Privacy-Preserving Classification Scheme Based ON SVM

IEEE SYSYEMS JOURNAL

2022/4/26 READING NOTES

1.Abstract

1.1 SVM is play a crucial part in ML

• data mining
• pattern recognition

1.2 privacy protection of senstive data in SVMS is become more and more important

• face recognition
• biometric information

1.3 problems in current research

• HE &SMPC
• computational effciency is low
• the scakability of the schemes is poor
• the user must stay online in some solutions

1.4 new methods

• this article designs a secure and efficient classification scheme based on SVM to protect the privacy of private data and support vectors in the calculation and transmission process.

• the distributed two trapdoors public-key cryptosystem proposed by Liu is used to realize the distributed double-key decryption function

  • weaken the decryption capability of a cloud server with the master key, prevent the server from launching active attacks

• design a universal secure computing protocol for non linear SVMS based on the Gaussian kernel function

  • also can be extend to polynomial kernel functions

1.5 new methods advantages

• reduces the amount of encrypted data
• simplifies the calculation process
• improves calculation efficiency
• an introduced cloud server realizes user offline function.
• verify its efficiency through experiments
• show that the scheme has the advantages of high efficiency
• good scalability
• user offline function.

2.Introduction

2.1 big data and machine learning is becoming more and more popular

• It has significant applications in
  • electronic commerce
  • financial services
  • transportation？
  • medical and health services

2.2 massive data will inevitably cause privacy-preserving problems

• in process of
  • storage
  • interaction
  • application

  machine learning service providers have access tothe users’ information in the training and prediction phase and can easily obtain private data, resulting in privacy leakage.

2.3 SVM is play a crucial part in ML

• where SVM from?

  • first used in recognition of handwritten digital
    library by Bell Laboratories [1]

• many applications

  • computer vision [2]
  • medical diagnosis [3]
  • information filtering [4]

• SVM plays an important role by

  • virtue of its ability to solve high-dimensional data
  • nonlinear feature problems
  • combine classification interval maximization with kernel method based on statistical learning theory
  • SVM solves the “overlearning” problem in a small sample space
  • remedies the defect of the local extremum.

2.4 problems in current research

• MPC

  • MPC has a large amount of data interaction
    • which cannot meet the practical requirements in terms of efficiency in SVM

• fully HE

  • the schemes based on fully HE have low efficiency
  • remain in the stage of theoretical experiments in SVM

• partial HE

  • partial HE are the mainstreams to satisfy the practical requirements.
    • the low effciency of ciphertext calculation
      • In order to realize complex ciphertext calculation
        • transform the relevant formulas of SVM
          • not only increases the amount of ciphertext calculation?
          • but also increases the amount of data interaction between users and servers?
    • poor scalability
      • SVM can be divided into linear SVM and nonlinear SVM
        - there are many kernel functions available for nonlinear SVM
        - Most of the existing partial HE schemes are designed for a certain type of SVM.
      • When the type of SVM changes, the schemes need to be redesigned.
    • long user online time
    • Since some schemes require the
      users and the servers to carry out cooperative computing4
    • so that the users must stay online.