the classifier and the generator is repeated while selecting the weight coefficients according to a predetermined distribution.

6. A learning program for causing a computer to function as the learning device according to any one of claims 1 to 4.

Images