Content uploaded by Hong Wen
Author content
All content in this area was uploaded by Hong Wen on Apr 11, 2022
Content may be subject to copyright.
1374 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
On Physical-Layer Authentication via Online
Transfer Learning
Yi Chen ,Student Member, IEEE, Pin-Han Ho ,Fellow, IEEE, Hong Wen ,Senior Member, IEEE,
Shih Yu Chang ,Senior Member, IEEE, and Shahriar Real
Abstract—This article introduces a novel physical-layer (PHY-
layer) authentication scheme, called transfer learning-based
PHY-layer authentication (TL-PHA), aiming to achieve fast
online user authentication that is highly desired for latency-
sensitive applications such as edge computing. The proposed
TL-PHA scheme is characterized by incorporating with a novel
convolutional neural network architecture, namely, the triple-
pool network (TP-Net), for achieving lightweight and online
classification, as well as effective data augmentation methods for
generation of data set samples for the network model training. To
assess the performance of the proposed scheme, we conducted two
sets of experiments, including the one using computer-simulated
channel data and the other utilizing real experiment data gen-
erated by our wireless testbed. The results demonstrate the
superiority of the proposed scheme in terms of authentication
accuracy, detection rate, and training complexity compared to
all the considered counterparts.
Index Terms—Channel state information (CSI), convolutional
neural network (CNN), edge computing (EC), physical-layer
(PHY-layer) authentication, transfer learning (TL).
I. INTRODUCTION
EDGE computing (EC) is an emerging technology aim-
ing to extend the legacy cloud computing services to
the network edge and resolve some of its inherent limitations,
such as real-time control incompetence, network traffic bottle-
necks, and cloud data privacy insecurity [1]–[3]. An EC system
is generally formed by a set of edge devices and Internet
of Things (IoT) terminal units, targeting at achieving a low-
latency and high-reliability service provisioning platform for
supporting heterogeneous EC applications, such as smart grid,
Manuscript received March 22, 2021; revised May 12, 2021; accepted
May 26, 2021. Date of publication June 4, 2021; date of current version
January 7, 2022. This work was supported in part by China Scholarship
Council and in part by the National Major Research and Development
Program, China, under Grant 2018YFB0904900 and Grant 2018YFB0904905.
This article was presented in part at the 2020 IEEE Globecom Workshops
(GC Wkshps), Taipei, Taiwan, December 2020. (Corresponding author:
Hong Wen.)
Yi Chen is with the National Key Laboratory of Science and Technology
on Communications, University of Electronic Science and Technology of
China, Chengdu 611731, China, and also with the Department of Electrical
and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1,
Canada (e-mail: chenyi1309@126.com).
Pin-Han Ho and Shahriar Real are with the Department of Electrical and
Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1,
Canada (e-mail: p4ho@uwaterloo.ca; shahriar.real@uwaterloo.ca).
Hong Wen is with the School of Aeronautics and Astronautics, University of
Electronic Science and Technology of China, Chengdu 611731, China (e-mail:
sunlike@uestc.edu.cn).
Shih Yu Chang is with the Department of Applied Data Science, San Jose
State University, San Jose, CA 95192 USA (e-mail: shihyu.chang@sjsu.edu).
Digital Object Identifier 10.1109/JIOT.2021.3086581
smart home, smart healthcare, and vehicle networks [3]–[5].
Some edge devices of high capacity may serve as small/micro
data centers that are equipped with the intelligence to ensure
Quality of Service (QoS), while the IoT terminals are gener-
ally with limited resources in terms of computational power
and memory that could be subject to various security threats
in the presence of the openness of wireless media.
Physical-layer (PHY-layer) authentication has been consid-
ered an effective alternative to the conventional key-based
approaches by exploring the rich PHY-layer channel char-
acteristics between the EC and client device (CD) [2].
Numerous researches on PHY-layer authentication have been
reported [6]–[10], mostly focusing on the manipulation of
novel statistical models/algorithms of binary hypothesis tests,
whose accuracy highly relies on the test thresholds.
Alternatively, machine learning (ML)-based PHY-layer
authentication has been studied, which aims to achieve
high accuracy of message authentication via a well-trained
ML model [11]–[14]. Although promising, most previously
reported ML-based PHY-layer authentication schemes have
employed ML models of a large number of layers and parame-
ters, leading to significant complexity and power consumption.
The size of the ML model serves as a key to the success of
the considered application scenario, where a fast and adaptive
online classification process is essential.
As a remedy, there have been some advanced convolutional
neural network (CNN) architecture and its activation functions,
such as the global-connected net (GC-Net) [15], reported to
achieve much improved efficiency in both training and classifi-
cation stages. Furthermore, sufficient data sets are essential to
obtain a well-trained model, which are nonetheless not always
available in the highly dynamic EC environment. As such,
efficient data augmentation schemes that encompass a suite of
techniques, such as geometric transformations, kernel filters,
random erasing, feature space augmentation, and adversarial
training, are highly desired [16]–[19]. With more training, data
sets definitely lead to longer training time. In order to fit into
the real-time scenario, we consider transfer learning (TL) for
the proposed EC system, where the EC is allowed to train the
model in off-line and the training result can be migrated to
the network model in real-time. This article introduces a novel
TL-based PHY-layer scheme, called, TL-PHA, for lightweight
user/message authentication. It is uniquely featured by a num-
ber of novel designs, aiming to resolve all the above-mentioned
issues, which are summarized as follows.
2327-4662 c
2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1375
1) We claim the proposed TL-PHA is the first PHY-
layer authentication scheme for the EC applications that
employs the TL strategy. Distinguished from the existing
TL strategies, the proposed TL-PHA scheme can deter-
mine the moment that online fine tuning for the network
model is needed according to the offline training results.
2) To enable a fast training and classification process,
we introduce a novel CNN interconnection architec-
ture, called TP-Net, to facilitate high-efficiency model
training and online classification.
3) To make up the training data sets insufficiency, two data
augmentation methods are investigated and incorporated
in the proposed TL-PHA scheme.
4) Extensive simulation is conducted to verify the proposed
TL-PHA scheme via both computer simulation and
USRP testbed, respectively.
The remainder of this article is organized as follows.
Section II reviews the related studies of PHY-layer authentica-
tion, TL, data augmentation, and CNN. Section III illustrates
the system model of the study. Section IV presents the
proposed TL-PHA scheme. Section V shows the introduction
of the triple pool network (TP-Net). Section VI introduces
the data augmentation methods employed in the proposed
TL-PHA. Section VII shows the experiment results and com-
parison with the counterparts. This article is concluded in
Section VIII.
II. RELATED WORK
A. PYH-Layer Authentication
A number of PHY-layer authentication schemes were
reported in [6]–[12]. The research approach of the binary
hypothesis test upon a given threshold is relatively mature.
A representative of this research direction is given in [6]–[8],
where PHY-layer authentication is performed via exploit-
ing the spatial variability of the radio channel between
unknown channel state information (CSI) and known legal
CSI. Xiao et al. [6] proposed a generalized likelihood ratio
test version for PHY-layer authentication. Wen et al. [7]
introduced two cross-layer authentication schemes for the
smart meter system, namely, the symmetric cryptography-
based PHY-layer-assisted authentication scheme and the
public-key infrastructure-based PHY-layer-assisted authentica-
tion scheme. Pan et al. [8] applied PHY-layer authentication
based on CSI to measurements from indoor, outdoor, moving,
and stationary industrial wireless communication scenarios,
respectively. They derived some meaningful insights on the
applicability of the PHY-layer method to industrial wireless
communications via the CSI analysis that still depended on the
threshold. Nevertheless, the authentication accuracy of these
approaches is highly rely on the test threshold values that are
hard to obtain in practical environment.
Adaptive classification approaches based on ML are gaining
popularity as new strategies for the PHY-layer authentication
scenarios. Pan et al. [9] proposed a threshold-free PHY-layer
authentication scheme based on ML for the industrial mobile
scenario, which can replace the traditional threshold-based
decision-making method. An adaptive PHY-layer authentica-
tion scheme based on ML as an intelligent process to learn
and use the complex time-varying environment was proposed
in [10]. A spoofing detection scheme based on reinforcement
learning process was introduced in [11]. Wang et al. [12] uti-
lized a deep neural network to complete the indoor location via
CSI. Liao et al. [2] proposed a data-augmented multiuser PHY-
layer authentication scheme based on the deep neural network.
The above schemes, although being claimed effective in the
considered scenarios, used conventional ML models with a
large number of layers and parameters. For an ML model to
be applied under latency-sensitive applications such as EC, the
efficiency of such an ML model cannot be just measured by
ML metric, we should also consider the required resources to
perform model training and prediction simultaneously.
B. Transfer Learning
TL has been reported to resolve the situation by using what
is learned for one problem to assist another different but related
problem. It also has been widely applied to the image pro-
cessing scenario [20]–[22]. The main two ways of doing TL
in deep neural networks are the fine-tuning method and the
freezing layer method, respectively. It requires to retrain the
whole network parameters in the fine-tuning method. Instead,
the freezing-layer method freezes most of the transferred
parameters [23].
The conventional ML algorithms need to be trained from
scratch every time to solve specific tasks. However, train-
ing a neural network from scratch may be cumbersome and
the available data sets may not be rich enough to effectively
capture model features. Therefore, the trained neural model
could not be generalized properly when applied in the practical
contexts.
Even if there are enough data sets to train the network online
in real time, it is also time consuming to train the model by
the server, which is not proper for latency-sensitive applica-
tions. Therefore, in this article, we exploit data sets to train the
ML model offline, and utilize the trained model parameters to
perform a fast and accurate online PHY-layer authentication.
This is our core idea to apply TL in PHY-layer authentication.
C. Data Augmentation
Data augmentation is the technique of increasing the size
of data set used for training a deep learning model [16]–
[19]. For desired predictions, the ML models often require
sufficient training data set, which is not always available.
Therefore, the existing data set is augmented in order to
make a better generalized model. Frid-Adar et al. [19] utilized
GAN-based image synthesis data augmentation to improve
classification performance for liver lesion classification in
2018. Li et al. [17] exploited data augmentation approaches
to enhance the authentication accuracy on smartphones. The
audio data augmentation for overcoming the problem of data
scarcity on environmental sound classification was introduced
in [18]. However, these conventional data augmentation meth-
ods mainly include flip, rotation, scale, kernel filters, random
erasing, and so on, which are not proper to increase the
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
1376 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
diversity of PHY-layer CSI, because they will destroy the
correlation structure between the channel matrices.
D. Convolutional Neural Network
CNN has been well recognized as a powerful tool for
intelligent decision making in the presence of complex data.
Krizhevsky et al. [24] first applied a CNN to achieve great
success on ImageNet competition in 2012. Simonyan and
Zisserman [25] found the deep convolutional network (i.e.,
VGG) has a significant effect on the increase in accuracy
for image recognition tasks. A residual-network architec-
ture (ResNet) was proposed to alleviate the training time
of networks in [26]. A deep CNN architecture code named
Inception [27] was proposed for the reduction of computa-
tional burden of the network architecture. Nonetheless, most
these reported CNN architectures still require numerous lay-
ers and a huge number of parameters to achieve the desired
accuracy, and are thus subject to extremely high computation
complexity.
GC-Net [15] is a recently introduced CNN architecture that
is reported to achieve similar performance to that by its pre-
decessors while taking much less parameters and, thus, much
less computation resources. The unique features of GC-Net
include a globally connected interconnection architecture and
a piecewise-linear activation function between the convolution
layers that can successfully mitigate the gradient-vanishing
problem. However, it still has room for improvement, espe-
cially in the application of PHY-layer authentication.
III. SYSTEM MODEL
We consider the scenario of latency-sensitive EC systems,
where an edge computing node (ECN) is associated with
multiple CDs. Authentication is needed when any message
is delivered between the ECN and each CD to avoid any
malicious attempt to the EC system.
The key element of PHY-layer authentication is the CSI (i.e.,
channel response matrix H) obtained via a channel estimation
procedure [28]–[30]. The signal of the receiver is given by
yp(t)=hp×xp(t)+n(t)(1)
where ypis the received signal at the ECN, tis the time interval
between every data frames, hprefers to a time-domain channel
matrix containing the channel coefficients, xprepresents the
pilot signal known to both CDs and ECN used to estimate
the CSI, and n(t)is the additive white Gaussian noise with
variance σ2. The channel time-domain response estimated by
the ECN through the channel estimation is as follows:
h(t)=yp(t)×x−1
p(t)
=hp×xp(t)×x−1
p(t)+n(t)×x−1
p(t)
=hp+n(t)×x−1
p(t)(2)
where x−1
p(t)is the inversion of xp(t). The ECN is able to
get the channel frequency response matrix, i.e., CSI H,bythe
discrete Fourier transform (DFT) of the time-domain signal.
For the sake of simplicity, the channel frequency response
matrix, i.e., CSI H, is denoted as
H=YpX−1
p=⎡
⎢
⎢
⎢
⎣
z11 z12 ··· z1n
z21 z22 ··· z2n
.
.
..
.
.....
.
.
zm1zm2··· zmn
⎤
⎥
⎥
⎥
⎦
(3)
where Xpis the known pilot and Ypis the pilot obtained
at the ECN. By assuming orthogonal frequency-division
multiplexing (OFDM), where H∈Cm×n,zuv =auv +jbuv,
a,b∈R,u=1,2,...,m,v=1,2,...,n,j
2=−1, and
m=Ns×Nt, with Nsbeing the number of subcarriers and Nt
being the number of transmitting antennas, whereas nis the
number of receiving antennas.
The ECN first validates the received signal Ypvia the con-
ventional cryptography-based approach implemented by the
upper layer, and goes through the CSI extraction. Once pre-
pared, the PHY-layer authentication can be performed on
the incoming messages. The detailed process of upper level
authentication is given as follows.
1) In the beginning, the timer is set to 0, i.e., t=0. The flag
bit representing the completion of PHY-layer authenti-
cation preparation is also set to 0, i.e., FPHY =0, where
FPHY is set to 1 when the PHY-layer authentication is
ready to perform. In addition, the counter of the data
frame is set to 0, i.e., numdata =0.
2) Upon receiving a message from a CD, the ECN first
checks the flag bit of FPHY, where the proposed PHY-
layer authentication is performed if FPHY =1, and the
upper level numerical authentication is relegated other-
wise. In the former, the ECN has to extract the CSI Ht
k,
the channel response vector of the kth node at time t.
3) If the upper level authentication is not successful, the
message is considered to be illegal and discarded; oth-
erwise, the ECN considers the message legal, measures,
and records the CSI Ht
k.
The collected channel response matrices shall be used to
train the TP-Net model offline. The process of upper level
authentication is summarized in Algorithm 1.
IV. PROPOSED TL-PHA
The proposed PHY-layer authentication scheme incorporates
with three functional modules as shown in Fig. 1: 1) offline
training; 2) online migration and fine tuning; and 3) online
decision making. Before offline training, ECN needs to col-
lect CSIs, whose legitimacy is determined via upper level
authentication (i.e., running Algorithm 1).
A. Offline Training
In the offline training phase, the ECN first uses the channel
frequency response matrices Hto generate the channel train-
ing vectors specific to each CD, which are, in turn, used for
training the TP-Net model. The details are as follows.
1) The ECN first adjusts the complex matrix Hto get a
real matrix Uwith a size of m×2n
U=[P,Q]=[(H), (H)](4)
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1377
Algorithm 1 Upper Level Authentication
Input: Wireless signal carrying information sent from CD.
Output: Channel frequency response matrix, i.e., CSI H.
1: Initialization parameters.
2: Upon receiving a message from a CD, ECN checks the
PHY-layer authentication flag bit:
3: if FPHY =1then
4: ECN extracts the CSI Ht
kand performs the PHY-layer
authentication with the well trained TP-Net model, i.e.,
running Algorithm 4.
5: else if FPHY =0then
6: ECN activates the upper level authentication.
7: if The authentication is unsuccessful then
8: The message is illegal and discarded.
9: else
10: ECN considers the message legal, measures and
records the CSI Ht
k.
11: end if
12: end if
13: Output channel frequency response matrices.
Fig. 1. Flowchart of TL-PHA.
where (·)represents the real part, (·)is the imaginary
part, P=(H),Q=(H), so the formula of Uis as
follows:
U=⎡
⎢
⎢
⎢
⎣
a11 ··· a1nb11 ··· b1n
a21 ··· a2nb21 ··· b2n
.
.
.....
.
..
.
.....
.
.
am1··· amn bm1··· bmn
⎤
⎥
⎥
⎥
⎦
.(5)
The superscript and subscript of Ut
kare specific to the
kth node at time tof the channel response, respectively.
Algorithm 2 Offline Training for PHY-Layer Authentication
Input: Channel frequency response matrices H.
Output: Authentication model.
1: ECN adjusts the complex channel matrix Ht
kto get a real
channel matrix Ut
k.
2: ECN gives the new real channel matrix Ut
ka label Ik.
3: ECN generates training set Dtrain.
4: ECN employs Dtrain to train the TP-Net model.
5: Output authentication model.
2) A label Ikis given to the newly obtained real channel
matrix Ut
k, which is a unit vector represented by an one-
hot code.
3) The training set Dtrain is generated as follows:
Dtrain ={Xtrain,Ytrain}(6)
where
Xtrain =⎡
⎢
⎢
⎢
⎣
Ut1
1
Ut2
2
.
.
.
Utk
k
⎤
⎥
⎥
⎥
⎦
=⎡
⎢
⎢
⎢
⎢
⎣
U1
1U2
1··· Uψ1
1
U1
2U2
2··· Uψ2
2
.
.
..
.
.....
.
.
U1
kU2
k··· Uψk
k
⎤
⎥
⎥
⎥
⎥
⎦
(7)
Ytrain =⎡
⎢
⎢
⎢
⎣
It1
1
It2
2
.
.
.
Itk
k
⎤
⎥
⎥
⎥
⎦
=⎡
⎢
⎢
⎢
⎢
⎣
I1
1I2
1··· Iψ1
1
I1
2I2
2··· Iψ2
2
.
.
..
.
.....
.
.
I1
kI2
k··· Iψk
k
⎤
⎥
⎥
⎥
⎥
⎦
.(8)
4) The training set Dtrain is used to train the TP-Net model
that will be used for the future online processes.
The offline training for PHY-layer authentication is summa-
rized in Algorithm 2.
B. Online Migration and Fine Tuning
The ECN periodically checks if the number of CDs is
changed. If not, the ECN does not need to update the trained
model parameters. Otherwise, the ECN uses some of the most
updated CSI to fine tune the TP-Net model based on the offline
training result. The detailed process of the online migration
and fine tuning is given as follows.
1) ECN first obtains the model parameters from the offline-
trained TP-Net model and resets the flag bit FPHY =0.
2) ECN online receives messages and runs Algorithm 1 for
collecting CSIs. If ECN collects δ(e.g., δ≥3) CSIs per
CD, it turns to the step of checking the number of CDs;
otherwise, waits for the next new message.
3) If Numonline =Numoffline, ECN directly adopts the
offline trained model as its online PHY-layer authentica-
tion model and sets FPHY =1, representing the readiness
of PHY-layer authentication for the subsequent incoming
messages.
4) Otherwise, the ECN uses the collected CSIs to
fine tune the model parameters in the last layer
in the offline trained model while freezing all the
others.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
1378 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
Algorithm 3 Online Migration and Fine Tuning for PHY-
Layer Authentication
Input: The offline well trained TP-Net model.
Output: Online authentication model.
1: ECN obtains model parameters from the offline trained
TP-Net model.
2: ECN online receives message and performs Algorithm 1.
3: ECN checks the number of CSIs:
4: if ECN collects δ(e.g., δ3) CSIs per CD. then
5: ECN checks the number of CDs:
6: if Numonline =Num
offline then
7: Adopt the offline trained model as its online PHY-
layer authentication model and set FPHY =1.
8: else
9: ECN utilizes the collected CSIs to fine-tune the
model of frozen parameters.
10: ECN takes the new trained model as its online
PHY-layer authentication model, activates the online
decision making and sets FPHY =1.
11: end if
12: else
13: ECN waits for next new message.
14: if The number of CDs is changed. then
15: ECN sets FPHY =0.
16: end if
17: end if
18: Output online authentication model.
5) When the fine tuning is completed, the ECN activates
the online decision making and sets FPHY =1, where
the desired PHY-layer authentication can be performed
on the subsequent messages.
The proposed online migration and fine-tuning mechanism
is summarized in Algorithm 3.
C. Online Decision Making
With FPHY =1, the proposed PHY-layer authentication can
start and is described as follows.
1) ECN starts to authenticate a new message via the most
updated model. ECN first extracts the channel response
matrix Ht
unknown via performing Algorithm 1.
2) The ECN adjusts the channel response matrix Ht
unknown
as the real number matrix Ut
unknown like formula (5).
3) The real number channel matrix is fed into the online
fine-tuned TP-Net for authentication, where the input of
TP-Net is marked as XAuc
XAuc =U1
unknown,U2
unknown,U3
unknown,...,
.(9)
4) Finally, the ECN makes the authentication decision
based on the output of XAuc of the TP-Net, where the
output is marked as ˆ
YAuc, but the real label of the output
is YAuc.
The online decision making for PHY-layer authentication is
summarized in Algorithm 4.
Algorithm 4 Online Decision Making for PHY-Layer
Authentication
Input: Channel matrices of unknown nodes.
Output: Labels of unknown nodes.
1: for ECN receives messages sent from unknown CD do
2: ECN checks the PHY-layer authentication flag bit:
3: if FPHY =1then
4: ECN estimates the CSI Ht
unknown;
5: ECN adjusts Ht
unknown as Ut
unknown like (5).
6: The real number channel matrix of unknown node
XAuc is fed into the TP-Net for authentication.
7: Output the label of unknown node.
8: else if FPHY =0then
9: Perform Algorithm 1.
10: end if
11: end for
Fig. 2. Architecture of TP-Net.
V. INTRODUCTION OF TP-NET
Inspired by the GC-Net [15], we propose a new CNN archi-
tecture, namely, TP-Net, whose interconnection architecture is
shown in Fig. 2. With nconvolutional blocks in total, each
block has two filters, followed by batch normalization and
activation, where any convolutional block is allowed to con-
nect with a hidden layer that feeds into the last hidden layer
and output (SoftMax) layer.
The proposed TP-Net employs an exponential linear unit
(ELU) and a standard unit softmax function (SoftMax) as its
activation functions, where the ELU is defined as a nonlin-
ear function as presented in formula (10), and the SoftMax
is defined by the formula (11), where i=1,2,...,N, and
w=(w1,w2,...,wN)∈RN
f(x)=x,if x≥0
α(ex−1),if x<0(10)
SoftMax(w)i=ewi
N
j=1ew
j
.(11)
Distinguished from any previously reported CNN architec-
ture, the proposed TP-Net is uniquely featured by consisting
of three types of pooling: 1) max; 2) average; and 3) global.
The max pooling is applied between two convolutional blocks
to reduce the feature map sizes. The average pooling aims
to tune the variance in the data set for possibly dimensional
reduction. The global average pooling (GAP) decreases the
dimension rapidly to possibly save computation cost. The main
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1379
reason behind using the triple pools is that these three types of
methods can utilize their strength parts to mitigate their weak-
ness parts mutually. Thus, TP-Net model can work on the CSI
database.
VI. PROPOSED DATA AUGMENTATION
In view of the application scenarios of edge servers with
different resources, two novel data augmentation methods,
namely, stochastic weight data augmentation (SWDA) and
block swap data augmentation (BSDA), are introduced for the
exploration of some unique features of the PHY-layer CSI,
such as its time continuity and correlation between subcarriers.
A. SWDA
To explore the correlation among the CSIs of consecu-
tive frames, SWDA generates new channel matrices using
the channel response vectors of two or more consecutive
frames. Stochastic weight averaging data augmentation (SWA-
DA) [2] is a recently proposed data augmentation method for
multiuser PHY-layer authentication. Different from SWA-DA
that requires all the original data sets (CSIs) to be within a
common coherence time [31], the proposed SWDA has adja-
cent CSIs to be within common coherence time, which not
only leads to easier implementation, but can also sufficiently
exploit the correlation of CSIs in adjacent frames. The detailed
SWDA is given as follows.
Let Dkdenote the original data set sample of the kth CD
Dk={Xk,Yk}(12)
where Xkis the original input samples
Xk=Ut
k=U1
k,U2
k,...,Uψk
k(13)
and Ykis the original output samples
Yk=It
k=I1
k,I2
k,...,Iψk
k.(14)
The new channel response matrix is constructed according to
the following formula:
Sμ
k=
μ+θ−1
i=μλi−μ+1×Ui
k,1<θ≤ψk(15)
where θ, which is a positive integer and 1 <θ≤ψk, rep-
resents the number of original adjacent samples involved in
generating a new sample, ψkis the total number of orig-
inal samples of the kth node, λis stochastic weight and
θ
i=1λi=1, and μis the index of new samples, natural
number and μ=1,2,...,ψ
k−θ+1. After reconstructing
the channel response matrix, the new input sample is
XSWDA_k=Ut
k,Sμ
k
=U1
k,U2
k,...,Uψk
k,S1
k,S2
k,...,Sμk
k.(16)
The new output sample is
YSWDA_k=I1
k,I2
k,...,Iψk
k,I1
k,I2
k,...,Iμk
k(17)
where μk=ψk−θ+1.
Algorithm 5 PHY-Layer Authentication With SWDA
Input: The channel matrices of unknown nodes XAuc.
Output: The labels of unkonwn nodes ˆ
YAuc.
1: Get the original CSIs samples Dtrain ={Xtrain,Ytrain}
and θ.
2: ECN reconstructs the new training data set DSWDA_train
according to equation (15), (16) and (17).
3: ECN uses the new training data set DSWDA_train by exe-
cuting the functional module of offline training as given
in Section IV-A.
4: Obtain XAuc.
5: XAuc is fed into the authentication model.
6: Output the labels of unkonwn nodes ˆ
YAuc.
The new training data set of the kth CD can be obtained as
follows:
DSWDA_k=XSWDA_k,YSWDA_k.(18)
The PHY-layer authentication with SWDA is summarized
in Algorithm 5.
B. BSDA
The conventional data augmentation techniques, such as flip,
rotation, and scale [16]–[18], may cause the receiver unable
to correctly demodulate/decode according to the new channel
matrix since the correlation structure of the channel matrix is
destroyed. Motivated by the observation, BSDA is designed to
generate new channel vectors with the prior and post adjacent
CSIs, where some elements of two or more channel response
vectors are swapped to obtain new channel response matri-
ces. Compared to the conventional data augmentation, the
BSDA scheme does not change the correlation between chan-
nel matrices, but only exchanges elements between the same
position of the channel matrices, which can generate more
new data samples that facilitate boosting the authentication
rate. The detailed BSDA is shown as follows.
Let Dkdenote the original training sample of the kth CD,
and Ut
kwithasizeofm×2ndenote the original real number
channel matrix of kth CD at ttime slot and t=1,2,...,ψ
k.
First, Ut
kis divided into two blocks equally according to
row vectors
Ut
k=⎡
⎢
⎣c1,c2,...,cm
2
1
,cm
2+1,...,cm
2
⎤
⎥
⎦
T
=r(t)(1)
k
r(t)(2)
k(19)
where r(t)(1)
kwith a size of (m/2)×2nrepresents the upper
half block of matrix Ut
kand r(t)(2)
kwith size of (m/2)×2nis
the lower half block of matrix Ut
k.
Second, the partial block elements of each pair of channel
matrices are swapped in adjacent η(2 ≤η≤ψk) channel
matrices according to the following formula:
⎧
⎨
⎩
r(t)(1)
k↔r(t+i−1)(1)
k
or
r(t)(2)
k↔r(t+i−1)(2)
k
(20)
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
1380 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
Algorithm 6 PHY-Layer Authentication With BSDA
Input: The channel matrices of unknown nodes XAuc.
Output: The labels of unknown nodes ˆ
YAuc.
1: Get the original CSI dataset Dtrain ={Xtrain,Ytrain}
and η.
2: ECN generates new CSI training dataset DBSDA_train
according to equation (19), (20), (21), (22) and (23).
3: ECN uses the new training dataset DBSDA_train by perform-
ing the functional module of offline training as given in
Section IV-A.
4: Obtain XAuc.
5: XAuc is fed into the authentication model.
6: Output the labels of unknown nodes ˆ
YAuc.
where iis just a positive integer index here and i=2,...,η.
Then, the new channel response matrix can be expressed as
ξ
k=1
k,
2
k,...,
(ψk−η+1)×(2η−2)
k
=r(i)(1)
k
r(t)(2)
km×2n
,r(t)(1)
k
r(t+1)(2)
km×2n(21)
where ξis the index of the newly generated sample and ξ=
1,2,...,(ψ
k−η+1)×(2η−2).
After generating the channel response matrix, the new input
sample is
XBSDA_k=Ut
k,ξ
k
=U1
k,U2
k,...,Uψk
k,
1
k,
2
k,...,
(ψk−η+1)×(2η−2)
k.(22)
The new output sample is
YBSDA_k=I1
k,I2
k,...,Iψk
k,
I1
k,I2
k,...,I(ψk−η+1)×(2η−2)
k.(23)
The new training data set of the kth CD can be denoted as
follows:
DBSDA_k=XBSDA_k,YBSDA_k.(24)
The PHY-layer authentication with BSDA is summarized in
Algorithm 6.
VII. EXPERIMENTAL RESULTS
A. Overall Settings
Two sets of experiments are conducted by using the chan-
nels generated by computer simulation and an NI USRP
testbed, respectively. A number of counterparts are consid-
ered for the comparison purpose, including the case without
using TL, the case without data augmentation, and the case
using conventional threshold-based PHY-layer authentication,
respectively. In addition to TP-Net, we also examine a con-
ventional CNN architecture [32], GC-Net [15], and VGG [24].
Furthermore, the threshold-based method in [8] is imple-
mented, too.
The TP-Net is implemented with four convolutions, among
which two are composed of two convolution layers with 1 ×1
filter and 16 and 32 feature maps, respectively; and the other
two are composed of two convolution layers with 3 ×3 filters
and 32 feature maps. The 2×2 max-pooling layer with a 2×2
stride as applied after both of the two 3×3 convolution layers.
GAP is applied to the output of the first convolution layer and
the collected parameters are fed as input to the SoftMax layer
for classification.
The conventional CNN architecture, namely, CNN-2,
employs ReLU as the activation function and contains two
convolutions, where one is composed of 4×4 filter and eight
feature maps, and the other is composed of 2×2 filters and 16
feature maps. The 4×4 average pooling layer with a stride of
4×4 is applied after the 4 ×4 convolution layers. The 2 ×2
average pooling layer with a stride of 2 ×2 is applied after
the 2 ×2 convolution layers.
The GC-Net considered in the experiment employs ELU as
the activation function and is composed of three convolution
layers with small 3×3 filters and 64, 64, and 64 feature maps,
respectively. The 2×2 max-pooling layer with a stride of 2×2
is applied after both of the first two convolution layers. GAP
is applied to the output of each convolution layer and the
collected averaged features are fed as input to a dense layer
with 64 neurons. The output of the dense layer is fed as input
to the SoftMax layer for classification.
The VGG architecture considered in the experiment, called
VGG-7, employs ELU as the activation function and is com-
posed of seven convolution layers with small 3 ×3 filters and
64, 64, 128, 128, 256, 256, and 256 feature maps, respectively.
The 2 ×2 max-pooling layer with a stride of 2 ×2 is applied
after the first two, the first four and the last convolution layer,
respectively. With two fully connected layers, one has 512 neu-
rons followed by an ELU activation function, while the other
has “classes” neurons accompanied with the SoftMax activa-
tion function, where classes is the total number of neurons to
be classified. The adaptive moment estimation (Adam) [33]
accelerated gradient algorithm is used for the acceleration of
all CNN training. To be fair, the parameter αof ELU is set
to 1.
The authentication rate is considered as the performance
metric, denoted as AucRate, which is the ratio of correctly
authenticated samples to the total launched ones
AucRate =1
Ψ
Ψ
Ωˆ
YAuc ◦YAuc(25)
where Ψis the total number of CSI that need to be verified, Ω
is the number of nodes, and ˆ
YAuc ◦YAuc denotes the Hadamard
product of the matrices ˆ
YAuc and YAuc.ˆ
YAuc is the label of
output of CNN, while the real label of the output is YAuc.
Training time is used to measure the computational com-
plexity of the model training, which is described in (26)
Training time =
N
i=1
ti(26)
where tidenotes the training time of the ith epoch, iis the
number of epoch here, and i=1,2,...,N.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1381
In addition, the true-positives rate (TPR) and false-positive
rate (FPR) are two performance matrices defined as follows:
TPR =TP
TP +FN (27)
FPR =FP
FP +TP (28)
where true positive (TP) and false positive (FP) are the num-
ber of malicious CSIs being detected and legitimate CSIs
detected as malicious, respectively, and false negative (FN)
is the number of malicious CSIs launched without being
detected. TP +FN depicts the total number of illegal CSIs,
while FP +TP is the total number of CSIs determined as
illegal.
We also define the detection rate, denoted as DetRate and
given by
DetRate =TP +TN
TP +TN +FP +FN (29)
where true negative (TN) represents the number of legitimate
CSIs being correctly detected.
B. Training With Simulated Channels
The tapped delay line (TDL) model is exploited to simulate
the Rayleigh fading channel [34] with multipath delay, where
a set of unit-power, independent nonfrequency selective fading
generators, such as the filtered white Gaussian noise (FWGN)
or Jakes model [35]–[37], are employed. It is based on the
following equation:
y(n)=
ND−1
d=0
hd(n)x(n−d)(30)
where y(n)and x(n)are the output and input at the nth sample
instant, respectively, and NDdenotes the number of taps of
the channel filters with filter coefficients hd. We use five paths
with different power delays to synthesize the channels of each
node. The time delay of the first four paths is the same, which
is 0 s (s), 2 ×10−6s, 4 ×10−6s, 8 ×10−6s, respectively.
When there are 40 CDs, the time delay of the fifth path is
3×10−6s, 4 ×10−6s,...,4.2×10−5s, respectively.
The least squares (LSs) algorithm is adopted to estimate CSI
under OFDM with a sampling interval tsampling =1×10−6s,
the number of subcarriers nsubcarrier =128, the pilot interval
npilot_interval =3, the cyclic prefix length lcp_length =16, and
the digital modulation method as QPSK. The number of origi-
nal channel samples for each node training and that for testing
CNN are both 100.
Our simulation program runs on a 64-bit Win10 Professional
system with an Intel Core i7-9750H of the main frequency
2.59 GHz and the physical memory 16 GB. The Python Keras
library is used to build the network.
1) Homogeneous Network Scenario: With the homoge-
neous network scenario, the number of nodes for the online
training and offline training is identical and, thus, there is no
need to perform online fine tuning during the processes of
offline training and online migration authentication.
Fig. 3 shows the authentication rates under 20 CDs with
different channel SNR and the number of antennas, and Fig. 4
Fig. 3. Authentication rate under different conditions. (a) 20 nodes, 2 dB,
eight antennas. (b) 20 nodes, 2 dB, 16 antennas. (c) 20 nodes, 8 dB, eight
antennas.
Fig. 4. Authentication rate of each scheme with channel SNR as 2 dB and
eight antennas, where the authentication models of all schemes are the training
results of the 50th epoch.
demonstrates the results of authentication rate under different
numbers of nodes with SNR as 2 dB and eight antennas, where
the proposed scheme is compared to the case of using CNN-2,
GC-Net, and VGG-7. We first find that increasing the number
of antennas and channel SNR can improve the authentication
rate, while the latter is even more dominant. In addition, the
proposed TL-PHA scheme can achieve the best performance,
and such an advantage remains when the number of nodes
increases.
Fig. 5 shows the comparison results among all the ML-based
and the threshold-based schemes, where the former provides
much better performance than the latter, while the performance
of threshold-based schemes highly relies on the channel SNR
and threshold values. It confirms that the proposed TL-PHA
scheme takes all the advantages against its counterparts.
Table I shows the detection performance with channel SNR
as 2 dB and number of antennas as 8, where 50% of the nodes
in the network are malicious. It is clear the proposed scheme
outperforms all the other considered counterparts.
2) Disparate Network Scenario: With the disparate
network scenario, the number of online authentication nodes
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
1382 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
TAB LE I
DETECTION PERFORMANCE UNDER DIFFERENT NODES WITH CHANNEL SNR AS 2DBAND NUMBER OF ANTENNAS AS 8
Fig. 5. Comparison of authentication rate under different numbers of nodes
among all the ML-based schemes and the threshold-based scheme, where the
authentication rate of ML-based schemes is the result of the 50th epoch.
is different from that of the offline training nodes and, thus,
the fine tuning of the online authentication model is necessary.
Without loss of generality, we focus on the case, where the
online authentication nodes are less than that for the offline
training. Specifically, the proposed offline training is con-
ducted on a 20-node CSI data set with 100 CSIs of each node.
The resultant authentication rates are compared to that without
TL under 8-dB channel SNR and a number of eight antennas.
Figs. 6 and 7 demonstrate the authentication rate
performance between online TL and training from the scratch
among all the ML-based schemes for 15 nodes with 5 and
10 training samples per node, respectively, where “TL, no
Freeze” stands for the case where no parameter is frozen, and
only the model trained offline is directly transferred for online
fine tuning; “TL, Freeze 1C” is the case where the parame-
ters of the first convolution layer of the migrated model are
frozen during online fine tuning; “TL, Freeze 2C” is the case
where the parameters of the first two convolution layers of the
migrated model are frozen, and so on; with “TL, Freeze 1-D,”
the parameters of the first “Dense” layer and all the convolu-
tion layers of the migrated model are frozen during the online
fine tuning; and with “no TL,” the model is trained from the
scratch with 5 or 10 samples per node.
Obviously, the authentication rate performance by using
online TL is better than that of the no-TL case. In addition,
the authentication rate curve of Fig. 7 is smoother than that
of Fig. 6 since the number of training data sets is different.
This also shows that with more data sets, the network is better
trained. What is more, freezing the parameters of different lay-
ers affect the results. For the TL of CNN-2, little difference
on authentication rate between the cases of frozen parame-
ters and unfrozen parameters, respectively, is observed. For
the TP-Net, the authentication rate of no freezing parameters
or freezing the parameters of the first three convolution layers
is higher than that of freezing all parameters. For the GC-Net,
the authentication rate performance by freezing the parame-
ters of the first three convolution layers and/or the parameters
of the first “Dense” layer is better than the other cases, while
the authentication rate of using TL is the lowest when the
parameters are not frozen. We can see that the use of migra-
tion learning is very helpful to improve the authentication rate
for GC-Net especially when the amount of training data sets
is limited. For the VGG-7, the authentication rate of no freez-
ing parameters or freezing the parameters of the first three
convolution layers is higher than that of other cases, while the
authentication rate of TL is the lowest when the parameters of
dense layer are frozen. This is due to the fact that the parame-
ters of the dense layer directly affect the authentication results
of the model, so the parameters of the dense layer should not
be frozen when the model is fine tuned.
Fig. 8 shows that the CNN-2 scheme takes the shortest
training time, and the proposed TL-PHA scheme consumes
slightly higher training time, which demonstrates a graceful
compromise with the gained authentication rate performance.
Along with Figs. 6–8, we conclude that using TL can not
only improve the authentication rate but also save the train-
ing time, which perfectly fits into the EC-related application
scenarios.
3) Analysis of Data Augmentation: The data augmentation
methods are examined on a network by taking 100 original
channel data frames in the training of the CNN of each node,
where the channel SNR and the number of antennas is 2 dB
and 8, respectively. With the SWDA method, the random num-
ber is generated by the normal distribution function N(μ, σ 2)
and the uniform distribution function U(a,b), respectively.
Fig. 9 shows the authentication rate by using the proposed
data augmentation methods on all the ML-based schemes for
15 nodes. Clearly, all the schemes with BSDA can yield better
authentication rates while with a higher convergence speed.
Such an advantage is gained thanks to the fact that BSDA
swaps the elements in the same position of the channel matrix
instead of the correlation between channel matrices and, thus,
can generate more new data samples that facilitate boosting
the authentication rate. On the other hand, SWDA using a
uniform distribution function can yield a higher authentication
rate than the case without data augmentation and the SWDA
method using a normal distribution function. This is because
the channel noise actually follows a Gaussian distribution and,
thus, the corresponding augmented training data sets can be
closer to the real channel data. Furthermore, with the increase
of θ, the number of newly generated channel samples will
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1383
Fig. 6. Comparison of authentication rate between the TL method and the method of training from the scratch among the ML-based schemes for a 15-node
network. The channel SNR is 8 dB and the number of antennas is 8. The number of original channel samples of each node for offline training is 100. The
number of new CSI per node of the online fine tuning authentication model is 5. (a) CNN-2. (b) TP-Net. (c) GC-Net. (d) VGG-7.
Fig. 7. Comparison of authentication rate between the cases of using TL and training from the scratch for 15 nodes. The channel SNR is 8 dB and the
number of antennas is 8. The number of original channel samples of each node for offline training is 100. The number of new CSI per node of the online
fine tuning authentication model is 10. (a) CNN-2. (b) TP-Net. (c) GC-Net. (d) VGG-7.
Fig. 8. Comparison of training time of each scheme under 15 wireless
nodes and different number of training samples with channel SNR as 8 dB
and number of antennas as 8.
decrease but the authentication rate does not changed much,
which is still higher than that without data augmentation.
C. Experiments on NI USRP Platform
Experiments are performed on an NI USRP Platform in an
office room of 8-m long, 7.5-m wide, and 3-m high. The ECN
is simulated by four USRPs, which configure eight transmit
and receive antennas, respectively. Four CDs are simulated
accordingly, where one is equipped with two transmit and
two receive antennas, while the others are simulated by two
USRPs, which configure four transmit and four receive anten-
nas, respectively. The number of original training and testing
frames per CD is 100, respectively. Multiple input and multiple
output-orthogonal frequency-division multiplexing (MIMO-
OFDM) at the ECN is assumed, and the improved-scaled LSs
(ILSs) is adopted for channel measurement [28], [29]. The
center frequency is fc=3.5 Giga Hertz (GHz), the num-
ber of subcarriers is nsubcarrier =128, the sampling interval
tsampling =5×10−7s, the digital modulation method is 4QAM,
and the transmitting power is 15 dBm and transmission gain
20 dB. The server parameters of training CNN are as follows:
the server CPU is with Intel Xeon Silver 4114 with the main
frequency as 2.2 GHz, the physical memory as 15 982 940 kB,
the operating system as Ubuntu 18.04.4 LTS.
Fig. 10 shows the authentication and detection performance,
where the notations follow Figs. 6 and 7. Fig. 10(a) shows
that after several epochs, the detection ability of all schemes
converges and the detection results are perfect, due to the small
number of malicious and legitimate nodes.
From the results of Fig. 10(b), offline training and online
training have no effect on the authentication rate. Fig. 10(c)
demonstrates the authentication rate between the methods of
online TL and training from the scratch among all the ML-
based schemes for four nodes with three training samples
per node. Obviously, the authentication rates of all schemes
are perfect after sufficient training epochs, while by using
the online TL scheme, the least amount of training epochs
is needed, thanks to much less parameters to be fine tuned
online. Fig. 10(d) shows that the consumed training time
highly depends on the size of training data sets, and the train-
ing time by the TP-Net scheme is slightly longer than that
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
1384 IEEE INTERNET OF THINGS JOURNAL, VOL. 9, NO. 2, JANUARY 15, 2022
Fig. 9. Comparison of authentication rate for different configurations of data augmentation methods. The number of nodes is 15, the channel SNR is 2 dB,
and the number of antenna is 8. (a) CNN-2. (b) TP-Net. (c) GC-Net. (d) VGG-7.
Fig. 10. Authentication and detection performance on the USRP testbed. (a) Comparison of detection rate. There are four nodes in total, including one
malicious node and three legitimate nodes. (b) Comparison of authentication rate between online migration authentication of offline training and online
authentication of online training for four nodes. (c) Comparison of authentication rate between the method of TL and the method of training from scratch
for four nodes. The number of original channel samples of each node for offline training is 100. The number of new CSI per node for the online fine tuning
authentication model is 3. (d) Comparison of training time of each scheme under four nodes and different number of training samples.
by CNN-2 but much shorter than that of VGG-7 and GC-Net.
This further attests the advantage of using TL in online appli-
cations such as the PHY-layer authentication considered in the
study.
VIII. CONCLUSION
In this work, we introduced TL-PHA for lightweight user
authentication for latency-sensitive applications such as the
EC. It is featured by incorporating a number of unique
designs, including the TP-Net, which is a novel CNN archi-
tecture jointly employing max, global, and average pooling;
the SWDA and BSDA, which are two novel data augmenta-
tion algorithms; as well as the TL mechanism, which enables
migration of offline training results and fine tuning of network
models for the desired online classification purpose. Extensive
experiments were conducted by using both computer sim-
ulation and an NI USRP testbed. The results showed that
the proposed TL-PHA significantly outperforms all the other
counterparts in terms of authentication accuracy and the abil-
ity of identifying the malicious nodes, while taking the least
computational complexity in the model training phase.
REFERENCES
[1] S. N. Shirazi, A. Gouglidis, A. Farshad, and D. Hutchison, “The
Extended cloud: Review and analysis of mobile edge computing and fog
from a security and resilience perspective,” IEEE J. Sel. Areas Commun.,
vol. 35, no. 11, pp. 2586–2595, Nov. 2017.
[2] R.-F. Liao et al., “Multiuser physical layer authentication in Internet of
Things with data augmentation,” IEEE Internet Things J., vol. 7, no. 3,
pp. 2077–2088, Mar. 2020.
[3] L. U. Khan, I. Yaqoob, N. H. Tran, S. M. A. Kazmi, T. N. Dang,
and C. S. Hong, “Edge-computing-enabled smart cities: A comprehen-
sive survey,” IEEE Internet Things J., vol. 7, no. 10, pp. 10200–10232,
Oct. 2020.
[4] M. Aazam, S. Zeadally, and K. A. Harras, “Health fog for smart
healthcare,” IEEE Consum. Electron. Mag., vol. 9, no. 2, pp. 96–102,
Mar. 2020.
[5] S.-C. Hung, X. Zhang, A. Festag, K.-C. Chen, and G. Fettweis,
“Vehicle-centric network association in heterogeneous vehicle-to-vehicle
networks,” IEEE Trans. Veh. Technol., vol. 68, no. 6, pp. 5981–5996,
Jun. 2019.
[6] L. Xiao, L. J. Greenstein, N. B. Mandayam, and W. Trappe, “Using
the physical layer for wireless authentication in time-variant channels,”
IEEE Trans. Wireless Commun., vol. 7, no. 7, pp. 2571–2579, Jul. 2008.
[7] H. Wen, Y. F. Wang, X. P. Zhu, J. Q. Li, and L. Zhou, “Physical layer
assist authentication technique for smart meter system,” IET Commun.,
vol. 7, no. 3, pp. 189–197, Feb. 2013.
[8] F. Pan et al., “Authentication based on channel state information for
industrial wireless communications,” in Proc. 44th Annu. Conf. IEEE
Ind. Electron. Soc. (IECON), Oct. 2018, pp. 4125–4130.
[9] F. Pan et al., “Threshold-free physical layer authentication based
on machine learning for industrial wireless CPS,” IEEE Trans. Ind.
Informat., vol. 15, no. 12, pp. 6481–6491, Dec. 2019.
[10] H. Fang, X. Wang, and L. Hanzo, “Learning-aided physical layer authen-
tication as an intelligent process,” IEEE Trans. Commun., vol. 67, no. 3,
pp. 2260–2273, Mar. 2019.
[11] L. Xiao, Y. Li, G. Han, G. Liu, and W. Zhuang, “PHY-layer spoofing
detection with reinforcement learning in wireless networks,” IEEE Trans.
Veh. Technol., vol. 65, no. 12, pp. 10037–10047, Dec. 2016.
[12] X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-based fingerprinting
for indoor localization: A deep learning approach,” IEEE Trans. Veh.
Technol., vol. 66, no. 1, pp. 763–776, Jan. 2017.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: ON PHYSICAL-LAYER AUTHENTICATION VIA ONLINE TRANSFER LEARNING 1385
[13] N. Wang, T. Jiang, S. C. Lv, and L. Xiao, “Physical-layer authentication
based on extreme learning machine,” IEEE Commun. Lett., vol. 21, no. 7,
pp. 1557–1560, Jul. 2017.
[14] H. Fang, X. Wang, and S. Tomasin, “Machine learning for intelligent
authentication in 5G and beyond wireless networks,” IEEE Wireless
Commun., vol. 26, no. 5, pp. 55–61, Oct. 2019.
[15] Z. Chen and P. Ho, “Cloud based content classification with global-
connected net (GC-Net),” in Proc. 21st Conf. Innovat. Clouds Internet
Netw. Workshops (ICIN), Feb. 2018, pp. 1–6.
[16] C. Shorten and T. M. Khoshgoftaar, “A survey on image data aug-
mentation for deep learning,” J. Big Data, vol. 6, no. 1, p. 60,
2019.
[17] Y. Li, H. Hu, and G. Zhou, “Using data augmentation in continuous
authentication on smartphones,” IEEE Internet Things J., vol. 6, no. 1,
pp. 628–640, Feb. 2019.
[18] J. Salamon and J. P. Bello, “Deep convolutional neural networks and
data augmentation for environmental sound classification,” IEEE Signal
Process. Lett., vol. 24, no. 3, pp. 279–283, Mar. 2017.
[19] M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger,
and H. Greenspan, “GAN-based synthetic medical image augmenta-
tion for increased CNN performance in liver lesion classification,”
Neurocomputing, vol. 321, pp. 321–331, Dec. 2018.
[20] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans.
Knowl. Data Eng., vol. 22, no. 10, pp. 1345–1359, Oct. 2010.
[21] L. Shao, F. Zhu, and X. Li, “Transfer learning for visual categoriza-
tion: A survey,” IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 5,
pp. 1019–1034, May 2015.
[22] R. Nassif, S. Vlaski, C. Richard, J. Chen, and A. H. Sayed, “Multitask
learning over graphs: An approach for distributed, streaming machine
learning,” IEEE Signal Process. Mag., vol. 37, no. 3, pp. 14–25, May
2020.
[23] G. Oligeri, S. Raponi, S. Sciancalepore, and R. Di Pietro, “PAST-AI:
physical-layer authentication of satellite transmitters via deep learning,”
2020. [Online]. Available: arXiv:2010.05470.
[24] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classifica-
tion with deep convolutional neural networks,” Adv. Neural Inf. Process.
Syst., vol. 25, pp. 1097–1105, Jan. 2012.
[25] K. Simonyan and A. Zisserman, “Very deep convolutional networks
for large-scale image recognition,” 2014. [Online]. Available:
arXiv:1409.1556.
[26] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proc. IEEE Conf. Comput. Vision Pattern Recognit.,
2016, pp. 770–778.
[27] C. Szegedy et al., “Going deeper with convolutions,” in Proc. IEEE
Conf. Comput. Vision Pattern Recognit. (CVPR), 2015, pp. 1–9.
[28] Y. Li, L. J. Cimini, and N. R. Sollenberger, “Robust channel estimation
for OFDM systems with rapid dispersive fading channels,” IEEE Trans.
Commun., vol. 46, no. 7, pp. 902–915, Jul. 1998.
[29] E. G. Larsson, G. Q. Liu, J. Li, and G. B. Giannakis, “Joint symbol tim-
ing and channel estimation for OFDM based WLANs,” IEEE Commun.
Lett., vol. 5, no. 8, pp. 325–327, Aug. 2001.
[30] Y. Liu, Z. Tan, H. Hu, L. J. Cimini, and G. Y. Li, “Channel estimation for
OFDM,” IEEE Commun. Surveys Tuts., vol. 16, no. 4, pp. 1891–1908,
4th Quart., 2014.
[31] A. Sorrentino, G. Ferrara, and M. Migliaccio, “On the coherence
time control of a continuous mode stirred reverberating chamber,”
IEEE Trans. Antennas Propag., vol. 57, no. 10, pp. 3372–3374,
Oct. 2009.
[32] R.-F. Liao et al., “Deep-learning-based physical layer authentication for
industrial wireless sensor networks,” Sensors, vol. 19, no. 11, p. 2440,
May 2019.
[33] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
2014. [Online]. Available: arXiv:1412.6980.
[34] R. Liao, H. Wen, J. Wu, H. Song, F. Pan, and L. Dong, “The rayleigh
fading channel prediction via deep learning,” Wireless Commun. Mobile
Comput., vol. 2018, Jul. 2018, Art. no. 6497340.
[35] R.-F. Liao et al., “Security enhancement for mobile edge com-
puting through physical layer authentication,” IEEE Access,vol.7,
pp. 116390–116401, Aug. 2019.
[36] R. H. Clarke, “A statistical theory of mobile-radio reception,” Bell Syst.
Techn. J., vol. 47, no. 6, pp. 957–1000, Jul.-Aug. 1968.
[37] W. C. Jakes and D. C. Cox, Microwave Mobile Communications.New
York, NY, USA: Wiley-IEEE Press, 1994.
Yi Chen (Student Member, IEEE) received the
master’s degree in communication and information
systems from Guizhou University, Guiyang, China,
in 2016. He is currently pursuing the Ph.D.
degree in communication and information system
with the National Key Laboratory of Science
and Technology on Communications, University
of Electronic Science and Technology of China,
Chengdu, China, under the supervision of Prof. H.
Wen .
He was a visiting Ph.D. student with the
Department of Electrical and Computer Engineering, University of Waterloo,
Waterloo, ON, Canada, from January 2020 to January 2021. His research
interests include communication network, Internet of Things, and physical-
layer authentication.
Pin-Han Ho (Fellow, IEEE) received the Ph.D.
degree from Queen’s University, Kingston, ON,
Canada, in 2002.
He is a Full Professor with the Department of
Electrical and Computer Engineering, University
of Waterloo, Waterloo, ON, Canada. He has
authored/coauthored of more than 400 referred
technical papers, several book chapters, and coau-
thored of two books on optical Internet design and
optimization. His current research interests cover a
wide range of topics in broadband wired and wire-
less communication networks, including survivable network design, wireless
communications, cyber–physical systems, and Internet of Things.
Hong Wen (Senior Member, IEEE) received the
M.Sc. degree in electrical and computer engineering
from Sichuan University, Chengdu, China, in 1997,
the first Ph.D. degree in electrical and computer
engineering from Southwest Jiaotong University,
Chengdu, China, in 2004, and the second Ph.D.
degree in electrical and computer engineering from
the University of Waterloo, Waterloo, ON, Canada,
in 2018.
She is currently a Professor with the School
of Aeronautics and Astronautics, University of
Electronic Science and Technology of China, Chengdu. Her current research
interests include communication systems, physical-layer security, smart grid,
and industrial communications.
Shih Yu Chang (Senior Member, IEEE) received
the B.S.E.E. degree from National Taiwan
University, Taipei, Taiwan, in 1998, and the Ph.D.
degree in electrical engineering and computer
engineering from the University of Michigan, Ann
Arbor, MI, USA, in 2006.
From August 2006 to February 2016, he was
the Faculty with the Department of Computer
Engineering, National Tsing Hua University,
Hsinchu, Taiwan. From July 2007 to August 2007,
he had been a Visiting Assistant Professor with
Television and Networks Transmission Group, Communications Research
Centre, Ottawa, ON, Canada. In June 2018, he began to provide lectures
about machine learning, data science, and AI with San Jose State University,
San Jose, CA, USA. Besides academic position, he also works as an AI
technical lead focusing on applying machine learning techniques to automate
office works.
Shahriar Real received the M.A.Sc. degree in pat-
tern analysis and machine intelligence from the
University of Waterloo, Waterloo, ON, Canada, in
2020.
Alongside completing his degree, he has two
years of intensive research experience work-
ing as a Graduate Research Student with the
University of Waterloo. He is currently work-
ing with KPMG, Amstelveen, The Netherlands, as
a Senior Software Consultant in their Scientific
Research and Experimental Development Team. His
research interest lies in machine learning, neural networks, and software
methodologies.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on January 31,2022 at 17:43:12 UTC from IEEE Xplore. Restrictions apply.
