ArticlePDF Available

# Deep Learning for Edge Computing Applications: A State-of-the-Art Survey

Authors:

## Abstract and Figures

With the booming development of Internet-of-Things (IoT) and communication technologies such as 5G, our future world is envisioned as an interconnected entity where billions of devices will provide uninterrupted service to our daily lives and the industry. Meanwhile, these devices will generate massive amounts of valuable data at the network edge, calling for not only instant data processing but also intelligent data analysis in order to fully unleash the potential of the edge big data. Both the traditional cloud computing and on-device computing cannot sufficiently address this problem due to the high latency and the limited computation capacity, respectively. Fortunately, the emerging edge computing sheds a light on the issue by pushing the data processing from the remote network core to the local network edge, remarkably reducing the latency and improving the efficiency. Besides, the recent breakthroughs in deep learning have greatly facilitated the data processing capacity, enabling a thrilling development of novel applications, such as video surveillance and autonomous driving. The convergence of edge computing and deep learning is believed to bring new possibilities to both interdisciplinary researches and industrial applications. In this article, we provide a comprehensive survey of the latest efforts on the deep-learning-enabled edge computing applications and particularly offer insights on how to leverage the deep learning advances to facilitate edge applications from four domains, i.e., smart multimedia, smart transportation, smart city, and smart industry. We also highlight the key research challenges and promising research directions therein. We believe this survey will inspire more researches and contributions in this promising field.
Content may be subject to copyright.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identiﬁer xxx/ACCESS.2017.DOI
Deep Learning for Edge Computing
Applications: A State-of-the-art Survey
FANGXIN WANG1(Student Member, IEEE), MIAO ZHANG1, XIANGXIANG WANG1,
XIAOQIANG MA2, AND JIANGCHUAN LIU1(Fellow, IEEE)
1School of Computing Science, Simon Fraser University, Burnaby, BC V5A 1S6, Canada
2School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China
Corresponding Author: Xiaoqiang Ma (mxqhust@gmail.com).
This work was supported in part by a Canada NSERC Discovery Grant and the National Natural Science Foundation of China under Grant
61702204.
ABSTRACT
With the booming development of Internet-of-Things (IoT) and communication technologies such as
5G, our future world is envisioned as an interconnected entity where billions of devices will provide
uninterrupted service to our daily lives and the industry. Meanwhile, these devices will generate massive
amounts of valuable data at the network edge, calling for not only instant data processing but also intelligent
data analysis in order to fully unleash the potential of the edge big data. Both the traditional cloud computing
and on-device computing cannot sufﬁciently address this problem due to the high latency and the limited
computation capacity, respectively. Fortunately, the emerging edge computing sheds a light on the issue by
pushing the data processing from the remote network core to the local network edge, remarkably reducing
the latency and improving the efﬁciency. Besides, the recent breakthroughs in deep learning have greatly
facilitated the data processing capacity, enabling a thrilling development of novel applications, such as video
surveillance and autonomous driving. The convergence of edge computing and deep learning is believed
to bring new possibilities to both interdisciplinary researches and industrial applications. In this article,
we provide a comprehensive survey of the latest efforts on the deep-learning-enabled edge computing
applications and particularly offer insights on how to leverage the deep learning advances to facilitate edge
applications from four domains, i.e., smart multimedia, smart transportation, smart city, and smart industry.
We also highlight the key research challenges and promising research directions therein. We believe this
survey will inspire more researches and contributions in this promising ﬁeld.
INDEX TERMS Internet of Things, Edge Computing, Deep Learning, Intelligent Edge Applications.
I. INTRODUCTION
With the explosive development of the Internet-of-Things
(IoT) as well as the communication technologies such as
WiFi and 5G, our future world is envisioned as an inter-
connected entity where billions of digital devices would
provide uninterrupted services to both our daily lives and the
industry. As reported by Cisco [1], there will be more than
50 billion IoT devices connected by the Internet by 2020.
Such numerous IoT devices will generate a myriad of valu-
able data which, once being well processed effectively and
efﬁciently, can empower many groundbreaking applications.
Traditional computing architecture relies on cloud computing
to provide sufﬁcient computation capacity and sustainable
energy. In this system, IoT devices are responsible to collect
the data and deliver it to the remote powerful cloud, and
the cloud servers will carry out the computation-intensive
tasks and distributed the result back. However, the large
latency caused by the long physical distance can sometimes
become unacceptable, especially for those latency-sensitive
applications like autonomous driving and highly interactive
applications such as VR gaming. In addition, the huge data
communication also greatly increases the pressure of the
backbone network, bringing large overhead and cost to ser-
vice providers.
The emerging edge computing [2] provides a promising
solution for this problem. Though with many representa-
tion forms, such as fog computing [3] and cloudlet [4],
the basic idea of edge computing is that the computation
capacity should be deployed close to the data source for
data processing, rather than transmitting the data to places
VOLUME 4, 2016 1
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Edge
Nodes
Edge Computing Deep Learning
Applications
FIGURE 1: The illustration of deep learning enabled edge computing applications.
with computation power. In this way, massive numbers of
servers are deployed at the edge of the network and the tasks
at IoT end devices can be ofﬂoaded to the edge servers for
instant processing. The paradigm of edge computing brings
many beneﬁts compared to cloud computing. First, since
data computing happens closer to the data source, the com-
munication latency can be largely reduced, facilitating the
development of latency-sensitive applications. Besides, local
computation can better protect data privacy and application
security. Last but not least, data processing at the network
edge can effectively reduce trafﬁc at the backbone network
so as to alleviate the network pressure.
Deep learning has made remarkable breakthroughs in re-
cent years due to the powerful perception ability. It has been
widely used in various ﬁelds such as computer vision [5] and
natural language processing [6]. Besides, its performance in
computer and chess games, e.g., Atari Games [7] and the
game of Go [8], even exceeds the best level of human players.
The conﬂuence of edge computing and deep learning will
undoubtedly sheds a light on address the current challenges,
enabling more desirable applications. On one hand, the ap-
plications of edge computing urgently need the powerful pro-
cessing capabilities of deep learning to handle various com-
plicated scenarios, such as video analytics [9], transporta-
tion control [10], etc. On the other hand, edge computing
has provided speciﬁcally designed hardware foundations and
platforms to better support deep learning running at the edge,
e.g., the light-weighted Nvidia Jetson TX2 developing kit1.
Though lots of pioneer efforts have been made towards deep-
learning-enabled edge computing applications, this ﬁeld is
still in the infant stage.
Several existing surveys have investigated the convergence
of deep learning and edge computing in the literature. Han et
1https://developer.nvidia.com/embedded/jetson-tx2
al. [11] presented their understanding on edge computing and
deep learning from ﬁve aspects, while they did not make a
comprehensive and in-depth overview from the perspective
of applications. Similarly, Chen et al. [12] focused on mul-
tiple aspects in deep learning and edge computing, but only
mentioned a general abstraction for those emerging applica-
tions. Zhou et al. [13] mainly focused on the deep learning
model training and inference with edge computing. There are
also a series of surveys for mobile edge computing [2], [14]–
[16] and deep learning [17], [18], respectively, while they
focused on either of them without a comprehensive review
on the combination. Therefore, a complete survey on the
current cutting-edge researches is required at this time to
provide a comprehensive review on deep-learning-enabled
edge computing applications and illuminate the potential
future directions.
To fulﬁll this gap, in this article, we focus on the con-
ﬂuence of edge computing and deep learning, and conduct
an up-to-date literature review on the latest advances of
leveraging deep learning to empower the edge computing
applications, as illustrated in Fig. 1. We ﬁrst provide a brief
overview of edge computing and deep learning on concepts,
advantages as well as representative technologies. We then
summarize the deep-learning-enabled edge computing appli-
cations into four representative domains, i.e., smart multi-
media, smart transportation, smart city and smart industry,
which cover a series of crucial applications like video analyt-
ics, autonomous driving, intelligent trafﬁc control, industrial
manufacturing, etc. At last, we discuss some key research
challenges and promising research directions to achieve sta-
ble, robust, and practical edge learning applications. Differ-
learning enabled edge computing applications, presenting a
comprehensive review and highlighting the challenges and
opportunities.
2VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
the basic paradigm understanding of edge computing and
its advantages in section II. We introduce some deep learn-
ing techniques in section III. The review of deep-learning-
enabled edge applications is summarized in section IV. We
highlight the challenges and research directions in section V.
II. EDGE COMPUTING OVERVIEW
The emerging edge computing in recent years has seen suc-
cessful development in various ﬁelds given its great poten-
tial in reducing latency and saving cost. Different from the
cloud computing architecture, edge computing enables data
processing at the edge of the network. On one hand, data
computing is put closer to the data source, which greatly
facilitates the development of delay-sensitive applications.
On the other hand, the network trafﬁc is largely reduced
since the local processing avoids much data transmission,
which remarkably saves the cost. In this section, we brieﬂy
introduce some edge computing paradigms and highlight the
The key component in edge computing is the edge devices,
which are usually edge servers located closer at the network
end for data processing, communication, caching, etc. There
are also some other paradigms or technologies that share
similar concepts with edge computing. We next discuss and
differentiate some typical paradigms that are related to edge
computing.
1) Cloudlet
Cloudlet, initiated by Carnegie Mellon University, is en-
visioned as small clusters with certain computation and
storage capabilities deployed near the mobile devices such
as buildings and shopping centers for assisted processing,
ofﬂoading, caching, etc. Cloudlet usually utilizes virtualiza-
tion management technologies [4] to better support mobile
applications. And an important target of cloudlet is to bring
the cloud advances to mobile users [19], achieving more
low-latency and resourceful processing. Micro data centers
(MDCs) [20], initiated by Microsoft that are similar to the
concept of cloudlet, are a small-scaled version of data centers
to extend the hyperspace cloud data centers. Different MDCs
are connected by the backbone network to achieve more efﬁ-
cient and intelligent computation, caching, and management.
MDC also serves as an important role in managing numerous
Internet of Things (IoT) devices [21].
2) Fog Computing
Fog computing [3], ﬁrst proposed by Cisco, is a computing
paradigm that aims to bring cloud computing services to the
end of the enterprise network. In fog computing, the data
processing is carried out at fog nodes, which are usually de-
ployed at the network gateway. The fog computing presents
a high-level platform that the numerous IoT devices can be
interconnected through the distributed fog nodes to provide
collaborative services [22]. The fog nodes are also mainly
designed to provide better support for the IoT devices. From
this perspective, compared to other similar edge computing
paradigms, fog computing often stands in alignment with IoT
and emphasizes more on the end side.
3) Mobile Edge Computing
The paradigm of mobile edge computing was ﬁrst standard-
ized by European Telecommunications Standards Institute
(ETSI) [23], which aims to provide sufﬁcient computing
capacities within the radio access network (RAN). It envi-
sions that the computing capacities are placed at the end of
the cellular network, e.g., the wireless base stations. Since
base stations are the important access gate for numerous
IoT devices, mobile edge computing could provide direct
service to the end devices through only one hop, bringing
great convenience for IoT data processing [16].
Compared to traditional cloud computing, edge computing
has many unique advantages, including low latency, energy
saving, context-aware service, and privacy as well as security.
We next summarize them as follows.
1) Low Latency
Since edge devices are placed closer to end devices, which
are usually both the data source and the transmission tar-
get of processing results, the transmission latency can be
largely reduced compared to the cloud computing scenario.
For example, the transmission latency is usually tens (or
hundreds) of milliseconds between an end user and a cloud
server, while this number is usually several milliseconds or
even at microsecond level. The emerging 5G technology
further enhances the advances of edge computing from the
perspective of low latency transmission, which empowers a
series of emerging applications, such as autonomous driv-
ing [24], virtual reality/augmented reality and healthcare-
related applications.
2) Energy Saving
Restricted by the size and usage scenarios, IoT devices
usually have quite limited energy supply, but they are also
expected to perform very complex tasks that are usually
power consuming. It is challenging to design a cost-efﬁcient
solution to well power the numerous distributed IoT devices
given that frequent battery charging/discharging is impracti-
cal in not possible [16]. Edge computing enables the billions
of IoT devices to ofﬂoad the most power-consuming compu-
tation tasks to the edge servers, which not only greatly reduce
the power consumption but also improves the processing
efﬁciency.
3) Context-Aware Service
Context-aware computing [25] is playing an important role in
IoT and edge computing applications, since good modeling
VOLUME 4, 2016 3
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Network
State
Action
observe state
Policy
Environment
(a) Restricted Boltzmann Machine
Network
State
Action
observe state
Policy
Environment
(b) Autoencoder
Network
State
Action
observe state
Policy
Environment
(c) Convolutional Neural Network
Network
State
Action
observe state
Policy
Environment
(d) Deep Neural Network
Network
State
Action
observe state
Policy
Environment
(e) Recurrent Neural Network
Network
State
Action
observe state
Policy
Environment
(f) Deep Reinforcement Learning
FIGURE 2: The structures of different deep learning models.
and reasoning of collected data can highly rely on the context
of the data. With the advantage of the proximity nature,
edge servers can collect more context information to support
the data processing. For example, in the Amazon Go super-
market, video cameras can not only record the goods that
customers select but also predict customers’ interest based
on their staying location, duration and behaviors.
4) Privacy and Security
Compared to cloud computing, edge computing is more
efﬁcient and effective in protecting the data privacy and
application security of users. On one hand, edge servers
are usually geographically distributed clusters that could
be managed and maintained by users themselves. Sensitive
information can be monitored and protected more strictly.
On the other hand, the small-scale nature makes it more
concealed than large-scale data centers, further making it less
likely to become a target of attacks [26].
III. DEEP LEARNING METHODS
Deep learning has been widely applied in many ﬁelds with
great success [27], such as computer vision (CV), natural
language processing (NLP), and artiﬁcial intelligence (AI).
Compared to traditional machine learning methods, deep
learning has demonstrated powerful information extraction
and processing capabilities, but also requires massive com-
putation resources. The breakthroughs of deep learning have
greatly expanded the edge computing applications in various
scenarios, improving performance, efﬁciency, and manage-
ment. In this section, we introduce some typical deep learning
models that are widely used for edge computing applications,
including restricted Boltzmann machine (RBM), autoencoder
(AE), deep neural network (DNN), convolutional neural net-
work (CNN), recurrent neural network (RNN), and deep
reinforcement learning (DRL). The basic architectures of
these learning models are illustrated in Fig. 2.
A. RESTRICTED BOLTZMANN MACHINE
Restricted Boltzmann machine (RBM) is a kind of proba-
bilistic graphical models that can be interpreted as stochastic
neural networks [28]. A typical two-layer RBM includes a
visible layer that contains the input we know and a hid-
den layer that contains the latent variables, as described in
Fig. 2(a). RBMs are organized as a bipartite graph, where
each visible neuron is connected to all hidden neurons
and vice versa, but any two units are not connected in
the same layer. RBMs have seen successful applications in
many ﬁelds, such as collaborative ﬁltering [29] and network
anomaly detection [30]. Multiple stacked RBM layers can
form a deep belief network (DBN), which consists of a
visible layer and multiple hidden layers. The training of a
DBN follows a layer-by-layer method, where each layer is
treated as an RBM trained on top of the previously trained
layer [31]. Many applications can beneﬁt from the structure
of DBNs, such as fault detection classiﬁcation in industrial
environments, threat identiﬁcation in security alert systems,
and emotional feature extraction out of images [17].
B. AUTOENCODER
An autoencoder includes an input layer and an output layer
that are connected by one or multiple hidden layers [32],
as illustrated in Fig. 2(b). The shape of the input layer and
the output layer are the same. The AE can be divided into
two parts, i.e., an encoder and a decoder. The encoder learns
the representative characteristics of the input and transforms
it into other latent features (usually in a compressing way).
4VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
And the decoder receives the latent features of the encoder
and aims to reconstruct the original form of the input data,
minimizing the reconstruction error. Similarly, an AE can be
formed as a deep architecture by stacking multiple layers into
the hidden layer. There are several variants and extensions
of AEs, such as sparse AE [33], denoising AE [34], and
variational AE [35].
C. DEEP NEURAL NETWORKS
Compared to the traditional artiﬁcial neural network (ANN)
that has shallow structure, deep neural network (DNN) (or
deep fully connected neural network) usually has a deeper
layer structure for more complicated learning tasks [32]. A
DNN consists of an input layer, several hidden layers, and
an output layer, where the output of each layer is fed to the
next layer with activation functions. At the last layer, the
ﬁnal output representing the model prediction is produced.
Optimization algorithms such as Stochastic Gradient Decent
(SGD) [36] and backpropagation [37] are mostly used in the
training process. DNNs are widely used in feature extraction,
classiﬁcation and function approximation.
D. CONVOLUTIONAL NEURAL NETWORKS
Convolutional neural networks (CNNs) are designed to pro-
cess data that comes in the form of multiple arrays, for exam-
ple, a color image composed of three 2D arrays containing
pixel intensities in the three color channels [27]. A CNN
receives 2D data structures and extracts high-level features
through convolutional layers as described in Fig. 2(c), which
is the core of CNN architecture. By going through the 2D
data with a set of moving ﬁlters and the pooling functions,
CNN extracts the spatial correlations between adjacent data
by calculating the inner product of the input and the ﬁlter. Af-
ter that, a pooling block is operated over the output to reduce
the spatial dimensions and generate a high-level abstraction.
Compared to traditional fully connected deep networks, CNN
can effectively decrease the parameter numbers of network
and extract the spatial correlations in the raw data, mitigating
the risk of overﬁtting [38]. The above advantages make
CNN achieve signiﬁcant results in many applications, such
as object detection [39] and health monitoring [40].
E. RECURRENT NEURAL NETWORKS
Different from CNNs that are good at abstracting spatial
features, recurrent neural networks (RNNs) are designed for
processing sequential or time-series data. The input to an
RNN includes both the current sample and the previously
observed samples. Speciﬁcally, each neuron of an RNN layer
not only receives the output of its previous layer but also
receives the stored state of from previous time steps, as
depicted in Fig. 2(e). With this special architecture, RNN is
able to remember previous information for integrated pro-
cessing with the current information. However, RNNs can
only look back for a few steps due to the gradient explosion
and long-term dependencies. To solve this problem, Long
Short-Term Memory (LSTM) network [41] is proposed to
control the ﬂow of information. In LSTM model, the forget
gate is utilized to control the cell state and decide what to
keep in the memory. Through the learning process, stored
computations in the memory cells are not distorted over time,
which particularly achieves better performance when data is
characterized in long dependency [42]. RNN and LSTM are
widely used in various sequential scenarios, such as language
processing [43] and activity recognition [44].
F. DEEP REINFORCEMENT LEARNING
Deep reinforcement learning (DRL) [7] is a combination of
deep learning (DL) and reinforcement learning (RL) [45]. It
aims to build an agent that is able to learn the best action
choices over a set of states through the interaction with the
environment, so as to maximize the long-term accumulated
rewards. Different from traditional RL, DRL utilizes a deep
neural network to represent the policy given its strong rep-
resentation ability to approximate the value function or the
direct strategy. DRL can be categorized into value-based
models, such as Deep Q-Learning (DQL), Double DQL [46]
and Duel DQL [47], and policy-gradient-based models, such
as deep deterministic policy gradient (DDPG) [48] and asyn-
chronous advantage actor-critic (A3C) [49]. The DRL has
been successfully applied in many ﬁelds, such as computer
gaming [7], chess gaming [8] and rate adaptation [50].
IV. EMPOWERING EDGE APPLICATIONS WITH DEEP
LEARNING
A. WHEN EDGE COMPUTING MEETS DEEP LEARNING
Recent years have witnessed the rapid development and the
achieved great success of edge computing and deep learning
in their respective ﬁelds. However, the massive amount of
invaluable data generated and collected at the edge side calls
for more powerful and intelligent processing capacities lo-
cally to fully unleash the underlying potentials of big data, so
as to satisfy the ever-increasing demands of various applica-
tions. Fortunately, the recent breakthroughs in deep learning
shed a light on the edge application scenarios, providing
strong ability in information perception, data management,
decision making, etc. The convergence of these two tech-
nologies can further create new opportunities, empowering
the development of many emerging applications. In fact,
artiﬁcial intelligence (AI) to achieve edge intelligence. In this
rest of this section, we conduct a comprehensive overview
of state-of-the-art research works on edge computing appli-
cations with deep learning and summarize them in several
aspects, including smart multimedia, smart transportation,
smart city, and smart industry.
B. SMART MULTIMEDIA
The Internet video content has been explosively increasing
in the past years. As estimated by Cisco, the global video
trafﬁc accounted for 75% of the Internet trafﬁc in 2017
and is estimated to grow four-fold by 2022 [63]. Mean-
while, people are having an increasingly higher demand for
VOLUME 4, 2016 5
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Application Work Model Summary
Video Analytics
Ren et al. [51] R-CNN An objection detection architecture and implementation that leverages Faster R-
CNN for objection detection.
Liu et al. [52] CNN A CNN-based visual food recognition algorithms with edge computing.
DeepDecision [53] CNN, Yolo [54] A distributed infrastructure that ties end and edges to improve Yolo with higher
frame rate and accuracy.
DeepCham [55] CNN Use edge and end user to collaboratively train CNN model for better object
recognition accuracy.
Nikouei et al. [56] L-CNN Develop L-CNN for resource-constrained edge devices with reduced ﬁlter num-
bers.
Adaptive Streaming Grazia [57] RBM, Liner Classiﬁer A multi-stage learning system integrating RMB and liner classiﬁer for simulta-
neous video transmission with guarantees for each user.
Wang et al. [58] DRL An edge computing-assisted framework that leverages DRL to intelligently
assign user to proper edge servers.
Caching
Li et al. [59] ARIMA, MLR,kNN Consider the video propagation and popularity evolution patterns with ARIMA,
MLP, and kNN.
Zhang et al. [60] LSTM-C Propose a caching framework LSTM-C to better learn to content popularity
patterns both at long and short time scale.
Zhu et al. [61] DRL Leverage DRL to automatically learn an end-to-end caching policy.
Jiang et al. [62] MADRL Formulate the D2D caching problem as a multi-agent MAB problem and
developing a multi-agent DRL solution.
TABLE 1: Summary of deep-learning-enabled edge computing applications in the ﬁeld of smart multimedia.
video content and video watching experience, calling for
more intelligent video processing, caching, and delivery, etc.
Nowadays, deep learning is integrated with edge computing
to provide both better video quality of experience (QoE) for
viewers and cost-effective functions to service providers. The
representative researches on smart multimedia applications
are summarized in Tab. 1.
1) Video Analytics
Nowadays, video analytics [9] are becoming more and more
widely used in different ﬁelds such as camera-based surveil-
lance [64] and augmented reality (AR). With the limited
processing capabilities of cameras, traditional video analytics
usually heavily rely on cloud computing for content process-
ing, i.e., the video contents are ﬁrst streamed to the backend
cloud servers and the processed results are then delivered to
the frontend devices. This processing mode however brings
high latency and consumes much bandwidth, unable to sat-
isfy those latency-sensitive applications, not to mention those
realtime requirements such as object detection [51] and track-
ing [65]. The emergence of edge computing pushes the video
analytics from the remote cloud to the local edge, allowing
the video content to be processed near the data source so
as to enable quick or even realtime response. For example,
Amazon has released the world’s ﬁrst deep-learning-enabled
video camera, where the locally executed deep learning func-
tion enables realtime objection even without the involvement
of the cloud.
Pioneer researches have conducted efforts towards intel-
ligent video analytics with edge computing. Ren et al. [51]
proposed an edge-computing-based object detection archi-
tecture as well as a preliminary implementation to achieve
distributed and efﬁcient object detection via wireless com-
munications for real-time surveillance applications. They
adopted Faster R-CNN [39] for model training and object
detection, with a well-designed RoI detection algorithm to
balance the detection accuracy and the data compression
rate. Liu et al. [52] developed a CNN-based visual food
recognition algorithms to achieve the best-in-class recog-
nition accuracy, where edge computing was employed to
overcome the system latency and low battery life of mobile
devices. In DeepDecision [53], a distributed infrastructure
was proposed to tie together computationally weak frontend
devices (assumed to be smartphones) with more powerful
back-end helpers (the edges) to allow deep learning to choose
local or remote execution. This approach boosts the perfor-
mance of CNN, in particular Yolo [54], to achieve higher
frame rate and accuracy. DeepCham [55] leveraged an edge
master server coordinated with several participating users
to collaboratively train CNN model to achieve better object
recognition accuracy. LAVEA [66] built up a client-edge
collaboration system and solved an optimization ofﬂoading
problem to minimize the response time. Nikouei et al. [56]
developed a lightweight convolutional neural network (L-
CNN), which leveraged the depthwise separable convolution
feature and tailored the CNN to be furnished in the resource-
constrained edge devices with reduced ﬁlter numbers in each
layer.
Adaptive video streaming [78] is becoming a critical issue
in today’s video delivery to provide the best quality of ex-
perience (QoE) over the Internet. The basic idea is to select
proper video bitrate considering the network states, stability,
fairness, user’s preference of video quality, etc. Most existing
adaptive streaming approaches [79]–[81] rely on the client-
ations based on several observed or predicted metrics such
as buffer size and bandwidth situation. The recent advances
in deep learning, particularly deep reinforcement learning,
6VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Application Work Model Summary
Autonomous Driving
SqueezeDet [67] CNN A careful designed CNN for objection with reduced model size.
Chen et al. [68] CNN A CNN-based monocular 3D object detection method based on the fact that objects
should be on the ground-plane.
MV3D [69] Fusion VGG A multi-view 3D deep learning network that takes both LIDAR point and camera
images as a fusion input.
Bojarski et al. [70] CNN An end-to-end learning architecture that directly mapped the raw pixels from a single
front-facing camera to the steering commands.
Trafﬁc Analysis
and Prediction
Polson et al. [71] DNN Leveraging deep neural network to mine the short term characteristics of the trafﬁc
situation for trafﬁc prediction.
Lv et al. [72] SAE Leverage a stacked autoencoder (SAE) to learn the generic trafﬁc features from the
historical data.
Koesdwiady et al. [73] DBN Consider the impact of weather on trafﬁc situation and incorporat the weather
information into a deep belief network.
DeepTransport [74] LSTM Consider the mobility analysis at the citywide scale with a LSTM model for
prediction.
Yao et al. [75] CNN,LSTM Revisit the spatiotemporal relationships in trafﬁc pattern and propose a novel spatial-
temporal dynamic network (STDN) based on CNN and LSTM for prediction.
Trafﬁc Signal Control
Wiering et al. [76] Tabular QL Leverage a tabular Q-learning model in an isolated intersection for signal control.
Abdoos et al. [10] MAQL Propose a multi-agent Q-learning (MAQL) method that considers the queue length
for cooperative scheduling.
Chu et al. [77] MA2C Propose a novel multi-agent actor-critic (MA2C) approach to comprehensively
combine the trafﬁc features for intelligent control.
TABLE 2: Summary of deep-learning-enabled edge computing applications in the ﬁeld of smart transportation.
are leveraged at the client end to automatically learn the
adaptation policy for better QoE [50], [82].
The emergence of edge computing provides an alternative
for adaptive video streaming. Given the intelligence and
computation capability, edge nodes can serve as a cache
server or transcoding server to provide edge-based (or edge-
cloud-based) adaptive streaming [83], [84]. This scheme can
usually collect more states from other users and achieve
higher fairness, stability, and collaborative intelligence. For
example, De Grazia [57] developed a multi-stage learning
system to manage simultaneous video transmission which
guarantees a minimum quality level for each user. In par-
ticular, they used an unsupervised Restricted Boltzmann
Machine (RBM) [28] to capture the latent features of the
input data and a supervised linear classiﬁer to estimate the
characteristics of unknown videos. Wang et al. [58] designed
an edge computing-assisted framework that leverages DRL
to intelligently assign users to proper edge servers to achieve
proper video streaming services.
3) Caching
Video content caching [85] is another important applica-
tion that has attracted continuous research efforts for years
given its great beneﬁts in improving multimedia services.
In traditional content delivery network (CDN) architecture,
video contents are usually placed or cached at remote servers,
where the high latency and limited bandwidth between the
end viewers and the remote servers can cause viewing delay
and congestion, seriously undermining the viewers’ QoE.
The emerging edge caching [86], [87] is able to alleviate this
problem by pushing the content close to the end users so as
to reduce the access latency and reduce the network pressure.
Traditional content providers may simply use rule-based so-
lutions such as Least Recently Used (LRU), Least Frequently
Used (LFU) and their variants [88], [89], or model-based
solutions such as [90], [91] given the easy implementation.
These solutions however heavily rely on dedicated features
and are not adaptive enough to the changing characteristics.
Deep learning brings new opportunities towards intelligent
edge caching using advanced learning techniques to well
capture the hidden features. Li et al. [59] considered the video
propagation as well as popularity evolution patterns and de-
veloped an integration of ARIMA, multiple linear regression
(MLR), and k-nearest neighbor regression (kNN) to predict
the social patterns to improve caching performance. Zhang et
al. [60] proposed an intelligent edge-assisted caching frame-
work LSTM-C based on LSTM to better learn to content
popularity patterns both at long and short time scale. Zhu et
al. [61] proposed to leverage DRL to automatically learn an
end-to-end caching policy, where the user requests, network
constraints, and external information are all embedded in the
learning environment. Besides individual caching decisions,
collaborative caching is also explored in recent years to
achieve collective intelligence. For example, Jiang et al. [62]
formulates the D2D caching problem as a multi-agent multi-
armed bandit (MAB) problem and developed a multi-agent
DRL (MADRL) solution to learn a coordinated caching
scheme among multiple agents.
C. SMART TRANSPORTATION
Vehicle is envisioned as the next intelligent information
carrier after smartphone. The coming era of 5G and mobile
edge computing (MEC) has enabled vehicle information to
be readily accessible anytime and anywhere with low latency,
forming an Internet of Vehicle (IoV) [92]. Integrated with
the latest advances in deep learning, IoV will enable more
VOLUME 4, 2016 7
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
intelligent transportation management, such as autonomous
driving [24], trafﬁc prediction, trafﬁc signal control, as sum-
marized in Tab. 2.
1) Autonomous Driving
Intelligent sensing and perception are of the most critical
issues in autonomous driving [24]. The vehicles ﬁrst collect
the information from various carried sensors such as cameras
and radars, and then conduct an intelligent perception and de-
cision. Purely using vehicle-based and cloud-based solutions
may not well satisfy the requirement of high computation
capacity, realtime feedback, enough redundancy, and security
for autonomous driving. Edge computing however provides
a promising solution with powerful computation and low-
latency communication [105]. With the beneﬁts of V2X
communications [106], part of the learning-based perception
can be ofﬂoaded to the edge server for processing.
Many existing works have conducted pioneer efforts to-
wards autonomous driving. SqueezeDet [67] proposed a care-
fully designed CNN-based learning pipeline that not only
achieves high object detection accuracy but also reduces the
model size for energy efﬁciency. To better understand the
captured object, Chen et al. [68] proposed a monocular 3D
object detection method using state-of-the-art CNN method
based on the fact that objects should be on the ground-plane.
To further improve the accuracy and robustness, MV3D [69]
developed a multi-view 3D deep learning network that takes
both LIDAR point and camera images as a fusion input to
predict 3D boundaries. Besides the road feature extraction,
researchers also dived deeply into the driving control based
on the sensing information. Bojarski et al. [70] proposed
an end-to-end learning architecture without detecting road
features. It directly mapped the raw pixels from a single
front-facing camera to the steering commands.
2) Trafﬁc Analysis and Prediction
Understanding the mobility patterns of the vehicles and
people is a critical problem for urban trafﬁc management,
city planning, and service provisioning. Given the distributed
features of mobile edge servers, edge computing is naturally
ideal for vehicle trafﬁc analysis and prediction [107]. Tra-
ditional approaches mostly used time-series analysis [108] or
probabilistic graph analysis [109], which may not sufﬁciently
capture the hidden spatiotemporal relationships therein. As
a powerful learning tool, deep learning stands out as an
effective method in this direction. Liu et al. [110] further
pointed out the potential of applying different deep learning
approaches in urban trafﬁc prediction. Polson et al. [71]
leveraged deep neural network to mine the short term charac-
teristics of the trafﬁc situation of a road segment to predict the
near future trafﬁc pattern. Lv et al. [72] leveraged a stacked
autoencoder (SAE) to learn the generic trafﬁc features from
the historical data. Koesdwiady et al. [73] further considered
the impact of weather on trafﬁc situations and incorporated
the weather information into a deep belief network for inte-
grated learning. In DeepTransport [74], the authors consid-
ered the mobility analysis at a larger scale, i.e., the citywide
scale. LSTM model is used for future movement prediction.
In [75], the authors revisited the spatiotemporal relationships
in trafﬁc patterns and proposed a novel spatial-temporal
dynamic network (STDN) based on CNN and LSTM, which
outperforms the existing prediction methods.
3) Trafﬁc Signal Control
With the above trafﬁc analysis and prediction, the combina-
tion of edge computing and deep learning actually can do
more things towards intelligent transportation management.
Among them, intelligent trafﬁc signal control [76] is one
of the most representative applications and has also been
explored by researchers for years. A good control policy is
able to reduce the average waiting time, trafﬁc congestion,
and trafﬁc accident. The early trafﬁc signal control methods
usually rely on fuzzy logic [111] or genetic algorithm [112].
The key challenge however lies in how to achieve collabora-
tive and intelligent control among multiple or even citywide
trafﬁc lights for large scale trafﬁc scheduling. Towards this
goal, reinforcement learning (RL) and multi-agent RL turns
out to be a promising solution where each agent (can be
implemented as an edge) will make control policy for a trafﬁc
light considering not only its local trafﬁc situation but also
other agents’ trafﬁc situations. In [76], tabular Q-learning
was ﬁrst applied in an isolated intersection for signal control.
To improve the collaboration among trafﬁc lights, Abdoos et
al. [10] proposed a multi-agent Q-learning (MAQL) method
that considered the queue length for cooperative scheduling.
The latest work [77] further integrated the state-of-the-art
actor-critic (A2C) RL algorithm and the multi-agent learning
as a multi-agent actor-critic (MA2C) approach to comprehen-
sively combine the trafﬁc features for intelligent control.
D. SMART CITY
Smart city [113] is another important application scenario for
deep-learning-enabled edge computing. The geo-distributed
big data [114] in a city naturally requires a distributed com-
puting paradigm for local processing and management. The
integration of edge computing and deep learning enables the
deep penetration of computing intelligence into every corner
of a city, forming a smart city that can provide more efﬁcient,
economic, energy-saving, and convenient services [115],
[116]. We next discuss the combinational advantages in the
main components of the smart city, including the smart home,
smart building, and smart grid, as in Tab. 3.
1) Smart Home
Smart IoT has been widely explored in smart home scenarios
to provide not only convenient control but also intelligent
sensing [117]. Considering the privacy issue in the home
scenario, edge computing is a good choice to provide local
computation and processing, especially for the computation-
intensive deep-learning-based applications. Dhakal et al. [93]
developed an automated home/business monitoring system
which resides on Network Function Virtualization (NFV)
8VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Application Work Model Summary
Smart Home
Dhakal et al. [93] kNN,DNN Develop an automated home/business monitoring system on NFV edge servers
performing online learning on streaming data from home.
SignFi [94] CNN Exploit the CSI signals of WiFi and using CNN to identify 276 sign language
gestures.
Wang et al. [95] CNN,LSTM Leverage a combined CNN and LSTM to recognize different gestures and
activities.
Mohammadi et al. [96] DRL Propose a novel semisupervised DRL-based method for indoor localization.
Smart Building
Zheng et al. [97], [98] Multi-task Learning Leverage multi-task learning to predict the performance of a chiller and strike
balance between electricity consumption and real-world deployment.
Yuce et al. [99] NN Propose a neural network based model to perform regression analysis of energy
consumption within a building.
Thokala [100] SVM,RNN Consider the heterogeneity in the electrical load and propose to use both SVM
and partial RNN to forecast future load.
Smart Grid
He et al. [101] DBN,RBM A deep learning based mechanism that integrates DBN and RBM to detect the
attack behavior of false data injection in realtime.
Yan et al. [102] RL A reinforcement learning-based approach to identify critical attack sequences
with consideration of physical system behaviors.
Shi et al. [103] RNN A novel pooling-based RNN network to forecast the household load addressing
the over-ﬁtting issue.
Wan et al. [104] DRL A model-free DRL-based model to automatically determine the charging policy.
TABLE 3: Summary of deep-learning-enabled edge computing applications in the ﬁeld of smart city.
edge servers performing online learning on streaming data
coming from homes and businesses in the neighborhood.
Leveraging the latest advances of deep learning, the ubiqui-
tous wireless signals can also be used for smart interaction
between human and devices. For example, SignFi [94] ex-
ploited the CSI signals of WiFi and was able to identify 276
sign language gestures including the head, arm, hand, and
ﬁnger with CNN for classiﬁcation. Wang et al. [95] analyzed
the impact patterns of moving humans on the WiFi signals
and leveraged a combined CNN and LSTM to recognize
different gestures and activities. Such method can be used
for remote control of home devices such as lights and tele-
visions [118]. Mohammadi et al. [96] explored more pos-
sibilities of deep learning approaches and proposed a novel
semisupervised DRL based method for indoor localization.
Such edge-intelligence-enabled solution can be widely used
for smart home, including intrusion detection, gesture-based
interaction, fall detection, etc.
2) Smart Building
Achieving intelligent monitoring, sensing, and control in the
building environment requires more comprehensive percep-
tion and processing compared to the home environment given
the complex architecture. In this context, edge computing
plays an important role in data processing, orchestration, and
privacy preserving. Recently, many efforts [119]–[121] have
been made towards smart building to reduce energy con-
sumption, improve the building security, enhance the sensing
capacity of buildings, etc. Zheng et al. [97], [98] focused
on the chiller sequencing problem to reduce the electricity
consumption in buildings. They leveraged multi-task learning
to predict the performance of a chiller and further strike a
good balance between the electricity consumption and ease
of use for real-world deployment. Yuce et al. [99] proposed a
neural-network-based model to perform regression analysis
of energy consumption within a building. Thokala [100]
further considered the heterogeneity in the electrical load and
proposed to use both SVM and partial RNN to forecast future
3) Smart Grid
The smart grid is an electricity distribution network with
smart meters deployed at various locations to measure the
realtime status information [127]. Smart grid is also an
important use case for edge computing or fog computing.
Edge collectors at the edge ingest the data generated by
grid sensors and devices, where some data for protection
and control loops even require real-time processing (from
milliseconds to sub-seconds) [3]. Deep learning together with
the edge computing empowers the grid with more intelli-
gent protection, control, and management. He et al. [101]
proposed a deep-learning-based mechanism that integrated
deep belief network (DBN) and RBM to detect the attack
behavior of false data injection in realtime. Considering the
sequential behavior of attacks for smart grid, Yan et al. [102]
further proposed a reinforcement learning-based approach
to identify critical attack sequences with consideration of
physical system behaviors. Shi et al. [103] proposed a novel
pooling-based RNN network to forecast the household load
addressing the over-ﬁtting issue. Pricing is another important
issue towards smart grid, which greatly affects customers’
using behaviors in many aspects, e.g., the economy-driven
electric vehicle charging [128]. For instance, Wan et al. [104]
jointly considered the electricity price and battery energy
of electric vehicles and proposed a model-free DRL based
model to automatically determine the charging policy.
E. SMART INDUSTRY
In the coming era of industry 4.0, we are experiencing a rev-
olution of smart industry, which has two main principles, i.e.,
VOLUME 4, 2016 9
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Application Work Model Summary
Smart Manufacturing
Weimer et al. [122] CNN A novel CNN-based architecture for fast and reliable industrial inspection.
Li et al. [123] CNN An edge-computing-based model that ofﬂoads computation burden to the fog
nodes and a CNN-based model with an early-exit design for accurate inspection.
Zhao et al. [124] CNN, bi-LSTM A system that combined CNN and bi-directional LSTM for machine health
monitoring.
Smart Industrial Analysis Wu et al. [125] vanilla LSTM A remaining life prediction for engineered system using vanilla LSTM neural
networks.
Wang et al. [126] DNN Proposing a DNN-based architecture to accurately predict the remaining energy
and remaining lifetime of batteries, which further enables an informed power
conﬁguration among base stations.
TABLE 4: Summary of deep-learning-enabled edge computing applications in the ﬁeld of smart industry.
production automation and smart data analysis [123]. The
former is one of our main objectives that could greatly liber-
ate the productivity and the latter is one of the most effective
methods towards our objectives. The recent advances of edge
computing migrate the massive computation from the remote
cloud to the local edge, enabling more low-latency and secure
manufacturing [129]. And deep learning further empowers
more effective local analysis and prediction [130] at the edge
node of industry instead of the cloud. We summarize them in
the next two parts as illustrated in Tab. 4.
1) Smart Manufacturing
Smart manufacturing is the key component of the smart
industry, which highly relies on the intelligent processing of
deep learning and quick response of edge computing. The
combination of deep learning and edge computing has been
applied in many aspects of industry manufacturing, such
as manufacture inspection and fault assessment. Weimer et
al. [122] developed a novel CNN-based architecture for fast
and reliable industrial inspection, which can automatically
generate meaningful features for a speciﬁc inspection task
from a huge amount of raw data with minimal human in-
teraction. Li et al. [123] proposed an edge-computing-based
model that is able to ofﬂoad the computation burden to
the fog nodes to deal with extremely large data. A CNN-
based model together with an early-exit design is used in
this model, which largely improved the inspection accuracy
and robustness. Zhao et al. [124] further combined CNN
with bi-directional LSTM to propose a novel machine health
monitoring system.
2) Smart Industrial Analysis
Besides the manufacture inspection and monitoring, the ap-
plication of edge computing and deep learning also enables
much intelligent industrial analysis. Wu et al. [125] focused
on the remaining useful life estimation of the engineered
system and proposed to use vanilla LSTM neural networks to
get good remaining lifetime prediction accuracy. Wang et al.
[126] focused on the remaining lifetime analysis of backup
batteries in the wireless base stations. They proposed to use
DNN-based architecture to accurately predict the remaining
energy and remaining lifetime of batteries, which further en-
ables an informed power conﬁguration among base stations.
V. RESEARCH CHALLENGES AND DIRECTIONS
Though the convergence of edge computing and deep learn-
ing has revealed great potentials and prompted the fast
development of many applications, there still exist various
problems in achieving stable, robust, and practical usage,
which calls for continuous efforts in this ﬁeld from many
perspectives. We next discuss some key research challenges
and promising directions.
A. MODEL TRAINING
The performance of deep-learning-enabled edge applications
highly relies on how the learning models are performed
in the edge computing architecture, where the model train-
ing is an important process. It is well known that model
training is often computation-intensive, consuming massive
CPU and GPU resources, especially for those deep models.
Edge servers are usually challenging, or at least not cost-
efﬁcient, to solely take the model training tasks. Besides,
in many applications, the data is distributed collected from
multiple edge servers and it is difﬁcult for a single edge
server to obtain the whole information for model training.
Sharing the raw data among the edge nodes is not a good
solution since it will consume massive communication re-
sources. Towards this direction, distributed learning [131]
and federated learning [132] are two promising models to
address this problem. The early idea of distributed learning is
to design a decentralized Stochastic Gradient Descent (SGD)
algorithm in the edge computing environment. The key chal-
lenge exists in reducing the communication cost for gradient
updates while preserving the training accuracy. Recent efforts
have been made towards this direction, e.g., delivering the
important gradient ﬁrst [133]. Federated learning is another
promising method emerging in recent years to train deep
neural networks as it leaves the raw data on clients and only
aggregates the intermediate updates from each client. In the
edge application scenario, it can also reduce the communica-
tion cost and improve resource utilization [134].
B. MODEL INFERENCE
As many outstanding learning models are getting bigger
and deeper, traditional large-scale learning models are often
deployed in a centralized cloud and receive the raw input
data from the distributed end devices, which can cause high
10 VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
delay. Edge servers provide alternative solutions for model
deployment, where the edge and cloud can work collabora-
tively to handle the massive amounts of learning takes. A
promising direction is model partition for deep neural net-
works, where the end, edge, and cloud will execute part of the
learning models, respectively. For example, Kang et al. [135]
developed a lightweight scheduler to automatically partition
DNN computation between mobile devices and data centers
at the granularity of neural network layers. And Huang et
al. [136] further explored the partitioning problem in the edge
computing scenario, and developed a partitioning strategy
among the device, edge, and cloud, which aimed to reduce
the execution delay. Another promising direction is the early
exit of inference (EEoI) [137] for deep neural networks.
Since passing through the whole deep networks is both time-
and energy-consuming for edge servers, EEoI allows the
inference to exit early if veriﬁed by some predeﬁned models.
C. APPLICATION ENHANCEMENT
The integration of deep learning and edge computing has
achieved remarkable improvement for many application sce-
narios, yet there are still some critical applications desir-
ing for real breakthroughs. Real-time VR gaming and au-
tonomous driving are two most representative applications.
Both these two applications require ultra low-delay interac-
tions and powerful computations. The emerging 5G commu-
nication technology together with edge learning brings new
possibilities towards a feasible solution, where the main com-
putation such as video rendering as well as video analytics
can be conducted at local edges and the processing results can
be delivered to the end in near real time, e.g., millisecond-
level interaction. Despite of the preliminary foundation of
feasibility, there is still a long way to go before the practical
application.
D. HARDWARE AND SOFTWARE OPTIMIZATION
In addition to the model and application enhancement, the
system-level optimization for deep learning and edge com-
puting is also a challenging yet promising direction. Most
of the existing hardware architecture, software platform,
and programming abstraction are particularly designed for
the cloud-based computing paradigm. Yet the edge learning
emphasizes some different aspects compared to cloud com-
puting, such as energy-efﬁciency, light-weight architecture,
and edge-oriented computation framework. For example,
from the perspective of the hardware architecture optimiza-
tion, Du et al. [138] studied Cortex-M micro-controllers
and proposed a streaming hardware accelerator to better
CNN in edge devices. Besides, FPGA-based edge comput-
ing platforms are also developed to support deep learning
computation ofﬂoading from mobile devices to the edge
FPGA platform [139]. For the software perspective, many
incorporations have proposed their own software platforms
or services to support edge-level learning and computing,
such as Amazon’s Greengrass2and Microsoft’s Azure IoT
Edge3. For the perspective of programming abstraction, there
are also some frameworks specially designed for edge sce-
nario, such as MXNet [140], Tensorﬂow Lite4and CoreML5.
Though with these existing systems and frameworks, there
still need a lot of efforts to integrate them to achieve a
more practical and high-performance system for general edge
learning applications.
VI. CONCLUSION
vances of deep learning can be leveraged to improve the
novel edge computing applications. We ﬁrst introduced the
basic concepts and paradigms of edge computing, highlight-
ing its key advantages. We then present some representative
deep learning models that can be used in edge computing,
such as autoencoder, CNN, RNN, DRL, etc. A comprehen-
sive survey on the latest deep learning empowered edge
computing applications is next conducted from four domains,
including smart multimedia, smart transportation, smart city,
and smart industry. Finally, we discussed the key challenges
and future research directions on improving the intelligent
edge computing applications. We hope this survey is able
to elicit more discussion and inspiration on the convergence
of edge computing and deep learning, so as to facilitate
the development of deep-learning-enabled edge computing
applications.
REFERENCES
[1] “Fog computing and the internet of things: Extend the cloud to where the
things are, white paper,” https://www.cisco.com/c/dam/en_us/solutions/
trends/iot/docs/computing-overview.pdf.
[2] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and
challenges,” IEEE Internet of Things Journal, vol. 3, no. 5, pp. 637–646,
2016.
[3] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its
role in the internet of things,” in Proceedings of ACM MCC Workshop.
ACM, 2012, pp. 13–16.
[4] M. Satyanarayanan, V. Bahl, R. Caceres, and N. Davies, “The case for
vm-based cloudlets in mobile computing,” IEEE pervasive Computing,
2009.
[5] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classiﬁcation
with deep convolutional neural networks,” in Proceedings of NeurIPS,
2012, pp. 1097–1105.
[6] T. Young, D. Hazarika, S. Poria, and E. Cambria, “Recent trends in
deep learning based natural language processing,” IEEE Computational
Intelligence Magazine, vol. 13, no. 3, pp. 55–75, 2018.
[7] V. Mnih, K. Kavukcuoglu et al., “Human-level control through deep
reinforcement learning,” Nature, vol. 518, no. 7540, p. 529, 2015.
[8] D. Silver, A. Huang et al., “Mastering the game of go with deep neural
networks and tree search,” Nature, vol. 529, no. 7587, p. 484, 2016.
[9] G. Ananthanarayanan, P. Bahl, P. Bodík, K. Chintalapudi, M. Philipose,
L. Ravindranath, and S. Sinha, “Real-time video analytics: The killer app
for edge computing,” IEEE Computer, vol. 50, no. 10, pp. 58–67, 2017.
[10] M. Abdoos, N. Mozayani, and A. L. Bazzan, “Trafﬁc light control in non-
stationary environments based on multi agent q-learning,” in Proceedings
of IEEE ITSC, 2011, pp. 1580–1585.
2https://aws.amazon.com/cn/greengrass/
3https://azure.microsoft.com/en-ca/services/iot- edge/
4https://www.tensorﬂow.org/lite
5https://developer.apple.com/documentation/coreml?language=objc
VOLUME 4, 2016 11
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
[11] Y. Han, X. Wang, V. Leung, D. Niyato, X. Yan, and X. Chen, “Conver-
gence of edge computing and deep learning: A comprehensive survey,”
arXiv preprint arXiv:1907.08349, 2019.
[12] J. Chen and X. Ran, “Deep learning with edge computing: A review,”
Proceedings of the IEEE, vol. 107, no. 8, pp. 1655–1674, 2019.
[13] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelli-
gence: Paving the last mile of artiﬁcial intelligence with edge computing,
arXiv preprint arXiv:1905.10083, 2019.
[14] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A
survey on mobile edge networks: Convergence of computing, caching
and communications,” IEEE Access, vol. 5, pp. 6757–6779, 2017.
[15] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge com-
puting: A survey,” IEEE Internet of Things Journal, vol. 5, no. 1, pp.
450–465, 2017.
[16] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on
mobile edge computing: The communication perspective,” IEEE Com-
munications Surveys & Tutorials, vol. 19, no. 4, pp. 2322–2358, 2017.
[17] M. Mohammadi, A. Al-Fuqaha, S. Sorour, and M. Guizani, “Deep
learning for iot big data and streaming analytics: A survey,” IEEE
Communications Surveys & Tutorials, vol. 20, no. 4, pp. 2923–2960,
2018.
[18] X. Ma, T. Yao, M. Hu, Y. Dong, W. Liu, F. Wang, and J. Liu, “A survey
on deep learning empowered iot applications,” IEEE Access, vol. 7, pp.
181 721–181 732, 2019.
[19] T. Verbelen, P. Simoens, F. De Turck, and B. Dhoedt, “Cloudlets: Bring-
ing the cloud to the mobile user,” in Proceedings of ACM MCS, 2012,
pp. 29–36.
[20] K. Bilal, O. Khalid, A. Erbad, and S. U. Khan, “Potentials, trends, and
prospects in edge technologies: Fog, cloudlet, mobile edge, and micro
data centers,” Computer Networks, vol. 130, pp. 94–120, 2018.
[21] M. Aazam and E.-N. Huh, “Fog computing micro datacenter based
dynamic resource estimation and pricing model for iot,” in Proceedings
of IEEE AINA, 2015, pp. 687–694.
[22] F. Bonomi, R. Milito, P. Natarajan, and J. Zhu, “Fog computing: A
platform for internet of things and analytics,” in Big data and internet
of things: A roadmap for smart environments. Springer, 2014, pp. 169–
186.
[23] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile edge
computing - a key technology towards 5g,” ETSI white paper, vol. 11,
no. 11, pp. 1–16, 2015.
[24] S. Liu, L. Liu, J. Tang, B. Yu, Y. Wang, and W. Shi, “Edge computing for
autonomous driving: Opportunities and challenges,” Proceedings of the
IEEE, vol. 107, no. 8, pp. 1697–1716, 2019.
[25] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Context
aware computing for the internet of things: A survey,” IEEE communica-
tions surveys & tutorials, vol. 16, no. 1, pp. 414–454, 2013.
[26] C. Wang, G. Liu, H. Huang, W. Feng, K. Peng, and L. Wang, “Miasec:
Enabling data indistinguishability against membership inference attacks
in mlaas,” IEEE Transactions on Sustainable Computing, 2019.
[27] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, p. 436, 2015.
[28] A. Fischer and C. Igel, “An introduction to restricted boltzmann ma-
chines,” in Iberoamerican Congress on Pattern Recognition. Springer,
2012, pp. 14–36.
[29] R. Salakhutdinov, A. Mnih, and G. Hinton, “Restricted boltzmann ma-
chines for collaborative ﬁltering,” in Proceedings of ACM ICML, 2007,
pp. 791–798.
[30] U. Fiore, F. Palmieri, A. Castiglione, and A. De Santis, “Network
anomaly detection with the restricted boltzmann machine,” Neurocom-
puting, vol. 122, pp. 13–23, 2013.
[31] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554,
2006.
[32] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,
2016, http://www.deeplearningbook.org.
[33] H. Lee, A. Battle, R. Raina, and A. Y. Ng, “Efﬁcient sparse coding
algorithms,” in Proceedings of NeurIPS, 2007, pp. 801–808.
[34] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, “Extracting
and composing robust features with denoising autoencoders,” in Proceed-
ings of ACM ICML, 2008, pp. 1096–1103.
[35] C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint
arXiv:1606.05908, 2016.
[36] L. Bottou, “Large-scale machine learning with stochastic gradient de-
scent,” in Proceedings of COMPSTAT. Springer, 2010, pp. 177–186.
[37] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning represen-
tations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–
536, 1986.
[38] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolu-
tional networks,” in Proceedings of ECCV. Springer, 2014, pp. 818–
833.
[39] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-
time object detection with region proposal networks,” in Proceedings of
NeurIPS, 2015, pp. 91–99.
[40] N. Y. Hammerla, S. Halloran, and T. Plötz, “Deep, convolutional, and
recurrent models for human activity recognition using wearables,” arXiv
preprint arXiv:1604.08880, 2016.
[41] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[42] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of
gated recurrent neural networks on sequence modeling,” arXiv preprint
arXiv:1412.3555, 2014.
[43] Q. Chen, X. Zhu, Z. Ling, S. Wei, H. Jiang, and D. Inkpen, “Enhanced
lstm for natural language inference,” arXiv preprint arXiv:1609.06038,
2016.
[44] Y. Guan and T. Plötz, “Ensembles of deep lstm learners for activity
recognition using wearables,” Proceedings of ACM IMWUT, vol. 1,
no. 2, p. 11, 2017.
[45] R. S. Sutton, A. G. Barto et al., Introduction to reinforcement learning.
MIT press Cambridge, 1998, vol. 2, no. 4.
[46] H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning
with double q-learning,” in Proceedings of AAAI, 2016.
[47] Z. Wang, T. Schaul, M. Hessel, H. Van Hasselt, M. Lanctot, and
N. De Freitas, “Dueling network architectures for deep reinforcement
learning,” arXiv preprint arXiv:1511.06581, 2015.
[48] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver,
and D. Wierstra, “Continuous control with deep reinforcement learning,
arXiv preprint arXiv:1509.02971, 2015.
[49] V. Mnih, A. P. Badia et al., “Asynchronous methods for deep reinforce-
ment learning,” in Proceedings of ICML, 2016, pp. 1928–1937.
[50] H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video streaming
with pensieve,” in Proceedings of ACM SIGCOMM, 2017, pp. 197–210.
[51] J. Ren, Y. Guo, D. Zhang, Q. Liu, and Y. Zhang, “Distributed and efﬁcient
object detection in edge computing: Challenges and solutions,” IEEE
Network, vol. 32, no. 6, pp. 137–143, 2018.
[52] C. Liu, Y. Cao, Y. Luo, G. Chen, V. Vokkarane, M. Yunsheng, S. Chen,
and P. Hou, “A new deep learning-based food recognition system for
dietary assessment on an edge computing service infrastructure,” IEEE
Transactions on Services Computing, vol. 11, no. 2, pp. 249–261, 2017.
[53] X. Ran, H. Chen, X. Zhu, Z. Liu, and J. Chen, “Deepdecision: A mobile
deep learning framework for edge video analytics,” in Proceeding of
IEEE INFOCOM, 2018, pp. 1421–1429.
[54] J. Redmon and A. Farhadi, “Yolo9000: better, faster, stronger,” in Pro-
ceedings of IEEE CVPR, 2017, pp. 7263–7271.
[55] D. Li, T. Salonidis, N. V. Desai, and M. C. Chuah, “Deepcham: Collab-
orative edge-mediated adaptive deep learning for mobile object recogni-
tion,” in Proceedings of IEEE/ACM SEC, 2016, pp. 64–76.
[56] S. Y. Nikouei, Y. Chen, S. Song, R. Xu, B.-Y. Choi, and T. Faughnan,
“Smart surveillance as an edge network service: From harr-cascade, svm
to a lightweight cnn,” in Proceedings of IEEE CIC, 2018, pp. 256–265.
[57] M. D. F. De Grazia, D. Zucchetto, A. Testolin, A. Zanella, M. Zorzi, and
M. Zorzi, “Qoe multi-stage machine learning for dynamic video stream-
ing,” IEEE Transactions on Cognitive Communications and Networking,
vol. 4, no. 1, pp. 146–161, 2017.
[58] F. Wang, C. Zhang, J. Liu, Y. Zhu, H. Pang, L. Sun et al., “Intelligent
edge-assisted crowdcast with deep reinforcement learning for personal-
ized qoe,” in Proceedings of IEEE INFOCOM, 2019, pp. 910–918.
[59] H. Li, X. Ma, F. Wang, J. Liu, and K. Xu, “On popularity prediction of
videos shared in online social networks,” in Proceedings of ACM CIKM,
2013, pp. 169–178.
[60] C. Zhang, H. Pang, J. Liu, S. Tang, R. Zhang, D. Wang, and L. Sun,
“Toward edge-assisted video content intelligent caching with long short-
term memory learning,” IEEE Access, vol. 7, pp. 152832–152 846, 2019.
[61] H. Zhu, Y. Cao, W. Wang, T. Jiang, and S. Jin, “Deep reinforcement
learning for mobile edge caching: Review, new features, and open issues,
IEEE Network, vol. 32, no. 6, pp. 50–57, 2018.
[62] W. Jiang, G. Feng, S. Qin, T. S. P. Yum, and G. Cao, “Multi-agent
reinforcement learning for efﬁcient content caching in mobile d2d net-
12 VOLUME 4, 2016
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
works,” IEEE Transactions on Wireless Communications, vol. 18, no. 3,
pp. 1610–1622, 2019.
[63] “Cisco visual networking index: Forecast and trends,
2017 to 2022 white paper,” https://www.cisco.com/c/en/us/
solutions/collateral/service-provider/visual- networking-index-vni/
white-paper-c11-741490.html.
[64] P. Natarajan, P. K. Atrey, and M. Kankanhalli, “Multi-camera coordina-
tion and control in surveillance systems: A survey,” ACM Transactions
on Multimedia Computing, Communications, and Applications, vol. 11,
no. 4, p. 57, 2015.
[65] C.-C. Hung, G. Ananthanarayanan, P. Bodik, L. Golubchik, M. Yu,
P. Bahl, and M. Philipose, “Videoedge: Processing camera streams using
hierarchical clusters,” in Proceedings of IEEE/ACM SEC. IEEE, 2018,
pp. 115–131.
[66] S. Yi, Z. Hao, Q. Zhang, Q. Zhang, W. Shi, and Q. Li, “Lavea: Latency-
aware video analytics on edge computing platform,” in Proceedings of
ACM/IEEE SEC, 2017, p. 15.
[67] B. Wu, F. Iandola, P. H. Jin, and K. Keutzer, “Squeezedet: Uniﬁed, small,
low power fully convolutional neural networks for real-time object detec-
tion for autonomous driving,” in Proceedings of IEEE CVPR Workshops,
2017, pp. 129–137.
[68] X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler, and R. Urtasun,
“Monocular 3d object detection for autonomous driving,” in Proceedings
of IEEE CVPR, 2016, pp. 2147–2156.
[69] X. Chen, H. Ma, J. Wan, B. Li, and T. Xia, “Multi-view 3d object
detection network for autonomous driving,” in Proceedings of the IEEE
CVPR, 2017, pp. 1907–1915.
[70] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal,
L. D. Jackel, M. Monfort, U. Muller, J. Zhang et al., “End to end learning
for self-driving cars,” arXiv preprint arXiv:1604.07316, 2016.
[71] N. G. Polson and V. O. Sokolov, “Deep learning for short-term trafﬁc ﬂow
prediction,” Transportation Research Part C: Emerging Technologies,
vol. 79, pp. 1–17, 2017.
[72] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Trafﬁc ﬂow prediction
with big data: a deep learning approach,” IEEE Transactions on Intelli-
gent Transportation Systems, vol. 16, no. 2, pp. 865–873, 2014.
[73] A. Koesdwiady, R. Soua, and F. Karray, “Improving trafﬁc ﬂow pre-
diction with weather information in connected cars: A deep learning
approach,” IEEE Transactions on Vehicular Technology, vol. 65, no. 12,
pp. 9508–9517, 2016.
[74] X. Song, H. Kanasugi, and R. Shibasaki, “Deeptransport: Prediction and
simulation of human mobility and transportation mode at a citywide
level.” in Proceedings of IJCAI, vol. 16, 2016, pp. 2618–2624.
[75] H. Yao, X. Tang, H. Wei, G. Zheng, and Z. Li, “Revisiting spatial-
temporal similarity: A deep learning framework for trafﬁc prediction,
in Proceedings of AAAI, 2019.
[76] M. Wiering, J. v. Veenen, J. Vreeken, and A. Koopman, “Intelligent trafﬁc
light control,” 2004.
[77] T. Chu, J. Wang, L. Codecà, and Z. Li, “Multi-agent deep reinforcement
learning for large-scale trafﬁc signal control,” IEEE Transactions on
Intelligent Transportation Systems, 2019.
[78] A. Bentaleb, B. Taani, A. C. Begen, C. Timmerer, and R. Zimmermann,
“A survey on bitrate adaptation schemes for streaming media over http,
IEEE Communications Surveys & Tutorials, vol. 21, no. 1, pp. 562–585,
2018.
[79] K. Spiteri, R. Urgaonkar, and R. K. Sitaraman, “Bola: Near-optimal
bitrate adaptation for online videos,” in Proceedings of IEEE INFOCOM,
2016, pp. 1–9.
[80] Y. Sun, X. Yin, J. Jiang, V.Sekar, F. Lin, N. Wang, T. Liu, and B. Sinopoli,
“Cs2p: Improving video bitrate selection and adaptation with data-driven
throughput prediction,” in Proceedings of ACM SIGCOMM, 2016, pp.
272–285.
[81] T.-Y. Huang, R. Johari, N. McKeown, M. Trunnell, and M. Watson, “A
buffer-based approach to rate adaptation: Evidence from a large video
streaming service,” in ACM SIGCOMM Computer Communication Re-
view, vol. 44, no. 4, 2014, pp. 187–198.
[82] T. Huang, R.-X. Zhang, C. Zhou, and L. Sun, “Qarc: Video quality aware
rate control for real-time video streaming based on deep reinforcement
learning,” in Proceedings of ACM Multimedia, 2018, pp. 1208–1216.
[83] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborative
mobile edge computing in 5g networks: New paradigms, scenarios, and
challenges,” IEEE Communications Magazine, vol. 55, no. 4, pp. 54–61,
2017.
[84] D. Wang, Y. Peng, X. Ma, W. Ding, H. Jiang, F. Chen, and J. Liu, “Adap-
tive wireless video streaming based on edge computing: Opportunities
and approaches,” IEEE Transactions on services Computing, vol. 12,
no. 5, pp. 685–697, 2018.
[85] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. Leung, “Cache in the
air: Exploiting content caching and delivery techniques for 5g systems,
IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, 2014.
[86] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role
of proactive caching in 5g wireless networks,” IEEE Communications
Magazine, vol. 52, no. 8, pp. 82–89, 2014.
[87] E. Zeydan, E. Bastug, M. Bennis, M. A. Kader, I. A. Karatepe, A. S. Er,
and M. Debbah, “Big data caching for networking: Moving from cloud to
edge,” IEEE Communications Magazine, vol. 54, no. 9, pp. 36–42, 2016.
[88] S. Podlipnig and L. Böszörmenyi, “A survey of web cache replacement
strategies,” ACM Computing Surveys, vol. 35, no. 4, pp. 374–398, 2003.
[89] Q. Huang, K. Birman, R. Van Renesse, W. Lloyd, S. Kumar, and H. C. Li,
“An analysis of facebook photo caching,” in Proceedings of ACM SOSP,
2013, pp. 167–181.
[90] J. Hachem, N. Karamchandani, and S. Diggavi, “Content caching and
delivery over heterogeneous wireless networks,” in Proceedings of IEEE
INFOCOM, 2015, pp. 756–764.
[91] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch, and
G. Caire, “Femtocaching: Wireless content delivery through distributed
caching helpers,” IEEE Transactions on Information Theory, vol. 59,
no. 12, pp. 8402–8413, 2013.
[92] F. Wang, F. Wang, X. Ma, and J. Liu, “Demystifying the crowd intelli-
gence in last mile parcel delivery for smart cities,” IEEE Network, vol. 33,
no. 2, pp. 23–29, 2019.
[93] A. Dhakal and K. Ramakrishnan, “Machine learning at the network
edge for automated home intrusion monitoring,” in Proceedings of IEEE
ICNP, 2017, pp. 1–6.
[94] Y. Ma, G. Zhou, S. Wang, H. Zhao, and W. Jung, “Signﬁ: Sign language
recognition using wiﬁ,” Proceedings of ACM IMWUT, vol. 2, no. 1,
p. 23, 2018.
[95] F. Wang, W. Gong, and J. Liu, “On spatial diversity in wiﬁ-based human
activity recognition: A deep learning-based approach,” IEEE Internet of
Things Journal, vol. 6, no. 2, pp. 2035–2047, 2018.
[96] M. Mohammadi, A. Al-Fuqaha, M. Guizani, and J.-S. Oh, “Semisu-
pervised deep reinforcement learning in support of iot and smart city
services,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 624–635,
2017.
[97] Z. Zheng, Q. Chen, C. Fan, N. Guan, A. Vishwanath, D. Wang, and F. Liu,
“Data driven chiller sequencing for reducing hvac electricity consump-
tion in commercial buildings,” in Proceedings of ACM e-Energy, 2018,
pp. 236–248.
[98] Z. Zheng, Q. Chen, C. Fan, N. Guan et al., “An edge based data-driven
chiller sequencing framework for hvac electricity consumption reduction
in commercial buildings,” IEEE Transactions on Sustainable Computing,
2019.
[99] B. Yuce and Y. Rezgui, “An ann-ga semantic rule-based system to reduce
the gap between predicted and actual energy consumption in buildings,
IEEE Transactions on Automation Science and Engineering, vol. 14,
no. 3, pp. 1351–1363, 2015.
[100] N. K. Thokala, A. Bapna, and M. G. Chandra, “A deployable electrical
load forecasting solution for commercial buildings,” in Proceedings of
IEEE ICIT, 2018, pp. 1101–1106.
[101] Y. He, G. J. Mendis, and J. Wei, “Real-time detection of false data injec-
tion attacks in smart grid: A deep learning-based intelligent mechanism,”
IEEE Transactions on Smart Grid, vol. 8, no. 5, pp. 2505–2516, 2017.
[102] J. Yan, H. He, X. Zhong, and Y. Tang, “Q-learning-based vulnerability
analysis of smart grid against sequential topology attacks,” IEEE Trans-
actions on Information Forensics and Security, vol. 12, no. 1, pp. 200–
210, 2016.
[103] H. Shi, M. Xu, and R. Li, “Deep learning for household load forecasting
- a novel pooling deep rnn,” IEEE Transactions on Smart Grid, vol. 9,
no. 5, pp. 5271–5280, 2017.
[104] Z. Wan, H. Li, H. He, and D. Prokhorov, “Model-free real-time ev charg-
ing scheduling based on deep reinforcement learning,” IEEE Transactions
on Smart Grid, 2018.
[105] Q. Yuan, H. Zhou, J. Li, Z. Liu, F. Yang, and X. S. Shen, “Toward efﬁcient
content delivery for automated driving services: An edge computing
solution,” IEEE Network, vol. 32, no. 1, pp. 80–86, 2018.
[106] L. Hobert, A. Festag, I. Llatser, L. Altomare, F. Visintainer, and A. Ko-
vacs, “Enhancements of v2x communication in support of cooperative
VOLUME 4, 2016 13
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
autonomous driving,” IEEE communications magazine, vol. 53, no. 12,
pp. 64–70, 2015.
[107] I. Lana, J. Del Ser, M. Velez, and E. I. Vlahogianni, “Road trafﬁc fore-
casting: recent advances and new challenges,” IEEE Intelligent Trans-
portation Systems Magazine, vol. 10, no. 2, pp. 93–109, 2018.
[108] B. Ghosh, B. Basu, and M. O’Mahony, “Bayesian time-series model for
short-term trafﬁc ﬂow forecasting,” Journal of transportation engineering,
vol. 133, no. 3, pp. 180–189, 2007.
[109] S. Sun and X. Xu, “Variational inference for inﬁnite mixtures of gaussian
processes with applications to trafﬁc ﬂow prediction,” IEEE Transactions
on Intelligent Transportation Systems, vol. 12, no. 2, pp. 466–475, 2010.
[110] Z. Liu, Z. Li, K. Wu, and M. Li, “Urban trafﬁc prediction from mobility
data using deep learning,” IEEE Network, vol. 32, no. 4, pp. 40–46, 2018.
[111] B. P. Gokulan and D. Srinivasan, “Distributed geometric fuzzy multiagent
urban trafﬁc signal control,” IEEE Transactions on Intelligent Transporta-
tion Systems, vol. 11, no. 3, pp. 714–727, 2010.
[112] H. Ceylan and M. G. Bell, “Trafﬁc signal timing optimisation based on
genetic algorithm approach, including drivers’ routing,” Transportation
Research Part B: Methodological, vol. 38, no. 4, pp. 329–342, 2004.
[113] F. Cicirelli, A. Guerrieri, G. Spezzano, and A. Vinci, “An edge-based plat-
form for dynamic smart city applications,” Future Generation Computer
Systems, vol. 76, pp. 106–118, 2017.
[114] X. He, K. Wang, H. Huang, and B. Liu, “Qoe-driven big data architecture
for smart city,” IEEE Communications Magazine, vol. 56, no. 2, pp. 88–
93, 2018.
[115] Y. Liu, C. Yang, L. Jiang, S. Xie, and Y. Zhang, “Intelligent edge com-
puting for iot-based energy management in smart cities,” IEEE Network,
vol. 33, no. 2, pp. 111–117, 2019.
[116] M. Mohammadi and A. Al-Fuqaha, “Enabling cognitive smart cities
using big data and machine learning: Approaches and challenges,” IEEE
Communications Magazine, vol. 56, no. 2, pp. 94–101, 2018.
[117] H. Jiang, C. Cai, X. Ma, Y. Yang, and J. Liu, “Smart home based on wiﬁ
sensing: A survey,” IEEE Access, vol. 6, pp. 13 317–13 325, 2018.
[118] Q. Pu, S. Gupta, S. Gollakota, and S. Patel, “Whole-home gesture
recognition using wireless signals,” in Proceedings of ACM MobiCom,
2013, pp. 27–38.
[119] W. Kleiminger, S. Santini, and F. Mattern, “Smart heating control with
occupancy prediction: how much can one save?” in Proceedings of ACM
Ubicomp, 2014, pp. 947–954.
[120] A. Vishwanath, V. Chandan, C. Mendoza, and C. Blake, “A data driven
pre-cooling framework for energy cost optimization in commercial build-
ings,” in Proceedings of ACM e-Energy, 2017, pp. 157–167.
[121] F.-J. Ferrández-Pastor, H. Mora, A. Jimeno-Morenilla, and B. Volckaert,
“Deployment of iot edge and fog computing technologies to develop
smart building services,” Sustainability, vol. 10, no. 11, p. 3832, 2018.
[122] D. Weimer, B. Scholz-Reiter, and M. Shpitalni, “Design of deep convo-
lutional neural network architectures for automated feature extraction in
industrial inspection,” CIRP Annals, vol. 65, no. 1, pp. 417–420, 2016.
[123] L. Li, K. Ota, and M. Dong, “Deep learning for smart industry: Efﬁcient
manufacture inspection system with fog computing,” IEEE Transactions
on Industrial Informatics, vol. 14, no. 10, pp. 4665–4673, 2018.
[124] R. Zhao, R. Yan, J. Wang, and K. Mao, “Learning to monitor machine
health with convolutional bi-directional lstm networks,” Sensors, vol. 17,
no. 2, p. 273, 2017.
[125] Y. Wu, M. Yuan, S. Dong, L. Lin, and Y. Liu, “Remaining useful life
estimation of engineered systems using vanilla lstm neural networks,
Neurocomputing, vol. 275, pp. 167–179, 2018.
[126] F. Wang, X. Fan, F. Wang, and J. Liu, “Backup battery analysis and allo-
cation against power outage for cellular base stations,” IEEE Transactions
on Mobile Computing, vol. 18, no. 3, pp. 520–533, 2018.
[127] S. Yi, Z. Hao, Z. Qin, and Q. Li, “Fog computing: Platform and applica-
tions,” in Proceedings of IEEE HotWeb, 2015, pp. 73–78.
[128] W. Shuai, P. Maillé, and A. Pelov, “Charging electric vehicles in the smart
city: A survey of economy-driven approaches,” IEEE Transactions on
Intelligent Transportation Systems, vol. 17, no. 8, pp. 2089–2106, 2016.
[129] J. Wang, Y. Ma, L. Zhang, R. X. Gao, and D. Wu, “Deep learning for
smart manufacturing: Methods and applications,” Journal of Manufactur-
ing Systems, vol. 48, pp. 144–156, 2018.
[130] F. Wang, F. Wang, X. Fan, and J. Liu, “Batalloc: Effective battery
allocation against power outage for cellular base stations,” in Proceedings
of ACM e-Energy, 2017, pp. 234–241.
[131] G. Kamath, P. Agnihotri, M. Valero, K. Sarker, and W.-Z. Song, “Pushing
analytics to the edge,” in Proceedings of IEEE GLOBECOM, 2016, pp.
1–6.
[132] J. Koneˇ
cn`
y, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and
D. Bacon, “Federated learning: Strategies for improving communication
efﬁciency,” arXiv preprint arXiv:1610.05492, 2016.
[133] Y. Lin, S. Han, H. Mao, Y. Wang, and W. J. Dally, “Deep gradient
compression: Reducing the communication bandwidth for distributed
training,” arXiv preprint arXiv:1712.01887, 2017.
[134] W. Y. B. Lim, N. C. Luong, D. T. Hoang, Y. Jiao, Y.-C. Liang, Q. Yang,
D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A
comprehensive survey,” arXiv preprint arXiv:1909.11875, 2019.
[135] Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and
L. Tang, “Neurosurgeon: Collaborative intelligence between the cloud
and mobile edge,” in Proceedings of ACM SIGARCH, vol. 45, no. 1,
2017, pp. 615–629.
[136] Y. Huang, F. Wang, F. Wang, and J. Liu, “Deepar: A hybrid device-edge-
cloud execution framework for mobile deep learning applications,” in
Proceedings of IEEE INFOCOM WKSHPS, 2019, pp. 892–897.
[137] S. Teerapittayanon, B. McDanel, and H.-T. Kung, “Branchynet: Fast
inference via early exiting from deep neural networks,” in Proceedings
of IEEE ICPR, 2016, pp. 2464–2469.
[138] L. Du, Y. Du, Y. Li, J. Su, Y.-C. Kuan, C.-C. Liu, and M.-C. F. Chang, “A
reconﬁgurable streaming deep convolutional neural network accelerator
for internet of things,” IEEE Transactions on Circuits and Systems,
vol. 65, no. 1, pp. 198–208, 2017.
[139] S. Jiang, D. He, C. Yang, C. Xu, G. Luo, Y. Chen, Y. Liu, and J. Jiang,
“Accelerating mobile applications at the network edge with software-
programmable fpgas,” in Proceedings of IEEE INFOCOM, 2018, pp. 55–
62.
[140] T. Chen, M. Li, Y. Li, M. Lin, N. Wang, M. Wang, T. Xiao, B. Xu,
C. Zhang, and Z. Zhang, “Mxnet: A ﬂexible and efﬁcient machine
learning library for heterogeneous distributed systems,” arXiv preprint
arXiv:1512.01274, 2015.
14 VOLUME 4, 2016
... It is computationally expensive, consuming massive CPU and GPU resources, power, memory, and time. However, edge devices have not yet become cost-efficient enough to take training DL models due to limited resources [7][8][9]. For instance, in [10] it is shown that the training of DL is the main obstacle in integrating DL with the industries internet of things (IIoT), due to the complications of DL models which take up time in the training phase. ...
... Currently, DL is partially deployed on edge devices and the remaining data are transferred to processing in the cloud, or DL models are deployed in the cloud to process raw data that are received from edge devices, thus causing a delay in latency [9]. For instance, Ref. [11] edge devices are used to sense water data and then transfer the data to the cloud to analyze; data can also be forecast using DL models. ...
Article
Full-text available
Recently, the Internet of Things (IoT) has gained a lot of attention, since IoT devices areplaced in various fields. Many of these devices are based on machine learning (ML) models, which render them intelligent and able to make decisions. IoT devices typically have limited resources,which restricts the execution of complex ML models such as deep learning (DL) on them. In addition,connecting IoT devices to the cloud to transfer raw data and perform processing causes delayedsystem responses, exposes private data and increases communication costs. Therefore, to tackle theseissues, there is a new technology called Tiny Machine Learning (TinyML), that has paved the wayto meet the challenges of IoT devices. This technology allows processing of the data locally on thedevice without the need to send it to the cloud. In addition, TinyML permits the inference of MLmodels, concerning DL models on the device as a Microcontroller that has limited resources. The aim of this paper is to provide an overview of the revolution of TinyML and a review of tinyML studies, wherein the main contribution is to provide an analysis of the type of ML models used in TinyML studies; it also presents the details of datasets and the types and characteristics of the devices with an aim to clarify the state of the art and envision development requirements.
... However, even a low-resolution image requires billions of multiply and accumulation operations (MACs) for inference [24]. While offloading inference to cloud centres with powerful GPU servers is possible in some applications, such as biomedical image analysis systems, it results in increased latency, which can significantly degrade the performance of many applications and pose potential privacy and security risks [25]. For example, transmitting high-resolution video from remote security camera nodes to a remote cloud centre consumes considerable energy and results in network congestion and increased latency, which is prohibitively slow for a real-time security monitoring system. ...
Thesis
Object detection in images is one of the most successful applications of convolutional neural networks (CNNs). However, applying deep CNNs to large numbers of video frames has recently emerged as a new challenge beyond image data due to the high computational requirements. Due to their similar appearances, CNNs often extract similar features from video frames. Conventional video object detection pipelines extract features of individual frames with a fixed computational effort, resulting in numerous redundant computations and an inefficient use of energy resources, particularly for edge computing. By exploiting frame-to-frame similarity, this thesis shows that the computational complexity of video object detection pipelines can be reduced. Similarity-aware CNNs are proposed to identify and avoid computations on similar feature pixels across frames. The proposed similarity-aware quantization scheme (SQS) increases the average number of unchanged feature pixels across frame pairs by up to 85% with a loss of less than 1% in detection accuracy. Second, by minimising redundant computations and memory accesses across frame pairs, the proposed similarity-aware row stationary (SRS) dataflow reduces energy consumption. According to simulation experiments, the proposed dataflow reduces video frame processing energy consumption by up to 30%. To further improve the efficiency of video object detection, a new temporal early exit module (TEEM) is proposed. Semantic differences between consecutive frames can be detected using TEEM with low computation overhead, avoiding redundant video frame feature extraction. Multiple TEEMs are inserted into the pipelines’ feature network at various early layers. TEEM-enabled pipelines only require full computation effort when a frame is determined to be semantically distinct from previous frames; otherwise, previous frame detection results are reused. Experiments on ImangeNet VID and TVnet demonstrate that TEEMs accelerate SOTA video object detection pipelines by 1.7× while maintaining a < 1% mean average precision reduction.
... [3]. The authors of [4] have highlighted the advantages of edge computing as low latency, energy saving, context awareness, privacy, and security. Scale reduction in edge computing makes these networks less prone to attacks as compared to the largescale data centers of the cloud. ...
Article
The pivotal role of data security in mobile edge-computing environments forms the foundation for the proposed work. Anomalies and outliers in the sensory data due to network attacks will be a prominent concern in real time. Sensor samples will be considered from a set of sensors at a particular time instant as far as the confidence level on the decision remains on par with the desired value. A “true” on the hypothesis test eventually means that the sensor has shown signs of anomaly or abnormality and samples have to be immediately ceased from being retrieved from the sensor. A deep learning Actor-Criticbased Reinforcement algorithm proposed will be able to detect anomalies in the form of binary indicators and hence decide when to withdraw from receiving further samples from specific sensors. The posterior trust value influences the value of the confidence interval and hence the probability of anomaly detection. The paper exercises a single-tailed normal function to determine the range of the posterior trust metric. The decision taken by the prediction model will be able to detect anomalies with a good percentage of anomaly detection accuracy.
... Hence, how to deal with the trade-off between optimality and efficiency shall be considered in depth according to different scenarios and requests [62]. Since how DNNs serve edge intelligence is not the focus of this paper, more information could be obtained in papers [10,33,62,66]. ...
... Still, better QoS is provided by them and also the latency which is low for the end users. The focus on the architecture of both is important to understand the advantages and the disadvantages and also compare them (Wang et al. 2020). The formation of edge computing is split generally into three features such as the front-end, near-end and far-end. ...
Article
Full-text available
The current text is based on the survey on Edge Computing and Cloud Computing in IoT. In the study, the aims and objectives are mentioned. The conventional components of IoT are discussed in the current study. Edge computing is also described in the present context. The edge computing architecture is also depicted in the current study. In this study, a secondary data collection method is used to get the data. Quantitative data makes the research accurate and also maintains the entire research.
Chapter
Researchers have proposed various structured Convolutional Neural Networks (CNNs) pruning strategies to make them work efficiently on edge devices. However, most of them focus on reducing the number of filter channels per layer without considering the redundancy within individual filter channels. In this work, we explore pruning from another dimension, the kernel size. We develop a CNN pruning framework called SMOF, which Squeezes More Out of Filters by reducing both kernel size and the number of filter channels. Notably, SMOF is friendly to standard hardware devices without any customized low-level implementations, and the pruning effort by kernel size reduction does not suffer from the fixed-size width constraint in SIMD units of general-purpose processors. The pruned networks can be deployed effortlessly with significant running time reduction. We also support these claims via extensive experiments on various CNN structures and general-purpose processors for mobile devices.
Article
Full-text available
The Internet of Things (IoT) is widely regarded as a key component of the Internet of the future and thereby has drawn significant interests in recent years. IoT consists of billions of intelligent and communicating “things”, which further extend borders of the world with physical and virtual entities. Such ubiquitous smart things produce massive data every day, posing urgent demands on quick data analysis on various smart mobile devices. Fortunately, the recent breakthroughs in deep learning have enabled us to address the problem in an elegant way. Deep models can be exported to process massive sensor data and learn underlying features quickly and efficiently for various IoT applications on smart devices. In this article, we survey the literature on leveraging deep learning to various IoT applications. We aim to give insights on how deep learning tools can be applied from diverse perspectives to empower IoT applications in four representative domains, including smart healthcare, smart home, smart transportation, and smart industry. A main thrust is to seamlessly merge the two disciplines of deep learning and IoT, resulting in a wide-range of new designs in IoT applications, such as health monitoring, disease analysis, indoor localization, intelligent control, home robotics, traffic prediction, traffic monitoring, autonomous driving, and manufacture inspection. We also discuss a set of issues, challenges, and future research directions that leverage deep learning to empower IoT applications, which may motivate and inspire further developments in this promising field.
Article
Full-text available
Nowadays video content has contributed to the majority of Internet traffic, which brings great challenge to the network infrastructure. Fortunately, the emergence of edge computing has provided a promising way to reduce the video load on the network by caching contents closer to users.But caching replacement algorithm is essential for the cache efficiency considering the limited cache space under existing edge-assisted network architecture. To investigate the challenges and opportunities inside, we first measure the performance of five state-of-the-art caching algorithms based on three real-world datasets. Our observation shows that state-of-the-art caching replacement algorithms suffer from following weaknesses: 1) the rule-based replacement approachs (e.g., LFU,LRU) cannot adapt under different scenarios; 2) data-driven forecast approaches only work efficiently on specific scenarios or datasets, as the extracted features working on one dataset may not work on another one. Motivated by these observations and edge-assisted computation capacity, we then propose an edge-assisted intelligent caching replacement framework LSTM-C based on deep Long Short-Term Memory network, which contains two types of modules: 1) four basic modules manage the coordination among content requests, content replace, cache space, service management; 2) three learning-based modules enable the online deep learning to provide intelligent caching strategy. Supported by this design, LSTM-C learns the pattern of content popularity at long and short time scales as well as determines the cache replacement policy. Most important, LSTM-C represents the request pattern with built-in memory cells, thus requires no data pre-processing, pre-programmed model or additional information. Our experiment results show that LSTM-C outperforms state-of-the-art methods in cache hit rate on three real-traces of video requests. When the cache size is limited, LSTM-C outperforms baselines by 20%~32% in cache hit rate. We also show that the training and predicting time of one iteration are $8.6~ms$ and $300~\mu s$ on average respectively, which are fast enough for online operations.
Article
Full-text available
Traffic prediction has drawn increasing attention in AI research field due to the increasing availability of large-scale traffic data and its importance in the real world. For example, an accurate taxi demand prediction can assist taxi companies in pre-allocating taxis. The key challenge of traffic prediction lies in how to model the complex spatial dependencies and temporal dynamics. Although both factors have been considered in modeling, existing works make strong assumptions about spatial dependence and temporal dynamics, i.e., spatial dependence is stationary in time, and temporal dynamics is strictly periodical. However, in practice the spatial dependence could be dynamic (i.e., changing from time to time), and the temporal dynamics could have some perturbation from one period to another period. In this paper, we make two important observations: (1) the spatial dependencies between locations are dynamic; and (2) the temporal dependency follows daily and weekly pattern but it is not strictly periodic for its dynamic temporal shifting. To address these two issues, we propose a novel Spatial-Temporal Dynamic Network (STDN), in which a flow gating mechanism is introduced to learn the dynamic similarity between locations, and a periodically shifted attention mechanism is designed to handle long-term periodic temporal shifting. To the best of our knowledge, this is the first work that tackle both issues in a unified framework. Our experimental results on real-world traffic datasets verify the effectiveness of the proposed method.
Article
Full-text available
Deep learning is currently widely used in a variety of applications, including computer vision and natural language processing. End devices, such as smartphones and Internet-of-Things sensors, are generating data that need to be analyzed in real time using deep learning or used to train deep learning models. However, deep learning inference and training require substantial computation resources to run quickly. Edge computing, where a fine mesh of compute nodes are placed close to end devices, is a viable way to meet the high computation and low-latency requirements of deep learning on edge devices and also provides additional benefits in terms of privacy, bandwidth efficiency, and scalability. This paper aims to provide a comprehensive review of the current state of the art at the intersection of deep learning and edge computing. Specifically, it will provide an overview of applications where deep learning is used at the network edge, discuss various approaches for quickly executing deep learning inference across a combination of end devices, edge servers, and the cloud, and describe the methods for training deep learning models across multiple edge devices. It will also discuss open challenges in terms of systems performance, network technologies and management, benchmarks, and privacy. The reader will take away the following concepts from this paper: understanding scenarios where deep learning at the network edge can be useful, understanding common techniques for speeding up deep learning inference and performing distributed training on edge devices, and understanding recent trends and opportunities.
Article
In recent years, mobile devices are equipped with increasingly advanced sensing and computing capabilities. Coupled with advancements in Deep Learning (DL), this opens up countless possibilities for meaningful applications, e.g., for medical purposes and in vehicular networks. Traditional cloud-based Machine Learning (ML) approaches require the data to be centralized in a cloud server or data center. However, this results in critical issues related to unacceptable latency and communication inefficiency. To this end, Mobile Edge Computing (MEC) has been proposed to bring intelligence closer to the edge, where data is produced. However, conventional enabling technologies for ML at mobile edge networks still require personal data to be shared with external parties, e.g., edge servers. Recently, in light of increasingly stringent data privacy legislations and growing privacy concerns, the concept of Federated Learning (FL) has been introduced. In FL, end devices use their local data to train an ML model required by the server. The end devices then send the model updates rather than raw data to the server for aggregation. FL can serve as an enabling technology in mobile edge networks since it enables the collaborative training of an ML model and also enables DL for mobile edge network optimization. However, in a large-scale and complex mobile edge network, heterogeneous devices with varying constraints are involved. This raises challenges of communication costs, resource allocation, and privacy and security in the implementation of FL at scale. In this survey, we begin with an introduction to the background and fundamentals of FL. Then, we highlight the aforementioned challenges of FL implementation and review existing solutions. Furthermore, we present the applications of FL for mobile edge network optimization. Finally, we discuss the important challenges and future research directions in FL.
Article
Ubiquitous sensors and smart devices from factories and communities are generating massive amounts of data, and ever-increasing computing power is driving the core of computation and services from the cloud to the edge of the network. As an important enabler broadly changing people’s lives, from face recognition to ambitious smart factories and cities, developments of artificial intelligence (especially deep learning, DL) based applications and services are thriving. However, due to efficiency and latency issues, the current cloud computing service architecture hinders the vision of “providing artificial intelligence for every person and every organization at everywhere”. Thus, unleashing DL services using resources at the network edge near the data sources has emerged as a desirable solution. Therefore, edge intelligence, aiming to facilitate the deployment of DL services by edge computing, has received significant attention. In addition, DL, as the representative technique of artificial intelligence, can be integrated into edge computing frameworks to build intelligent edge for dynamic, adaptive edge maintenance and management. With regard to mutually beneficial edge intelligence and intelligent edge, this paper introduces and discusses: 1) the application scenarios of both; 2) the practical implementation methods and enabling technologies, namely DL training and inference in the customized edge computing framework; 3) challenges and future trends of more pervasive and fine-grained intelligence. We believe that by consolidating information scattered across the communication, networking, and DL areas, this survey can help readers to understand the connections between enabling technologies while promoting further discussions on the fusion of edge intelligence and intelligent edge, i.e., Edge DL.
Article
It is well-known that the HVAC (heating, ventilation and air conditioning) dominates electricity consumption in commercial buildings. In this paper, we focus on one of the core problems in building operation, namely chiller sequencing to reduce HVAC electricity consumption. Our contributions are threefold. First, we make a case for why it is important to quantify the performance profile of a chiller, namely coefficient of performance (COP), at run-time, by developing a data-driven COP estimation methodology. Second, we show that predicting COP accurately is a non-trivial problem, requiring considerable computation time. To overcome this barrier, we develop a data-driven COP prediction model and an edge-based chiller sequencing framework integrating the COP predictions, and show that they strike a good balance between electricity saving and ease of use for real-world deployment. Finally, we evaluate the performance of our scheme by applying it to real-world data, spanning 4 years, obtained from multiple chillers across 3 large commercial buildings in Hong Kong. The results show an electricity saving of over 30% compared to baselines. We offer our edge based data-driven chiller sequencing framework as a cost-effective and practical mechanism to reduce electricity consumption associated with HVAC operation in commercial buildings.
Article
The emerging of machine learning has massively promoted the abilities of computational sustainability in natural resource management and allocation. Many Internet giants such as Google, Amazon, and Microsoft now provide Machine Learning as a Service (MLaaS) to meet the increasing demand for machine learning services. However, the prediction results of training data and testing data with the same machine learning model in MLaaS have remarkable differences, and thus the attackers can leverage machine learning technique to launch the so-called membership inference attacks, i.e., to infer whether a record is in the training data or not. In this paper, we propose MIASec that can guarantee the data indistinguishability of the training data and thereby has the ability to defend against membership inference attacks in MLaaS. The key idea of MIASec is to narrow the dynamic ranges of vital features in the training data, such that the training data, the testing data, and even the synthetic data have almost semblable prediction results by the same machine learning model. With elaborated design on modifying the values of vital features in the training data, MIASec can thus reduce the differences between the model's outcomes of training data and testing data, thereby protecting the training data in effect while keeping the model's accuracy stable. We empirically evaluate MIASec on machine learning models trained by off-line neural networks and on-line MLaaS. Using realistic data and classification tasks, our experiment results show that MIASec can defend the membership inference attacks effectively. In particular, MIASec can reduce the precision and recall of attacks respectively by 11:7% and 15:4% in average, and by 18:6% and 21:8% at best.