WO2022239200A1 - 学習装置、推論装置、学習方法、及びコンピュータ可読媒体 - Google Patents
学習装置、推論装置、学習方法、及びコンピュータ可読媒体 Download PDFInfo
- Publication number
- WO2022239200A1 WO2022239200A1 PCT/JP2021/018265 JP2021018265W WO2022239200A1 WO 2022239200 A1 WO2022239200 A1 WO 2022239200A1 JP 2021018265 W JP2021018265 W JP 2021018265W WO 2022239200 A1 WO2022239200 A1 WO 2022239200A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- learning
- data
- reasoner
- unit
- learning data
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 32
- 238000010801 machine learning Methods 0.000 claims abstract description 105
- 230000006870 function Effects 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 9
- 238000013500 data storage Methods 0.000 description 14
- 238000010586 diagram Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 238000004590 computer program Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 235000000332 black box Nutrition 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000007423 decrease Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Definitions
- This disclosure relates to machine learning.
- Non-Patent Documents 1 and 2 there is known a Membership Inference attack (MI attack) that leaks secret information (eg, customer information, trade secrets, etc.) used for learning from learned parameters of machine learning.
- MI attack Membership Inference attack
- Non-Patent Document 1 discloses a method of MI attack under the condition that access to the inference algorithm is possible. MI attacks are carried out using a phenomenon called "overfitting" in machine learning. Overfitting is a phenomenon in which machine learning excessively adapts to the data used for learning. Due to overfitting, the tendency of the output differs when the data used for learning is input to the input of the inference algorithm and when the other data is input. By exploiting this difference in tendency, MI attack attackers can determine whether the data at hand has been used for learning or not.
- Non-Patent Document 4 discloses a method called MemGuard. This method misleads the attacker's classifier as a countermeasure against black-box attacks under the condition that the learned parameters of the target's inference algorithm are not known.
- Non-Patent Document 5 discloses a learning algorithm that is resistant to MI attacks. Specifically, in Non-Patent Document 5, using any known machine learning inference algorithm f and a discriminator h that identifies whether the data input to f is the data used for learning f there is Each parameter is learned in an adversarial manner to raise the inference accuracy of the inference algorithm f and its resistance to MI attacks.
- the data used for learning may contain confidential information such as customer information and trade secrets.
- An MI attack may leak secret information used for learning from learned parameters of machine learning. For example, an attacker who illegally obtains learned parameters may guess learning data. Alternatively, even if the learned parameters are not leaked, an attacker can predict the learned parameters by repeatedly accessing the inference algorithm. Then, learning data may be predicted from predicted learned parameters.
- Non-Patent Document 4 is defended by adding noise to the inference result. Therefore, there is a problem that noise influences the inference result regardless of the defense performance.
- Non-Patent Document 5 there is a trade-off between accuracy and attack resistance. Specifically, parameters are set to determine the degree of trade-off between accuracy and attack resistance. Therefore, there is a problem that it is difficult to improve both accuracy and attack resistance.
- An object of the present disclosure is to provide an inference device, a learning device, a learning method, and a recording medium that are highly resistant to MI attacks and have high accuracy.
- a learning device includes a data dividing unit that generates n sets of divided data by dividing the first learning data by n (n is an integer equal to or greater than 2), and dividing the first learning data into one set.
- a learning data generation unit that generates second learning data by inputting data into the generating inference unit, and a learning unit that generates the inference unit by machine learning using the second learning data.
- the learning method generates n sets of divided data by dividing the first learning data by n (n is an integer of 2 or more), and removes one set of divided data from the first learning data. generating n learning data generating reasoners by machine learning using data, and inputting the set of divided data excluded from the machine learning to the n learning data generating reasoners, respectively; , the second learning data is generated, and the reasoner is generated by machine learning using the second learning data.
- a computer-readable medium is a computer-readable medium storing a program for causing a computer to execute a learning method, wherein the learning method stores first learning data as n (n is an integer equal to or greater than 2). ) generating n sets of divided data by dividing, generating n learning data generating reasoners by machine learning using data obtained by removing one set of divided data from the first learning data, Second learning data is generated by inputting the one set of divided data excluded from machine learning to each of the n learning data generation reasoners, and machine learning using the second learning data Generate an inferencer.
- a learning device an inference device, a learning method, and a computer-readable medium that are highly resistant to MI attacks and highly accurate.
- FIG. 1 is a block diagram showing the configuration of a machine learning system having a learning device;
- FIG. 4 is a flow chart showing the operation of the learning device;
- 4 is a flow chart showing the operation of the learning unit of the reasoner H;
- FIG. 11 is a block diagram showing the operation of a learning device according to another embodiment; It is a figure which shows the hardware constitutions of the apparatus concerning this Embodiment.
- FIG. 1 is a block diagram showing the configuration of a machine learning system including a learning device 100.
- the learning device 100 includes a data generation section 200 and a learning section 122 .
- learning data T are prepared in advance.
- the learning data T is also called member data, and data other than the learning data T is also called non-member data.
- the data generation unit 200 generates learning data for the reasoner H based on the learning data T.
- the learning data T prepared in advance will also be referred to as first learning data
- the learning data generated by the data generation unit 200 will also be referred to as second learning data.
- the learning unit 122 performs machine learning based on the second learning data generated by the data generation unit 200 . Thereby, the reasoner H is generated.
- the reasoner H is a machine learning model that makes inferences on input data.
- the reasoner H outputs an inference result when inference is made based on the input data.
- the reasoner H can be a classifier that performs image classification.
- the reasoner H outputs a score vector indicating the probability that it corresponds to each class.
- the learning data T is the first learning data, and is a data group including multiple data.
- the learning data T becomes a data set with a correct label (teacher data).
- the learning data T includes a plurality of pieces of input data, each of which is associated with a correct label.
- machine learning is not limited to supervised learning.
- the data generation unit 200 generates second learning data (training data) used for machine learning of the reasoner H.
- the data generation unit 200 includes a data division unit 220 , learning units 202 - 1 to 202 -n for F 1 to F n , and a learning data storage unit 250 .
- the data division unit 220 divides the learning data T into n (n is an integer equal to or greater than 2).
- the n-divided learning data are defined as divided data T 1 to T n . That is, the data dividing unit 220 divides the learning data T into n groups to generate n sets of divided data T 1 to T n . Assuming that the learning data T is one data set, each of the divided data T 1 to T n is a subset. As will be described later, each of the divided data T 1 -T n becomes input data of the inferencers F 1 -F n .
- Data sets included in the divided data T 1 to T n preferably do not overlap each other.
- data included in the divided data T 1 is preferably not included in the divided data T 2 to T n .
- the data included in the divided data T n is not included in the divided data T 1 to T n ⁇ 1 .
- the data division unit 220 equally divides the learning data T into n. Therefore, the divided data T 1 to T n contain the same number of data.
- the number of pieces of data included in the divided data T 1 to T n is not limited to be equal, and may be different.
- the data division unit 220 outputs some divided data extracted from the learning data T to the learning units 202-1 to 202-n.
- the data generation unit 200 extracts the learning data T ⁇ T 1 from the divided data T 1 to T n and inputs it to the learning unit 202-1 of F 1 .
- the learning data T ⁇ T 1 is a difference set obtained by excluding the divided data T 1 from the learning data T. That is, the learning data T ⁇ T 1 of F 1 includes T 2 to T n .
- the data generator 200 removes the divided data T1 from the learning data T to generate learning data T ⁇ T1.
- the learning unit 202-1 of F 1 performs machine learning to generate the reasoner F 1 using the learning data T ⁇ T 1 .
- the learning unit 202-1 trains the reasoner F1 based on the learning data T ⁇ T1.
- Various techniques such as supervised learning can be used for machine learning in the learning unit 202-1.
- a known method can be used for the machine learning of the learning unit 202-1, so a description thereof will be omitted.
- the learning unit 202-1 performs machine learning using all the data included in the learning data T ⁇ T1. Machine learning, for example, optimizes the parameters of each layer in a deep learning model. This produces the reasoner F1.
- the data generator 200 inputs the divided data T1 to the inference unit F1.
- the learning data storage unit 250 of the reasoner H stores the output of the reasoner F1 as H's learning data. That is, the inference result of the inference device F1 is stored in the memory or the like as learning data of the inference device H.
- FIG. The learning data of the reasoner H includes the inference result of the reasoner F1 when the divided data T1 is input to the reasoner F1. In this way, the learning data used during learning of the inference device F1 is different from the input data used during inference.
- the learning unit 202-n of F n performs machine learning to generate the reasoner F n using the learning data T ⁇ T n .
- the learning unit 202- n trains the reasoner F n based on the learning data T ⁇ T n.
- Various techniques such as supervised learning can be used for machine learning in the learning unit 202-n.
- a known method can be used for machine learning by the learning unit 202-n, so a description thereof will be omitted.
- the learning unit 202- n performs machine learning using all the data included in the learning data T ⁇ Tn. Machine learning, for example, optimizes the parameters of each layer in a deep learning model. This produces a reasoner Fn .
- the data generation unit 200 inputs the divided data Tn to the inference unit Fn .
- the learning data storage unit 250 of the reasoner H stores the output of the reasoner Fn as H's learning data. That is, the inference result of the inference device Fn is stored in the memory or the like as learning data of the inference device H.
- the learning data of the reasoner H includes the inference result of the reasoner Fn when the divided data Tn is input to the reasoner Fn. In this way, the learning data used during learning of the inference device Fn is different from the input data used during inference.
- the data generator 200 receives the entire set of training data T.
- FIG. The data dividing unit 220 divides the learning data T into n sets (n subsets) to generate divided data Ti.
- the learning unit of the data generation unit 200 machine-learns the inference device F i using the learning data T ⁇ T i .
- Learning data used for machine learning of the reasoner F i are T 1 to T i ⁇ 1 and T i+1 to T n .
- the reasoner F i makes an inference based on the divided data T i .
- the learning data storage unit 250 stores the inference result of the inference device F i as learning data.
- the learning units 202-1 to 202-n of F 1 to F n are learning data generation reasoner generation units that generate the reasoners F 1 to F n .
- the reasoners F 1 to F n can be machine learning models having a similar layer structure. That is, the reasoners F 1 to F n have the same number of layers, nodes, edges, and the like.
- Learning units 202-1 to 202-n generate reasoners F 1 to F n using different learning data. That is, the reasoners F 1 to F n are machine learning models generated using different learning data.
- the reasoners F 1 to F n are machine learning models that perform image classification and the like. In this case, the reasoners F 1 to F n output the same score vector as the reasoner H.
- the learning data storage unit 250 of the reasoner H stores the inference results of the reasoners F 1 , F 2 , . . . , F i , .
- the learning data storage unit 250 may store input data to the reasoners F 1 to F n and their inference results in association with each other.
- the learning data stored in the learning data storage unit 250 of the reasoner H becomes the second learning data as described above. Therefore, in the following description, the learning data stored in the learning data storage unit 250 of the reasoner H is also simply referred to as second learning data.
- the second learning data is a data set represented by Equation (1) below.
- the learning unit 122 of the reasoner H performs machine learning to generate the reasoner H using the second learning data.
- the learning unit 122 trains the reasoner H based on the second learning data.
- Machine learning in the learning unit 122 can use various techniques such as supervised learning.
- Machine learning by the learning unit 122 can be performed using a known technique, and thus description thereof is omitted.
- the learning unit 122 performs machine learning using all the data included in the second learning data. Machine learning, for example, optimizes the parameters of each layer in a deep learning model. Thereby, the reasoner H is generated.
- the learning unit 122 performs supervised learning using the inference result F i (x) for the input data x included in the divided data T i as the correct label.
- the inference result output from the reasoner H is expressed by the following equation (2).
- the data generating section 200 generates learning data for the reasoner H based on the outputs of the reasoners F 1 to F n .
- the reasoner H becomes a distilled model generated using the outputs of the reasoners F 1 to F n . That is, the reasoners F 1 to F n extract some information from the learning data T.
- the learning data storage unit 250 causes the reasoner H to learn using the information extracted by the reasoners F 1 to F n as learning data. Therefore, the reasoner H can obtain high estimation accuracy with a simple model.
- FIG. 2 is a flow chart showing the learning method according to this embodiment.
- step S201 the data generation unit 200 generates learning data for the reasoner H (S201).
- the processing of step S201 will be described in detail using FIG.
- FIG. 3 is a flow chart showing the process of generating learning data for the reasoner H.
- FIG. 3 is a flow chart showing the process of generating learning data for the reasoner H.
- the data division unit 220 divides the learning data T into n (S501). That is, the data dividing unit 220 generates divided data T 1 to T n .
- the learning units 202-1 to 202-n train the n reasoners F 1 to F n with the learning data excluding the divided data T 1 to T n (S502). That is, the learning unit of the data generation unit 200 machine-learns the reasoner F i using T ⁇ T i .
- the data generation unit 200 inputs the divided data not used for learning of the n reasoners F 1 to F n to the respective reasoners F 1 to F n (S503). That is, the data generation unit 200 inputs the divided data T i to the inference device F i .
- the divided data T i is input to the inference unit F i such that the input data during learning of the inference unit F i is different from the input data during inference.
- the divided data T i removed by machine learning in the learning unit 202-1 for F i is input to the inference unit F i .
- the learning data storage unit 250 stores the outputs of the reasoners F 1 to F n as the learning data of the reasoner H (S504). That is, the reasoner F i performs inference based on the divided data T i excluded from the machine learning that generates the reasoner F i .
- the learning data storage unit 250 stores the inference result of the inference device F i as learning data of the inference device H. FIG. This completes the generation of learning data.
- the learning unit 122 causes the inference device H to learn using the second learning data (S202).
- the learning unit 122 reads the learning data stored in the learning data storage unit 250 and uses it for machine learning of the reasoner H. FIG. Thereby, the reasoner H is generated.
- the data generator 200 generates the reasoner H.
- FIG. 1
- the inference device F i when the data included in the learning data T is input to the inference device H as input data, the inference device F i generated by machine learning excluding the input data performs inference. Therefore, only the reasoner H can provide sufficient security.
- the reasoner H reduces the classification accuracy for member data to the classification accuracy for non-member data. Therefore, higher safety can be obtained.
- the learning unit 122 performs supervised learning using the inference result obtained by the inference device F i as the correct label. When the member data is input to the reasoner H, the inference result of the reasoner Fi that has learned with the non-member data excluding the member data is output. Therefore, sufficient security can be obtained with the reasoner H alone.
- the data generator 200 generates learning data for the reasoner H based on the outputs of the reasoners F 1 to F n .
- the reasoner H becomes a distilled model generated using the outputs of the reasoners F 1 to F n . That is, the reasoners F 1 to F n extract some information from the learning data T.
- the learning data storage unit 250 causes the reasoner H to learn using the information extracted by the reasoners F 1 to F n as learning data. Therefore, the reasoner H can obtain high accuracy with a simple model.
- the learning unit 122 uses not only the second learning data but also the first learning data. That is, the learning unit 122 uses at least part of the learning data T to perform machine learning. In the learning data T, a true correct label y is associated with the input data x. In a modified example, the learning unit 122 can adjust the ratio of true correct labels y to be mixed with the second learning data.
- the second learning data is the data set shown in Equation (1) above.
- L0 be the loss function when learning using the data set shown in the above equation (1).
- L1 be the loss function when learning is performed using the learning data T, which is the first learning data.
- ⁇ be a parameter for adjusting security and accuracy against MI attacks. ⁇ is a real number of 0 or more and 1 or less.
- the parameter ⁇ indicates the ratio of the first learning data to the second learning data.
- the learning unit 122 generates the reasoner H based on the parameter ⁇ , the loss function L 1 and the loss function L 0 .
- the learning unit 122 performs machine learning based on the loss function L ⁇ .
- FIG. 4 is a block diagram showing a learning device 600 according to another embodiment.
- the learning device 600 includes a data division unit 602 , a reasoner generation unit 603 , a learning data generation unit 604 and a learning unit 605 .
- the data dividing unit 602 generates n sets of divided data by dividing the first learning data into n (n is an integer equal to or greater than 2).
- the reasoner generating unit 603 generates n learning data generation reasoners by machine learning using data obtained by removing one set of divided data from the first learning data.
- the learning data generation unit 604 generates the second learning data by inputting the set of divided data excluded from the machine learning to the n learning data generation reasoners.
- a learning unit 605 generates a reasoner by machine learning using the second learning data.
- each element of the machine learning system can be implemented by a computer program. That is, the reasoner H, the learning unit 122, the data generation unit 200, and the like can each be realized by computer programs. In addition, the reasoner H, the learning unit 122, the data generation unit 200, etc. may not physically be a single device, but may be distributed over a plurality of computers.
- FIG. 5 is a block diagram showing an example of the hardware configuration of the machine learning system 700.
- machine learning system 700 includes, for example, at least one memory 701 , at least one processor 702 , and network interface 703 .
- a network interface 703 is used to communicate with other devices via a wired or wireless network.
- Network interface 703 may include, for example, a network interface card (NIC).
- Machine learning system 700 transmits and receives data via network interface 703 .
- Machine learning system 700 may acquire learning data T via a network interface.
- NIC network interface card
- the memory 701 is configured by a combination of volatile memory and nonvolatile memory.
- Memory 701 may include storage located remotely from processor 702 .
- processor 702 may access memory 701 via an input/output interface (not shown).
- the memory 701 is used to store software (computer program) including one or more instructions to be executed by the processor 702 . Further, when the machine learning system 700 has the learning device 100, the memory 701 may store the reasoner H, the learning units 121 to 123, the data generation unit 200, and the like.
- a program includes a set of instructions (or software code) that, when read into a computer, cause the computer to perform one or more of the functions described in the embodiments.
- the program may be stored in a non-transitory computer-readable medium or tangible storage medium.
- computer readable media or tangible storage media may include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drives (SSD) or other memory technology, CDs - ROM, digital versatile disc (DVD), Blu-ray disc or other optical disc storage, magnetic cassette, magnetic tape, magnetic disc storage or other magnetic storage device.
- the program may be transmitted on a transitory computer-readable medium or communication medium.
- transitory computer readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.
- (Appendix 1) a data dividing unit that generates n sets of divided data by dividing the first learning data by n (where n is an integer equal to or greater than 2); a reasoner generating unit that generates n learning data generating reasoners by machine learning using data obtained by removing one set of divided data from the first learning data; a learning data generation unit configured to generate second learning data by inputting the set of divided data excluded from the machine learning to n learning data generation reasoners, respectively; a learning unit that generates a reasoner by machine learning using the second learning data.
- the learning device wherein the learning unit generates the reasoner based on the parameter ⁇ , the loss function L 1 , and the loss function L 0 .
- the learning device (1 ⁇ ⁇ ) L0 + ⁇ L1 (3) 3.
- the learning device (1 ⁇ ⁇ ) L0 + ⁇ L1 (3) 3.
- the learning device Appendix 3, wherein the reasoner is calculated based on the loss function L ⁇ .
- Appendix 6 A reasoning device generated by the learning device according to any one of appendices 1 to 5.
- a computer-readable medium storing a program for causing a computer to execute a learning method The learning method includes: By dividing the first learning data by n (where n is an integer of 2 or more), n sets of divided data are generated, generating n learning data generating reasoners by machine learning using data obtained by removing one set of divided data from the first learning data; generating second learning data by inputting the one set of divided data excluded from the machine learning to n learning data generation reasoners, respectively; A computer-readable medium for generating a reasoner by machine learning using the second training data. (Appendix 13) In the learning method, 13. The computer-readable medium of Clause 12, wherein the reasoner is generated by machine learning using the first training data.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Bioethics (AREA)
- General Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
本実施の形態にかかる機械学習(訓練)について、図1を用い説明する。図1は、学習装置100を備えた機械学習システムの構成を示すブロック図である。学習装置100は、データ生成部200と、学習部122とを備えている。推論器Hの機械学習には、予め学習データTが用意されている。学習データTをメンバデータとも称し、学習データT以外のデータをノンメンバデータとも称する。
推論器Hは、メンバデータに対する分類精度がノンメンバデータに対する分類精度まで低下する。よって、より高い安全性を得ることができる。また、学習部122は、推論器Fiで得られた推論結果を正解ラベルとして教師有り学習している。推論器Hにメンバデータを入力すると、あたかもメンバデータを除いたノンメンバデータで学習した推論器Fiの推論結果が出力される。よって、推論器H単体で十分な安全性を得ることができる。
変形例では、学習部122が、第2学習データだけでなく、第1学習データを用いている。つまり、学習部122が学習データTの少なくとも一部を用いて機械学習を行っている。学習データTでは、入力データxに対する真の正解ラベルyが対応付けられている。変形例では、学習部122が第2学習データに混ぜ合わせる真の正解ラベルyの割合を調整することができる。
Lα=(1-α)L0+αL1 ・・・(3)
図4はその他の実施形態にかかる学習装置600を示すブロック図である。学習装置600は、データ分割部602と、推論器生成部603と、学習データ生成部604と、学習部605とを備える。
(付記1)
第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成するデータ分割部と、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成する推論器生成部と、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成する学習データ生成部と、
前記第2学習データを用いた機械学習により推論器を生成する学習部と、を備えた学習装置。
(付記2)
前記学習部が前記第1学習データを用いた機械学習により前記推論器を生成する付記1に記載の学習装置。
(付記3)
前記第1学習データでは、入力データと正解ラベルとが対応付けられており、
前記学習部の機械学習において、前記第2学習データに対する前記第1学習データの割合が設定されている付記2に記載の学習装置。
(付記4)
前記第2学習データに対する前記第1学習データの割合を示すパラメータをαとし、
前記第1学習データでの機械学習での損失関数をL1とし、
前記第2学習データでの機械学習での損失関数をL0とした場合、
前記学習部がパラメータα、損失関数L1、及び損失関数L0に基づいて前記推論器を生成している付記3に記載の学習装置。
(付記5)
前記学習部が、
以下の式(3)に基づいて、損失関数Lαを算出し、
Lα=(1-α)L0+αL1 ・・・(3)
前記損失関数Lαに基づいて、前記推論器を算出している付記3に記載の学習装置。
(付記6)
付記1~5のいずれか1項に記載の学習装置で生成された推論装置。
(付記7)
第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成し、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成し、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成し、
前記第2学習データを用いた機械学習により推論器を生成する、学習方法。
(付記8)
前記第1学習データを用いた機械学習により前記推論器を生成する付記7に記載の学習方法。
(付記9)
前記第1学習データでは、入力データと正解ラベルとが対応付けられており、
前記学習部の機械学習において、前記第2学習データに対する前記第1学習データの割合が設定されている付記8に記載の学習方法。
(付記10)
前記第2学習データに対する前記第1学習データの割合を示すパラメータをαとし、
前記第1学習データでの機械学習での損失関数をL1とし、
前記第2学習データでの機械学習での損失関数をL0とした場合、
前記学習部がパラメータα、損失関数L1、及び損失関数L0に基づいて前記推論器を生成している付記9に記載の学習方法。
(付記11)
前記学習部が、
以下の式(3)に基づいて、損失関数Lαを算出し、
Lα=(1-α)L0+αL1 ・・・(3)
前記損失関数Lαに基づいて、前記推論器を算出している付記10に記載の学習方法。
(付記12)
コンピュータに対して学習方法を実行させるためのプログラムが格納されたコンピュータ可読媒体であって、
前記学習方法は、
第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成し、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成し、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成し、
前記第2学習データを用いた機械学習により推論器を生成する、コンピュータ可読媒体。
(付記13)
前記学習方法では、
前記第1学習データを用いた機械学習により前記推論器を生成する付記12に記載のコンピュータ可読媒体。
(付記14)
前記第1学習データでは、入力データと正解ラベルとが対応付けられており、
前記学習部の機械学習において、前記第2学習データに対する前記第1学習データの割合が設定されている付記13に記載のコンピュータ可読媒体。
(付記15)
前記第2学習データに対する前記第1学習データの割合を示すパラメータをαとし、
前記第1学習データでの機械学習での損失関数をL1とし、
前記第2学習データでの機械学習での損失関数をL0とした場合、
前記学習部がパラメータα、損失関数L1、及び損失関数L0に基づいて前記推論器を生成している付記14に記載のコンピュータ可読媒体。
(付記16)
前記学習部が、
以下の式(3)に基づいて、損失関数Lαを算出し、
Lα=(1-α)L0+αL1 ・・・(3)
前記損失関数Lαに基づいて、前記推論器を算出している付記15に記載のコンピュータ可読媒体。
T1~Tn 分割データ
121 学習部
122 学習部
123 学習部
200 データ生成部
220 データ分割部
202-1 F1の学習部
202-n Fnの学習部
250 学習データ記憶部
F1 推論器
Fn 推論器
H 推論器
Claims (10)
- 第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成するデータ分割部と、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成する推論器生成部と、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成する学習データ生成部と、
前記第2学習データを用いた機械学習により推論器を生成する学習部と、を備えた学習装置。 - 前記学習部が前記第1学習データを用いた機械学習により前記推論器を生成する請求項1に記載の学習装置。
- 前記第1学習データでは、入力データと正解ラベルとが対応付けられており、
前記学習部の機械学習において、前記第2学習データに対する前記第1学習データの割合が設定されている請求項2に記載の学習装置。 - 前記第2学習データに対する前記第1学習データの割合を示すパラメータをαとし、
前記第1学習データでの機械学習での損失関数をL1とし、
前記第2学習データでの機械学習での損失関数をL0とした場合、
前記学習部がパラメータα、損失関数L1、及び損失関数L0に基づいて前記推論器を生成している請求項3に記載の学習装置。 - 前記学習部が、
以下の式(3)に基づいて、損失関数Lαを算出し、
Lα=(1-α)L0+αL1 ・・・(3)
前記損失関数Lαに基づいて、前記推論器を算出している請求項3に記載の学習装置。 - 請求項1~5のいずれか1項に記載の学習装置で生成された推論装置。
- 第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成し、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成し、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成し、
前記第2学習データを用いた機械学習により推論器を生成する、学習方法。 - 前記第1学習データを用いた機械学習により前記推論器を生成する請求項7に記載の学習方法。
- コンピュータに対して学習方法を実行させるためのプログラムが格納されたコンピュータ可読媒体であって、
前記学習方法は、
第1学習データをn(nは2以上の整数)分割することで、nセットの分割データを生成し、
前記第1学習データから1セットの分割データを除いたデータを用いた機械学習によりn個の学習データ生成用推論器を生成し、
前記機械学習から除かれた前記1セットの前記分割データをn個の前記学習データ生成用推論器にそれぞれ入力することで、第2学習データを生成し、
前記第2学習データを用いた機械学習により推論器を生成する、コンピュータ可読媒体。 - 前記学習方法では、
前記第1学習データを用いた機械学習により前記推論器を生成する請求項9に記載のコンピュータ可読媒体。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2023520695A JPWO2022239200A5 (ja) | 2021-05-13 | 学習装置、推論装置、学習方法、及びプログラム | |
PCT/JP2021/018265 WO2022239200A1 (ja) | 2021-05-13 | 2021-05-13 | 学習装置、推論装置、学習方法、及びコンピュータ可読媒体 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/018265 WO2022239200A1 (ja) | 2021-05-13 | 2021-05-13 | 学習装置、推論装置、学習方法、及びコンピュータ可読媒体 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022239200A1 true WO2022239200A1 (ja) | 2022-11-17 |
Family
ID=84028064
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/018265 WO2022239200A1 (ja) | 2021-05-13 | 2021-05-13 | 学習装置、推論装置、学習方法、及びコンピュータ可読媒体 |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2022239200A1 (ja) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200311540A1 (en) * | 2019-03-28 | 2020-10-01 | International Business Machines Corporation | Layer-Wise Distillation for Protecting Pre-Trained Neural Network Models |
CN111931216A (zh) * | 2020-09-16 | 2020-11-13 | 支付宝(杭州)信息技术有限公司 | 一种基于隐私保护的方式获取联合训练模型的方法及系统 |
-
2021
- 2021-05-13 WO PCT/JP2021/018265 patent/WO2022239200A1/ja active Application Filing
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200311540A1 (en) * | 2019-03-28 | 2020-10-01 | International Business Machines Corporation | Layer-Wise Distillation for Protecting Pre-Trained Neural Network Models |
CN111931216A (zh) * | 2020-09-16 | 2020-11-13 | 支付宝(杭州)信息技术有限公司 | 一种基于隐私保护的方式获取联合训练模型的方法及系统 |
Non-Patent Citations (2)
Title |
---|
NASR, M. ET AL.: "Machine learning with membership privacy using adversarial regularization", PROCEEDINGS OF THE 2018 ACM SIGSAC CONFERENCE ON COMPUTER AND COMMUNICATIONS SECURITY, October 2018 (2018-10-01), pages 634 - 646, XP058701177, Retrieved from the Internet <URL:https://dl.acm.org/doi/abs/10.1145/3243734.3243855> [retrieved on 20210702], DOI: 10.1145/3243734.3243855 * |
TAKAESU ISAO: "How to infer the training data of a machine learning model", MBSD, 18 June 2020 (2020-06-18), pages 1 - 9, XP055883928, Retrieved from the Internet <URL:https://www.mbsd.jp/blog/20200618.html> [retrieved on 20220126] * |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022239200A1 (ja) | 2022-11-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yan et al. | A method of information protection for collaborative deep learning under GAN model attack | |
Bae et al. | Security and privacy issues in deep learning | |
US11689566B2 (en) | Detecting and mitigating poison attacks using data provenance | |
Mugunthan et al. | Smpai: Secure multi-party computation for federated learning | |
Reith et al. | Efficiently stealing your machine learning models | |
Barreno et al. | Can machine learning be secure? | |
Niu et al. | Toward verifiable and privacy preserving machine learning prediction | |
Hu et al. | Adaptive cyber defense against multi-stage attacks using learning-based POMDP | |
CN115659408B (zh) | 一种电力系统敏感数据共享方法、系统及存储介质 | |
CN115238827B (zh) | 保护隐私的样本检测系统训练方法及装置 | |
KR20220109418A (ko) | 신경 흐름 증명 | |
Zhang et al. | A survey on privacy inference attacks and defenses in cloud-based deep neural network | |
Shafee et al. | Privacy attacks against deep learning models and their countermeasures | |
Kadry et al. | Intrusion detection model using optimized quantum neural network and elliptical curve cryptography for data security | |
Sharifzadeh et al. | Convolutional neural network steganalysis's application to steganography | |
Zhang et al. | Privacy inference attacks and defenses in cloud-based deep neural network: A survey | |
CN113255526A (zh) | 基于动量的对人群计数模型的对抗样本生成方法及系统 | |
Guo et al. | Robust and privacy-preserving collaborative learning: A comprehensive survey | |
Zhang et al. | Visual object detection for privacy-preserving federated learning | |
Yu et al. | Security and Privacy in Federated Learning | |
WO2022239200A1 (ja) | 学習装置、推論装置、学習方法、及びコンピュータ可読媒体 | |
WO2022018867A1 (en) | Inference apparatus, inference method and computer-readable storage medium | |
WO2022239201A1 (ja) | 推論装置、学習装置、機械学習システム、推論方法、学習方法、及びコンピュータ可読媒体 | |
Luqman et al. | Privacy and Security Implications of Cloud-Based AI Services: A Survey | |
Alrawashdeh et al. | Optimizing Deep Learning Based Intrusion Detection Systems Defense Against White-Box and Backdoor Adversarial Attacks Through a Genetic Algorithm |
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: 21941933 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18559389 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023520695 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21941933 Country of ref document: EP Kind code of ref document: A1 |