IEEE Computational Intelligence Magazine - May 2020 - 51

private machine learning. Dwork et al.
[19] briefly summarized a limited num-
ber of the most basic problems on differ-
entially private machine learning, includ-
ing the sample complexity, online
learning and empirical risk minimization.
Ji et al. [20] focused on generalizing dif-
ferential privacy mechanisms to tradi-
tional machine learning models, while
missing many newly developed
approaches that update state-of-the-art
benchmarks. Furthermore, both of them
do not cover the advances of differential
privacy in deep learning. Inspired by this,
the purpose of this survey is to present a
systematic and comprehensive overview
of the extensive researches about differ-
entially private machine learning. More
specifically, this survey investigates on
applications of different differential pri-
vacy mechanisms from the perspectives

of both traditional machine learning and
recently booming deep learning. Fig-
ure 2 roughly provides a summary of dif-
ferentially private machine learning
research. In particular, this survey has
three major contributions as follows.
1) We propose a taxonomy of existing
techniques about differentially pri-
vate machine learning. To the best
of our knowledge, we are the first to
investigate the development of dif-
ferentially private deep learning.
2) We provide a detailed and thorough
study of the state-of-the-art methods.
As a result, this survey brings new
perspectives to better understand the
existing works and improve the priva-
cy level of machine learning models.
3) To facilitate the timely and potential
research of this area for researchers
from privacy preserving, especially

Laplace/Gaussian/
Exponential Mechanism

Output/Objective
Perturbation Mechanism

Traditional Machine
Learning

Traditional Machine
Learning

Naive Bayes
Model
Decision Tree
Learning
Boosting
Clustering
Dimensionality
Reduction
Online Learning
Deep Learning
Centralized
Training
Algorithms

machine learning, we summarize the
limitations and challenges of current
research works, and suggest several
promising future research directions.
The rest of this paper is organized as
follows. Section II introduces the related
backgrounds. In Section III, we give a
theoretical description of the application
of differential privacy in traditional
machine learning in detail. Section IV
carefully covers recent models of differen-
tially private deep learning. In Section V,
we discuss the existing challenges and
point out promising future directions. Sec-
tion VI summarizes this review. Due to
the limitation of space, five summary tables
are shown in the supplementary material1.
II. Backgrounds

In this section, we will introduce the def-
inition of differential privacy, the princi-
ple differential privacy mechanisms and
then present a brief overview of machine
learning. The notations used in this sur-
vey are summarized in Table I.
A. Differential Privacy

Differential privacy [14] is a solid priva-
cy-preserving model presented by
Dwork et al. in 2006. It aims to preserve

Linear Regression
Logistic
Regression

1
The supplementary material that includes five summa-
ry tables is available at http://see.xidian.edu.cn/faculty/
mggong/publication.htm

Linear SVM
Kernel SVM

TABLE 1 The common notations
used in this survey.

Deep Learning
Deep
Auto-Encoders
Convolutional
Deep Belief
Networks
DeepWalk

Distributed
Training
Algorithms
Long Short-Term
Memory Networks

NOTATIONS

EXPLANATION

D

DATASET

X

SAMPLE SPACE

Y

OUTPUT SPACE

M

ALGORITHM

n

SIZE OF THE DATASET

d

DIMENSION OF THE SAMPLE SPACE

L

LOSS FUNCTION

e

PRIVACY BUDGET

d

POSSIBILITY OF VIOLATING
e -DIFFERENTIAL PRIVACY

S^ $ h

Generative
Neural Networks

SENSITIVITY

Lap ^ $ h

LAPLACE DISTRIBUTION
2

N (0 , v )

FIGURE 2 The proposed taxonomy to summarize the methods about differentially private
machine learning.

GAUSSIAN DISTRIBUTION
WITH MEAN 0 AND VARIANCE v 2

MAY 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

51
http://see.xidian.edu.cn/faculty/mggong/publication.htm http://see.xidian.edu.cn/faculty/mggong/publication.htm
IEEE Computational Intelligence Magazine - May 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2020
