WO2024116402A1 - 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム - Google Patents

連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム Download PDF

Info

Publication number
WO2024116402A1
WO2024116402A1 PCT/JP2022/044548 JP2022044548W WO2024116402A1 WO 2024116402 A1 WO2024116402 A1 WO 2024116402A1 JP 2022044548 W JP2022044548 W JP 2022044548W WO 2024116402 A1 WO2024116402 A1 WO 2024116402A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
learning
dual
variables
model learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/JP2022/044548
Other languages
English (en)
French (fr)
Japanese (ja)
Inventor
匠 深見
一凡 張
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nippon Telegraph and Telephone Corp
Original Assignee
Nippon Telegraph and Telephone Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nippon Telegraph and Telephone Corp filed Critical Nippon Telegraph and Telephone Corp
Priority to PCT/JP2022/044548 priority Critical patent/WO2024116402A1/ja
Priority to JP2024561121A priority patent/JPWO2024116402A1/ja
Publication of WO2024116402A1 publication Critical patent/WO2024116402A1/ja
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/098—Distributed learning, e.g. federated learning

Definitions

  • Non-Patent Document 2 only supports a basic network configuration consisting of a server and a client, and has the problem that it does not support complex network configurations, such as when the number of other terminals connected varies from terminal to terminal.
  • the present disclosure aims to provide a federated learning technology that can prevent degradation of learning accuracy not only in basic network configurations, but also in other complex network configurations, and even when attacked by an attacker.
  • the federated learning system of the present disclosure has a plurality of model learning devices, each of which is connected to one of the other model learning devices via a network.
  • the model parameters, dual variables, learning rate, learning rate of the dual variables, and contribution are set to predetermined initial values.
  • Each model learning device has a mini-batch extraction unit, a gradient calculation unit, a dual variable restriction unit, a model parameter update unit, a dual variable calculation/transmission unit, a dual variable reception unit, and a dual variable setting unit.
  • the mini-batch extraction unit extracts a predetermined amount of data as a mini-batch from the model learning data.
  • the gradient calculation unit calculates the gradient of the cost function from the model parameters and the mini-batch of the model learning data.
  • the dual variable restriction unit restricts the magnitude of the dual variables based on a predetermined condition.
  • the model parameter update unit performs learning using the model parameters, the restricted dual variables, the learning rate, the learning rate of the dual variables, the contribution, the gradient of the cost function, the constraint parameters, and the number of other model learning devices connected to the model learning device, and updates the model parameters.
  • the dual variable calculation and transmission unit calculates and transmits dual variables for each of the other model learning devices connected to the model learning device using the updated model parameters, the learning rate of the dual variables, the restricted dual variables, the contributions, and the constraint parameters.
  • the dual variable reception unit receives dual variables from the other model learning devices connected to the model learning device.
  • the dual variable setting unit sets the received dual variables as dual variables to be used in the next learning.
  • the values of the dual variables used are restricted by the dual variable restriction unit 140. This makes it possible to reduce the impact on model learning even if an attacker transmits dual variables with large values. Therefore, it is possible to prevent degradation of learning accuracy due to attacks not only in basic network configurations but also in other complex network configurations.
  • FIG. 1 is a diagram showing algorithm 2 and equations 7, 8, 10, and 11 shown in Non-Patent Document 2.
  • FIG. 2 is a diagram showing an example of the functional configuration of the associative learning system according to this embodiment.
  • FIG. 3 is a diagram showing an example in which model learning devices according to this embodiment are connected in a ring shape.
  • FIG. 4 is a diagram showing an example in which model learning devices according to this embodiment are randomly connected.
  • FIG. 5 is a diagram showing an example of a processing flow of the federated learning system according to this embodiment.
  • FIG. 6 is a diagram showing the algorithm of the federated learning system according to this embodiment in the same format as that of Non-Patent Document 2.
  • FIG. 7 is a diagram illustrating an example of the functional configuration of a computer.
  • FIG. 2 shows an example of the functional configuration of the federated learning system according to this embodiment.
  • FIG. 3 shows an example of the model learning devices according to this embodiment connected in a ring shape, and
  • FIG. 4 shows an example of the model learning devices according to this embodiment connected randomly.
  • FIG. 5 shows an example of the processing flow of the federated learning system according to this embodiment.
  • FIG. 6 shows an algorithm of the federated learning system according to this embodiment in the same description format as Non-Patent Document 2.
  • the federated learning system 10 of this embodiment has N model learning devices 100 1 , ..., 100 N , and each model learning device 100 i is connected to one of the other model learning devices 100 j via a network 900.
  • Each model learning device 100 i includes an initial setting unit 110, a mini-batch extraction unit 120, a gradient calculation unit 130, a dual variable restriction unit 140, a model parameter update unit 150, a dual variable calculation and transmission unit 160, a dual variable reception unit 170, and a dual variable setting unit 180.
  • 3 and 4 are merely examples, and the network configuration, the number N of model learning devices, and the number E i of other model learning devices connected to the model learning device 100 i can be determined arbitrarily.
  • j , the learning rate ⁇ of the learning model, the learning rate ⁇ i of the dual variables, and the contribution ⁇ i are set to predetermined initial values (S110).
  • w 1 , ..., w N are set to the same value
  • j z i
  • j 0.
  • ⁇ and ⁇ i are set to arbitrary values.
  • ⁇ i is determined by 1/( ⁇ E i K)
  • E i is the number of other model learning devices (100 j ) connected to the model learning device (100 i )
  • K is the number of iterations in the inner loop process.
  • the mini-batch extraction unit 120 of each model learning device 100 1 , ..., 100 N extracts a predetermined amount of data as a mini-batch ⁇ i r,k from the model learning data x i (S120). For example, when there are 10,000 pieces of learning data, 500 or 1,000 mini-patches may be arbitrarily extracted.
  • the i at the bottom right of ⁇ indicates the number of the model learning device.
  • the k at the top right of ⁇ is an integer indicating the number of repetitions of the repetitive process (inner loop process) performed by each model learning device 100 1 , ..., 100 N. In the inner loop process, K repetitive processes are performed.
  • K is an integer of 2 or more
  • k is an integer of 1 to K.
  • the r at the top right of ⁇ is an integer indicating the number of repetitions of the repetitive process of the inner loop process (outer loop process) in the entire associative learning system.
  • R repetitive processes are performed.
  • R is an integer of 2 or more
  • r is an integer of 1 to R.
  • the i and j in the lower right of the symbols indicate the model learning device number
  • the r in the upper right of the symbols indicates the number of iterations in the outer loop process
  • the k in the upper right of the symbols indicates the number of iterations in the inner loop process. Note that in the text of the specification, the symbols in the lower right and the upper right cannot be written in the same position horizontally, so they are written offset from each other. On the other hand, in the formulas and figures in the specification, they can be written, so they are written in the same position horizontally.
  • the gradient calculation unit 130 of each model learning device 100 1 , ..., 100 N calculates the gradient g i (w i r,k ) of the cost function from the model parameters w i r , k and the mini-batch ⁇ i r ,k of the model learning data (step S130). Specifically, the gradient calculation unit 130 calculates the gradient g i (w i r,k ) of the cost function of the learning model as shown in the following formula.
  • f i is the cost function
  • ⁇ indicates the gradient.
  • the calculated gradient g i (w i r,k ) of the cost function of the learning model is used to calculate the update of the model parameters w i described later.
  • the arrows in the formula indicate substitution.
  • the dual variable restriction unit 140 of each model learning device 100 1 , ..., 100 N restricts the magnitude of the dual variables based on a predetermined condition (step S140). Specifically, the dual variable restriction unit 140 calculates the dual variable z i
  • the symbol “ ⁇ " (superscript hat) on the left side of z indicates that it is before restriction. Note that in the text of the specification, it cannot be written above z, so it is written shifted. On the other hand, it can be written in the formulas and figures of the specification, so it is written above z.
  • j r is calculated from the following equation.
  • C is an arbitrary value that is set in advance, and is referred to as a clipping variable here.
  • j r is calculated from the following equation.
  • j r ⁇ 1,k is a dual variable calculated by the model learning device 100 i by the dual variable calculation and transmission unit 160 described later, and is a dual variable calculated in the r ⁇ 1th outer loop process and the kth inner loop process.
  • Clipping is a method for restricting the received pre-restriction dual variable ⁇ z i
  • j r-1,k are compared for each element, and if an element of ⁇ z i
  • Reducing is a method for reducing the received pre-restriction dual variable ⁇ z i
  • j r is restricted by substituting a value obtained by multiplying ⁇ z i
  • the dual variable calculation/transmission unit 160 restricts the magnitude of the value of the dual variable z i
  • j r is clipped or reduced using y i
  • j r may be clipped or reduced using the value (y i
  • j r may be clipped or reduced using the value (y i
  • clipping or shrinking may be performed using the average or median of the dual variables (y j
  • the model parameter update unit 150 of each of the model learning devices 100 1 , ..., 100 N performs learning using the model parameters w i r,k , the restricted dual variables z i
  • the model parameter update unit 150 updates the model parameters as shown in the following formula, and obtains the model parameters w i r,k+1 to be used in the next process (the k+1th process of the inner loop process).
  • the learning rate ⁇ i of the dual variable may be 1/( ⁇ E i K).
  • the contribution ⁇ i is calculated from 1+ ⁇ i .
  • the introduction of the hyperparameter ⁇ is optional, by introducing ⁇ (i.e., by calculating the contribution ⁇ i from 1+ ⁇ i ), it is possible to control the influence of the received dual variable ⁇ z i
  • ⁇ is increased to increase the value of the contribution ⁇ i
  • the contribution of the information of the other model learning device 100 j which is the attacker, to the learning can be reduced as much as possible.
  • model parameters w i r,k For details of the model parameters w i r,k , the dual variables z i
  • the dual variable calculation and transmission unit 160 of each model learning device 100 1 , ..., 100 N calculates a dual variable y i
  • j is an intermediate value for calculating the dual variable y i
  • the dual variable receiving unit 170 of each of the model learning devices 100 1 , ..., 100 N receives the dual variable y j
  • i r, k+1 received by the dual variable receiving unit 170 is calculated and transmitted by the dual variable calculation and transmission unit 160 of the model learning device 100 j , and therefore the positions of "j" and "i" are reversed from those of the dual variable y i
  • the dual variable setting unit 180 of each model learning device 100 1 , ..., 100 N sets the received dual variable y j
  • the dual variable setting unit 180 may set the received dual variable y j
  • Each of the model learning devices 100 1 , ..., 100 N checks whether the processing of the inner loop process has been completed, and if not (No), continues the repeated processing, and if completed (Yes), proceeds to checking the outer loop process (S190). If step S190 is Yes, each of the model learning devices 100 1 , ..., 100 N checks whether the processing of the outer loop process has been completed, and if not (No), continues the repeated processing, and if completed (Yes), ends the processing (S195).
  • j used are restricted in advance by the dual variable restriction unit 140 based on a predetermined condition. Therefore, even if an attacker transmits a dual variable with a large value, the influence on the model learning can be reduced. Therefore, even if an attacker attacks a network with a basic network configuration consisting of a server and a client, for example, other complex network configurations such as those shown in FIG. 3 or FIG. 4, the degradation of learning accuracy can be prevented.
  • the program describing this processing can be recorded on a computer-readable recording medium.
  • Examples of computer-readable recording media include magnetic recording devices, optical disks, magneto-optical recording media, and semiconductor memories.
  • the program may be distributed, for example, by selling, transferring, or lending portable recording media such as DVDs or CD-ROMs on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network.
  • a computer that executes such a program for example, first stores in its own storage device the program recorded on a portable recording medium or the program transferred from a server computer. Then, when executing a process, the computer reads the program stored on its own recording medium and executes the process according to the read program. As another execution form of the program, the computer may read the program directly from the portable recording medium and execute the process according to the program, or may execute the process according to the received program each time a program is transferred from the server computer to the computer.
  • the above-mentioned process may also be executed by a so-called ASP (Application Service Provider) type service that does not transfer the program from the server computer to the computer, but realizes the processing function only by issuing an execution instruction and obtaining the results.
  • ASP Application Service Provider
  • the program in this form includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct command to the computer but has properties that specify the processing of the computer).
  • the device is configured by executing a specific program on a computer, but at least a portion of the processing may be realized by hardware.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
PCT/JP2022/044548 2022-12-02 2022-12-02 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム Ceased WO2024116402A1 (ja)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/JP2022/044548 WO2024116402A1 (ja) 2022-12-02 2022-12-02 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム
JP2024561121A JPWO2024116402A1 (enrdf_load_stackoverflow) 2022-12-02 2022-12-02

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2022/044548 WO2024116402A1 (ja) 2022-12-02 2022-12-02 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム

Publications (1)

Publication Number Publication Date
WO2024116402A1 true WO2024116402A1 (ja) 2024-06-06

Family

ID=91323166

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2022/044548 Ceased WO2024116402A1 (ja) 2022-12-02 2022-12-02 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム

Country Status (2)

Country Link
JP (1) JPWO2024116402A1 (enrdf_load_stackoverflow)
WO (1) WO2024116402A1 (enrdf_load_stackoverflow)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017520824A (ja) * 2014-05-12 2017-07-27 クゥアルコム・インコーポレイテッドＱｕａｌｃｏｍｍ Ｉｎｃｏｒｐｏｒａｔｅｄ 共通特徴にわたる分類子の更新
WO2022249436A1 (ja) * 2021-05-28 2022-12-01 日本電信電話株式会社 変数最適化システム

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017520824A (ja) * 2014-05-12 2017-07-27 クゥアルコム・インコーポレイテッドＱｕａｌｃｏｍｍ Ｉｎｃｏｒｐｏｒａｔｅｄ 共通特徴にわたる分類子の更新
WO2022249436A1 (ja) * 2021-05-28 2022-12-01 日本電信電話株式会社 変数最適化システム

Also Published As

Publication number Publication date
JPWO2024116402A1 (enrdf_load_stackoverflow) 2024-06-06

Similar Documents

Publication Publication Date Title
Jere et al. A taxonomy of attacks on federated learning
Bonte et al. Privacy-preserving logistic regression training
Chou et al. Villandiffusion: A unified backdoor attack framework for diffusion models
Niu et al. Toward verifiable and privacy preserving machine learning prediction
CN110537191A (zh) 隐私保护机器学习
Jiang et al. Towards a more reliable privacy-preserving recommender system
US20240259181A1 (en) Computational network conversion for fully homomorphic evaluation
WO2021252294A1 (en) Protection of transformations by intermediate randomization in cryptographic operations
Vaishnavi et al. Transferring adversarial robustness through robust representation matching
CN118350452A (zh) 基于注意力机制的联邦学习方法、客户端、服务器和系统
You et al. Single-trace fragment template attack on a 32-bit implementation of keccak
Gohr et al. Subsampling and knowledge distillation on adversarial examples: New techniques for deep learning based side channel evaluations
Krček et al. Autoencoder-enabled model portability for reducing hyperparameter tuning efforts in side-channel analysis
Zhou et al. Breaking symmetric cryptosystems using the offline distributed Grover-meets-Simon algorithm
Li et al. A systematic study of data augmentation for protected AES implementations
WO2024116402A1 (ja) 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム
Luo et al. Distributed Differentially Private Matrix Factorization for Implicit Data via Secure Aggregation
JP7732591B2 (ja) 連合学習システム、モデル学習装置、連合学習方法、モデル学習プログラム
Wagh New directions in efficient privacy-preserving machine learning
Poudel et al. Machine Learning-Based AES Key Recovery via Side-Channel Analysis on the ASCAD Dataset
Li et al. On approximating optimal weighted composite likelihood method for spatial models
Fan et al. Robust enhanced collaborative filtering without explicit noise filtering
Steward et al. Risk‐adjusted monitoring of healthcare quality: model selection and change‐point estimation
Deng et al. Privacy Preservation in User Behavior Analysis for Mobile Edge Computing
Bergmann Cryptanalysis using nature-inspired optimization algorithms

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: 22967232

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2024561121

Country of ref document: JP

NENP Non-entry into the national phase

Ref country code: DE