Machine learning method, machine learning system and program

The method generates pseudo-real data using a conditional generative model and a discriminator to protect against MI attacks, ensuring secure machine learning by distinguishing real from simulated data.

JP7582290B2Active Publication Date: 2024-11-13NEC CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2022504874
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-03-05
Publication Date
2024-11-13
Estimated Expiration
2040-03-05

AI Technical Summary

Technical Problem

As machine learning becomes more widespread, devices with inference algorithms and learned parameters are increasingly accessible, making them vulnerable to Membership Inference (MI) attacks that can leak secret information.

Method used

A machine learning method using a conditional generative model generates pseudo-real data corresponding to input labels, and a discriminator distinguishes between real and pseudo-real data to enhance security against MI attacks.

Benefits of technology

The method effectively protects real data from MI attacks by making it difficult for attackers to distinguish between genuine and simulated data, thereby enhancing security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007582290000004
    Figure 0007582290000004
  • Figure 0007582290000005
    Figure 0007582290000005
  • Figure 0007582290000006
    Figure 0007582290000006
Patent Text Reader

Abstract

The present invention enhances security against membership inference attacks (MI attacks), which infer data that was used for machine learning. Pseudo-real data corresponding to a random label is generated by entering the label into a generator G that uses a conditional generation model which is created using real data as learning data, and which outputs pseudo-real data corresponding to an input label. The labeled pseudo-real data output from this generator G is used to perform machine learning and create a learning model.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a machine learning method, a machine learning system, and a program. [Background technology]

[0002] Membership Inference attacks (MI attacks) are known that leak confidential information (e.g., customer information, trade secrets, etc.) used in learning from learned parameters of machine learning. For example, Non-Patent Document 1 discloses a method of MI attack under the condition that the inference algorithm is accessible. MI attacks are carried out by utilizing a phenomenon called "overlearning" in machine learning. Overlearning is a phenomenon in which machine learning over-fits to data used in learning, which causes the output tendency to differ when data used in learning is input as input to the inference algorithm and when data that is not used in learning is input. Attackers in MI attacks exploit this difference in tendency to determine whether data at hand is used for learning or not.

[0003] Non-Patent Document 2 is a document that proposes a learning algorithm that is resistant to MI attacks.

[0004] Non-Patent Document 3 discloses a method called MemGuard, which performs processing to mislead the attacker's classifier, as a countermeasure against such black box attacks. Non-Patent Document 4 discloses a countermeasure against what could be called a gray box attack, in which the attacker knows the inference algorithm. Non-Patent Document 5 discloses a method of training using a trained learning machine that has been trained using secret information and public data that is completely different from the secret information to be trained.

[0005] Non-Patent Documents 6 to 9 and Patent Document 1 are documents that disclose conditional generative models that, contrary to inference machines, generate fictitious data (pseudo-real data) of a specified label. Non-Patent Document 6 discloses a method called CGAN, Non-Patent Document 7 discloses a method called InfoGAN, Non-Patent Document 8 discloses a method called ACGAN, and Non-Patent Document 9 discloses a method called Conditional VAE. Patent Document 2 discloses the above-mentioned CGAN (referred to as C-GAN), as well as auxiliary classifier CGAN (AC-GN) and semi-supervised stacked GAN (SS-GAN). Patent Document 2 also proposes an improvement to GAN. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2019-91440 A [Patent Document 2] JP 2019-56975 A [Non-patent literature]

[0007] [Non-Patent Document 1] Reza Shokri, Marco Stronati, Congzheng Song, Vitaly Shmatikov: “Membership Inference Attacks Against Machine Learning Models” , IEEE Symposium on Security and Privacy 2017: 3-18, [online], [Retrieved February 24, 2020], Internet〈URL:https: / / arxiv.org / pdf / 1610.05820.pdf〉 [Non-Patent Document 2] Milad Nasr, Reza Shokri, Amir Houmansadr, “Machine Learning with Membership Privacy using Adversarial Regularization”, [online], [Retrieved February 24, 2020], Internet〈URL:https: / / arxiv.org / pdf / 1807.05852.pdf〉 [Non-Patent Document 3] Jinyuan Jia, Ahmed Salem, Michael Backes, Yang Zhang, Neil Zhenqiang Gong, "MemGuard:Defending against Black-Box Membership Inference Attacks via Adversarial Examples", [online], [Retrieved February 24, 2020], Internet〈URL:https: / / arxiv.org / pdf / 1909.10594.pdf〉 [Non-Patent Document 4] N. Papernot, M. Abadi, U. Erlingsson, I. Goodfellow, and K. Talwar, "Semi-supervised knowledge transfer for deep learning from private training data", [online], [Retrieved February 24, 2020], Internet〈URL:https: / / arxiv.org / pdf / 1610.05755.pdf〉 [Non-Patent Document 5] Sravanti Addepalli, Gaurav Kumar Nayak, Anirban Chakraborty, R. Venkatesh Babu, "DeGAN : Data-Enriching GAN for Retrieving Representative Samples from a Trained Classifier", [online], [Retrieved February 24, 2020], Internet <URL: https: / / arxiv.org / pdf / 1912.11960.pdf>

Non-licensed Document 6

Non-licensed Document 7

Non-licensed Document 8

Non-licensed literature 9

[0008] The following analysis is provided by the inventor. As machine learning becomes more widespread, devices equipped with inference algorithms and learned parameters, and terminals that can access them, will be placed in various places in the future. For this reason, MI attacks in the form of the above-mentioned black box or even gray box may be carried out, and secret information may be inferred.

[0009] An object of the present invention is to provide a machine learning method, a machine learning system, and a program that can contribute to improving security against the above-mentioned MI attacks. [Means for solving the problem]

[0010] According to a first aspect, a machine learning method is provided, which includes inputting a random label to a generator G using a conditional generative model that is created using real data as learning data and outputs pseudo-real data corresponding to the input label, generating pseudo-real data corresponding to the label, and performing machine learning using the labeled pseudo-real data output from the generator G to create a learning model.

[0011] According to a second aspect, a machine learning method is provided that includes a generator G using a conditional generative model that is created using real data as learning data and outputs pseudo-real data corresponding to an input label, an inference unit F that performs inference on input data using a learning model created by performing machine learning using the pseudo-real data output from the generator, and a discriminator D that identifies whether the input data is real data or pseudo-real data. First, either pseudo-real data corresponding to the label obtained by inputting a random label to the generator G or learning data using real data is selected, and input to the inference unit F to output an inference result. Next, data inferred by the inference unit F and an inference result of the learning model are input to the discriminator D, and a process of discriminating whether the data inferred by the inference unit F is real data or pseudo-real data is repeated. In the process of the repetition, the parameters of the learning model of the inference unit F are updated using the pseudo-real data output from the generator G with the updated parameters. Independently from the update of the parameters of the learning model of the inference unit F, the parameters of the conditional generative model of the generator G are updated using the classification result of the classifier D so that classification by the classifier D becomes more difficult, and the parameters of the classifier D are updated based on whether the classification result by the classifier D is correct, thereby updating the learning model of the classifier D. Each of the above methods is linked to a specific machine, that is, a computer that performs machine learning using training data and real data and its labels as teaching data.

[0012] According to a third aspect, a machine learning system capable of implementing each of the above-mentioned methods is provided.

[0013] According to a fourth aspect, there is provided a computer program for implementing the functions of the computer described above. The program is input to the computer device from an input device or from the outside via a communication interface, stored in a storage device, drives the processor according to a predetermined step or process, and can display the process result including an intermediate state at each stage as necessary via a display device, or can communicate with the outside via the communication interface. For example, a computer device for this purpose typically includes a processor, a storage device, an input device, a communication interface, and a display device as necessary, which are connectable to each other via a bus. Moreover, this program can be recorded on a computer-readable (non-transitive) storage medium. Effect of the Invention

[0014] According to the present invention, it is possible to carry out machine learning in a form that improves security against the above-mentioned MI attacks. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram illustrating a device configuration capable of implementing a machine learning method according to a first embodiment of the present invention. [Diagram 2] FIG. 11 is a diagram illustrating a device configuration capable of implementing a machine learning method according to a second embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram showing the configuration of a computer constituting a terminal device or a server of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] [First embodiment] First, a first embodiment of the present invention will be described with reference to the drawings. Note that the reference symbols in the drawings attached to the following embodiments are attached to each element for convenience as an example to aid understanding, and are not intended to limit the present invention to the illustrated form. Also, the connection lines between blocks in the drawings and the like referred to in the following description include both bidirectional and unidirectional. A unidirectional arrow is a schematic representation of the flow of a main signal (data), and does not exclude bidirectionality. A program is executed via a computer device, and the computer device includes, for example, a processor, a storage device, an input device, a communication interface, and a display device as necessary. Also, this computer device is configured to be able to communicate with devices inside the device or external devices (including computers) via the communication interface, regardless of whether they are wired or wireless. Also, although there are ports or interfaces at the connection points of input and output of each block in the drawings, they are not shown. Also, in the following description, "A and / or B" is used to mean A or B, or A and B.

[0017] 1, the machine learning method according to the first embodiment can be implemented by a machine learning system 100 including a generator G and a parameter learning unit 10. The machine learning system 100 can also be configured using a computer capable of operating as each of the generator G and the parameter learning unit 10.

[0018] More specifically, the generator G is configured using a conditional generative model that is created using learning data using real data and outputs pseudo real data corresponding to an input label. Such a generator G can be configured using various GANs described in Non-Patent Documents 6 to 9 and Patent Document 1.

[0019] The parameter learning unit 10 performs machine learning using the input labeled data to create a learning model. Specifically, the learning model is created by setting the inference parameters output by the parameter learning unit 10 in Fig. 1 in the learning model F (inference unit) in the lower part of Fig. 1. The created learning model F (inference unit) is then used to infer inference target data, as shown in the lower part of Fig. 1. Hereinafter, parameters typified by the weights of a neural network will be simply referred to as "parameters".

[0020] In the machine learning system of this embodiment, a random label R is input to the generator G to generate pseudo-real data corresponding to the label. Then, the parameter learning unit 10 performs machine learning to create a learning model using the labeled pseudo-real data output from the generator G. Specifically, the learning model is created by setting the inference parameters output by the parameter learning unit 10 in Fig. 1 to the learning model F (inference unit) in the lower part of Fig. 1.

[0021] For example, if the real data is a face image of a person and the label is gender, the generator G generates a non-existent face image of either male or female as pseudo real data. The parameter learning unit 10 then performs machine learning using the face image output from the generator G with either male or female set as the label to calculate inference parameters.

[0022] According to the machine learning method of the first embodiment described above, machine learning is performed using pseudo real data that is different from real data. Therefore, even if repeated access to the learning model F (inference unit) in the lower part of FIG. 1 occurs or learned parameters (inference parameters) are leaked, the attacker who uses the MI attack can only obtain the result of estimating the pseudo real data. Therefore, real data is protected from the MI attack.

[0023] [Second embodiment] Next, a second embodiment in which some elements are added to the configuration of the first embodiment to enable creation of a more accurate learning model will be described in detail with reference to the drawings.

[0024] Fig. 2 is a diagram showing a device configuration capable of implementing the machine learning method of the second embodiment of the present invention. Referring to Fig. 2, a configuration including a generator G, a parameter learning unit 10a, a random selection unit 11, a learning model F (inference unit), a discriminator D, and a second parameter learning unit 12 is shown.

[0025] As in the first embodiment, the generator G is configured using a conditional generative model that is created by machine learning using real data as learning data and outputs pseudo real data corresponding to an input label.

[0026] The parameter learning unit 10a performs machine learning using the pseudo-real data generated by the generator G and its corresponding labels, and updates the inference parameters of the learning model (inference unit) F.

[0027] The random selection unit 11 randomly selects either the pseudo real data or the real data generated by the generator G. The random selection unit 11 inputs the selected data to a learning model F (inference unit) and a discriminator D, respectively.

[0028] The learning model F (inference unit) performs inference on the data input from the random selection unit 11, and outputs the inference result. The inference result by the learning model F (inference unit) is input to the classifier D.

[0029] The discriminator D receives as input the label (correct inference data), the inference result of the learning model F (inference unit), and the data input from the random selection unit 11, and discriminates whether the data input from the random selection unit 11 is real data or pseudo-real data.

[0030] The second parameter learning unit 12 updates the parameters of the generator G and the discriminator D by using the pseudo real data generated by the generator G and its corresponding label.

[0031] Here, detailed operations of parameter learning unit 10a and second parameter learning unit 12 of the second embodiment will be described.

[0032] In the following, the loss functions for the learning of the discriminator D, the generator G, and the learning model F (inference part) are respectively L D , L G , L F Let us assume that.

[0033] Moreover, the parameters of the discriminator D, the generator G, and the learning model F (inference unit) are denoted as ω, τ, and θ, respectively.

[0034] In addition, ∇ in the formulas described below represents the nabla (= differential for each component). m D ,m G ,m F ,k D ,k G ,k F represents a predetermined constant.

[0035] The parameter learning unit 10a and the second parameter learning unit 12 perform learning according to the following algorithm. (1) A set S consisting of pairs of real data or pseudo-real data and their labels is received as input. (2) Initialize the parameters ω, τ, and θ. (3) Then, learning of ω, θ, and τ is performed. (3-1) The following are k D Repeat times. From the set S, select the data and its label pairs (x1, y1),...,(x mD ,y mD ) is randomly selected. This process corresponds to the operation of the random selection unit 11 described above. Random numbers z1,…z mD Randomly choose i=1, ,m D Whereas, x bari =G(τ,z i ,y i ) is calculated by using the current parameter τ and inputting z i ,y i Enter pseudo real data x bar i Note that the x in the following formula is expressed as x bar It is written as follows. The update amount Δω of the parameter ω is calculated by the following equation [Mathematical formula 1]. The parameter ω of the classifier D is updated using this update amount Δω.

[0036]

number

[0037]

number

[0038]

number

[0039] As can be seen from the above [Equation 2], according to the configuration of the second embodiment, the pseudo real data x bar i Therefore, it is impossible for information about real data to be leaked from the learning model F (inference part).

[0040] In addition, in the second embodiment, as is clear from [Mathematical Expression 1], the output F(θ, x bar i On the other hand, the generator G uses the output D(ω,x bar i ,y i ,F(θ,x bar i That is, the generator G in the second embodiment is strengthened to generate pseudo real data that is difficult to distinguish from real data even by a powerful classifier D that receives the output of F and attempts to classify it.

[0041] In the second embodiment, learning is performed using such "more realistic" pseudo-real data, and learned parameters θ of the learning model F (inference unit) are obtained. This is the difference from the first embodiment.

[0042] Although each embodiment of the present invention has been described above, the present invention is not limited to the above-mentioned embodiments, and further modifications, substitutions, and adjustments can be made without departing from the basic technical idea of ​​the present invention. For example, the system configuration, the configuration of each element, and the representation form of data, etc. shown in each drawing are examples to help understand the present invention, and the present invention is not limited to the configurations shown in these drawings.

[0043] For example, in the above embodiment, an example was given in which the real data is a person's face image and the inference device identifies the gender, but as is clear from the above description, the scope of application of the present invention is not limited to this. For example, the present invention can also be applied to a configuration in which the inference device infers the age group of a person by setting the age group of the person as the label.

[0044] Similarly, various methods can be adopted for the learning algorithm of the learning model (inference unit) F in the above-described embodiment.

[0045] The procedure shown in the above-described embodiment can be realized by a program that causes a computer (9000 in FIG. 3) functioning as a machine learning system to realize the functions of the device. Such a computer is exemplified by a configuration including a CPU (Central Processing Unit) 9010, a communication interface 9020, a memory 9030, and an auxiliary storage device 9040 in FIG. 3. That is, the machine learning program and the encryption / decryption program are executed by the CPU 9010 in FIG. 3, and an update process of each calculation parameter held in the auxiliary storage device 9040 or the like is performed.

[0046] In other words, the learning process of the machine learning system shown in the above-mentioned embodiments and the inference process obtained by this learning can be realized by a computer program that causes a processor installed in these devices to execute each of the above-mentioned processes using its hardware.

[0047] The disclosures of the above patent documents and non-patent documents are incorporated herein by reference and may be used as the basis or part of the present invention as necessary. Within the scope of the entire disclosure of the present invention (including the scope of claims), modifications and adjustments of the embodiments and examples are possible based on the basic technical ideas. Furthermore, within the scope of the disclosure of the present invention, various combinations or selections (including partial deletions) of various disclosed elements (including each element of each claim, each element of each embodiment or example, each element of each drawing, etc.) are possible. In other words, the present invention naturally includes various modifications and corrections that a person skilled in the art would be able to make according to the entire disclosure including the scope of claims and the technical ideas. In particular, with regard to the numerical ranges described in this document, any numerical value or small range included within the range should be interpreted as being specifically described even if not otherwise specified. Furthermore, the disclosures of the above cited documents may be used in part or in whole in combination with the disclosures of this document as part of the disclosure of the present invention in accordance with the spirit of the present invention as necessary, and are deemed to be included in the disclosures of this application. [Explanation of symbols]

[0048] G generator F Learning model (inference part) D Discriminator 10, 10a Parameter learning section 12 Second parameter learning section 11 Random Selection Section 100 Machine Learning Systems 9000 Computers 9010 CPU 9020 Communication Interface 9030 Memory 9040 Auxiliary storage device

Claims

1. A generator G using a conditional generative model that is created using real data as learning data and outputs pseudo real data corresponding to an input label; An inference unit F that performs inference on input data using a learning model created by machine learning using only the pseudo-real data output from the generator; A classifier D that classifies input data as real data or pseudo real data; A machine learning method using Selecting either pseudo real data corresponding to a label obtained by inputting a random label to the generator G or learning data using real data, inputting the selected pseudo real data to the inference unit F and outputting an inference result; The data on which the inference unit F performed the inference and the inference result of the learning model are input to the classifier D, and a process of classifying whether the data on which the inference unit F performed the inference is real data or pseudo real data is repeated; updating parameters of a learning model of the inference unit F using the pseudo-real data output from the generator G with updated parameters; Using the classification result of the classifier D, update the parameters of the conditional generative model of the generator G so that the classification by the classifier D becomes more difficult; and updating a learning model of the classifier D by updating parameters of the classifier D based on whether the classification result by the classifier D is correct or not; Machine learning methods.

2. A generator G using a conditional generative model that is created using real data as learning data and outputs pseudo real data corresponding to an input label; An inference unit F that performs inference on input data using a learning model created by machine learning using only the pseudo-real data output from the generator; and a discriminator D for discriminating whether input data is real data or pseudo real data; Selecting either pseudo real data corresponding to a label obtained by inputting a random label to the generator G or learning data using real data, inputting the selected pseudo real data to the inference unit F and outputting an inference result; The data on which the inference unit F performed the inference and the inference result of the learning model are input to the classifier D, and a process of classifying whether the data on which the inference unit F performed the inference is real data or pseudo real data is repeated; updating parameters of a learning model of the inference unit F using the pseudo-real data output from the generator G with updated parameters; Using the classification result of the classifier D, update the parameters of the conditional generative model of the generator G so that the classification by the classifier D becomes more difficult; and updating a learning model of the classifier D by updating parameters of the classifier D based on whether the classification result by the classifier D is correct or not; Machine learning systems.

3. A generator G using a conditional generative model that is created using real data as learning data and outputs pseudo real data corresponding to an input label; An inference unit F that performs inference on input data using a learning model created by machine learning using only the pseudo-real data output from the generator; A computer operable as a discriminator D for discriminating whether input data is real data or pseudo real data, A process of selecting either pseudo real data corresponding to a label obtained by inputting a random label to the generator G or learning data using real data, inputting the selected pseudo real data to the inference unit F, and outputting an inference result; the data on which the inference unit F has performed inference and the inference result of the learning model are input to the classifier D, and a process of classifying whether the data on which the inference unit F has performed inference is real data or pseudo real data is repeated; a process of updating parameters of a learning model of the inference unit F using pseudo real data output from the generator G with updated parameters, updating parameters of a conditional generative model of the generator G using a discrimination result of the discriminator D so that discrimination by the discriminator D becomes more difficult, and updating the parameters of the discriminator D based on whether the discrimination result by the discriminator D is correct, thereby updating the learning model of the discriminator D; A program that executes the following.

Citation Information

Patent Citations

  • Improved generative adversarial network achievement program, improved generative adversarial network achievement device, and learned model generation method

    JP2019056975A

  • System and method for semi-supervised conditional generation modeling using hostile network

    JP2019091440A