Received June 7, 2021, accepted July 4, 2021, date of publication July 8, 2021, date of current version July 16,

2021.
Digital Object Identifier 10.1109/ACCESS.2021.3095546

Indoor Localization Using Data Augmentation via

Selective Generative Adversarial Networks
WAFA NJIMA 1 , (Member, IEEE), MARWA CHAFII 1 , (Member, IEEE),
ARSENIA CHORTI 1 , (Senior Member, IEEE), RAED M. SHUBAIR 2 , (Senior Member, IEEE),
AND H. VINCENT POOR 3 , (Life Fellow, IEEE)
1 ETIS UMR 8051, CY Paris Université, ENSEA, CNRS, 95000 Cergy, France.
2 Department of Electrical and Computer Engineering, New York University (NYU) Abu Dhabi, Abu Dhabi, United Arab Emirates
3 Department of Electrical and Computer Engineering, Princeton University, Princeton, NJ 08544, USA
Corresponding author: Wafa Njima ([email protected])
This work was supported in part by CY Initiative of Excellence (grant ‘‘Investissements d’Avenir’’ ANR-16-IDEX-0008) and by the
project DELICATE funded by CNRS/INS2I.

ABSTRACT Several location-based services require accurate location information in indoor environments.
Recently, it has been shown that deep neural network (DNN) based received signal strength indicator (RSSI)
fingerprints achieve high localization performance with low online complexity. However, such methods
require a very large amount of training data, in order to properly design and optimize the DNN model,
which makes the data collection very costly. In this paper, we propose generative adversarial networks
for RSSI data augmentation which generate fake RSSI data based on a small set of real collected labeled
data. The developed model utilizes semi-supervised learning in order to predict the pseudo-labels of the
generated RSSIs. A proper selection of the generated data is proposed in order to cover the entire considered
indoor environment, and to reduce the data generation error by only selecting the most realistic fake RSSIs.
Extensive numerical experiments show that the proposed data augmentation and selection scheme leads to
a localization accuracy improvement of 21.69% for simulated data and 15.36% for experimental data.

INDEX TERMS Indoor localization, received signal strength indicator (RSSI), deep neural network (DNN),
generative adversarial network (GAN), semi-supervised learning.

I. INTRODUCTION geometric mapping are heavily affected by multipath prop-

The proliferation of smart wireless devices leveraging agation effects which blur increasingly the relations between
ubiquitous connectivity, particularly the internet of things physical measurements and distances. Consequently, propa-
(IoT), has driven increased interest in a wide range of ser- gation modeling becomes complicated and localization accu-
vices including indoor localization [1], [2]. In particular, racy is degenerated. As an alternative to geometric methods,
the availability of location information has become a key fac- fingerprinting-based methods have been proposed which
tor in location-based services such as healthcare, homecare, adopt a pattern-matching process. Such methods begin by
monitoring, tracking, etc [3], [4]. To provide user location a site survey task collecting signal features at training posi-
information, different wireless and emerging technologies tions in the area of interest to build a fingerprint database
(e.g. Bluetooth [5], WiFi [6], ultra wide-band (UWB) [7], called also a radio map. Localization is then performed
visible light [8], magnetic field [9]) and signal parameters online calculating the similarity between the fingerprint at an
(e.g. time of arrival (ToA) [10], angle of arrival (AoA) [11], unknown location and the fingerprint database. The estimated
channel state information (CSI) [12], and received signal location corresponds to the best-fitted fingerprint applying
strength indicator (RSSI) [13]) have been investigated. different estimation algorithms. Classical similarity evalua-
Traditional indoor localization systems are mainly based tion and estimation algorithms are very demanding in terms
on geometric [14] (e.g. trilateration and triangulation) and of energy and computation time, since the whole process is
fingerprinting-based methods [15]. The performance of online and it requires to browse huge collected fingerprint
databases. To overcome these issues, machine learning (ML)
The associate editor coordinating the review of this manuscript and based indoor localization systems have been proposed [16],
approving it for publication was Kegen Yu . in particular deep learning (DL) methods [17]–[19].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 98337
 W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

Using such methods, a saved DL model based on collected reference positions as proposed in [34], but we generate RSSI
fingerprint database, trained offline, is used online for accu- data for new unknown positions to cover the whole environ-
rate information location prediction which minimizes the ment. In addition, we propose in this work to apply a selection
online complexity. A variety of localization systems based process of the generated fake data that considers constraints
on DL methods (e.g. for collected data completion [20], on both the coverage and the credibility of the generated data.
noise minimization [21], and location estimation [22]) have The proposed algorithm is validated and tested on simulated
been proposed in the literature and provide good localization environments using realistic propagation model parameters
performance. However, in order to build an efficient DL measured from experiments. To support simulation results,
model, these methods are data intensive, requiring very large we apply our proposed system on real measurements from
amounts of training data, the acquisition of which is a time the public UJIndoorLoc database [36].
and energy consuming task. To address this issue, some The remainder of this paper is organized as follows:
recent approaches that leverage semi-supervised learning Section II formulates the localization problem and describes
have been introduced, combining a small set of collected the system model. The proposed selective semi-supervised
labeled data and a large set of collected unlabeled data GANs for data augmentation is presented in Section III. The
[23]–[25]. Nevertheless, collecting unlabeled data can still environmental setting and obtained simulation results are
be expensive, and thus, in this paper we investigate using provided and discussed in Section IV. Experimental results
generative models [26]–[28] for data augmentation without based on real measurements from the UJIndoorLoc database
any extra data collection. are presented in Section V. Finally, Section VI concludes our
In the last few years, generative models have attracted work.
significant interest in the research community due to their
promising benefits in different fields. Generative models II. INDOOR LOCALIZATION SCENARIOS AND
have shown a good ability to produce realistic supplement FINGERPRINTING TECHNIQUE
various types of data such as images, texts and sounds. In this section, we briefly describe the limitations of exist-
In particular, generative adversarial networks (GANs) have ing indoor localization systems based on RSSI fingerprints.
been employed [29]–[31] to generate additional measure- And, we detail different steps of the proposed system model
ments, which reduces collection time and saves human developed to address the problem of interest.
effort. Such networks generate samples with improved diver-
sity and expand the training database in order to ensure a A. SYSTEM MODEL DESCRIPTION
proper design of deep neural networks (DNNs) in different We consider an indoor environment covering (L × W ) m2 ,
fields including localization. In [32], GANs are used with where WiFi technology is deployed using M access points
semi-supervised learning where GANs use both labeled and (APs). As depicted in Fig. 1, the system consists of two main
unlabeled data to share weights with a localization classifier parts: the central unit (CU) which performs all the data pro-
in order to benefit from data contained in unlabeled infor- cessing and localization, and the mobile sensor nodes. There
mation when labeled data are not sufficient. The authors are two types of mobile sensor nodes: those used during the
in [33] aim to construct an efficient radio map covering training phase in order to collect RSSI measurements for the
free space (e.g, open spaces and corridors) and constrained training database construction and those requesting online to
spaces. In this experiment, the data corresponding to the free be localized. To enhance the richness of the fingerprint RSSI
space environment are measured, whereas the data for the database, GANs are used to generate fake RSSI fingerprints.
constrained space are artificially generated. This is justified Let pij be the RSSI measured at the i-th position of the signal
by the difficulties in performing measurement under space transmitted by the j-th AP.
constraints. In the works described in [34], and [35], the ini-
tial training fingerprint database is expanded increasing the B. FINGERPRINTING BASED LOCALIZATION
amount of training data collected at each reference point In this section, we present the standard fingerprinting tech-
based on GANs in order to enhance its richness. For this, nique when used with DNN models. Then, we detail briefly
a collection of signal measurements have been done at dif- our system model steps including data collection and data
ferent reference points and GANs have been used in order training.
to generate additional measurements at each reference point
to enhance the diversity of collected data at each reference 1) DNN-BASED FINGERPRINTING
position. RSSI-based positioning systems can be divided into two cat-
In this paper, unlike [32], we do not assume the availability egories: geometric methods and fingerprints-based methods.
of unlabeled data to enhance the training of the generative Among them, fingerprinting is widely studied and adopted
model, and we do not assume having sufficient collected because of its localization accuracy and implementation sim-
data for particular regions as assumed in [33] for free space. plicity. It consists of two main phases: an offline phase
We consider an extreme and realistic case where only a small called also a training phase and an online phase called a
amount of labeled data is available. We also do not use GANs test phase. During the offline phase, RSSI measurements are
to increase the diversity of the RSSI data for specific known collected at each measurement location (training position)

98338 VOLUME 9, 2021

W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

FIGURE 1. The pipeline of the proposed selective GANs based localization method during the training phase.

from different radio signal transmitters, essentially APs when a proper data selection protocol should be developed to over-
considering WiFi signals. These collected RSSIs with their come these issues.
location coordinates are transferred to a CU and stored in a
training database. In the online phase, RSSIs from different 2) DATA COLLECTION AND DATA TRAINING
APs are measured and compared with the ones collected and In this part, we present in details the collection of labeled data,
stored in the training database to predict the location of the as well as the whole process of collected data pre-processing
target. This prediction can be performed by a DNN model and training.
which can be trained offline and applied directly online. The
use of DNN models for location prediction is recommended a. Fingerprint database collection: In this paper, we con-
since most of the online complexity is shifted to the offline sider a noisy indoor environment as the area of interest
phase. DNN is recommended for accurate indoor localization in order to collect RSSI fingerprints during the offline
due to the fact that it is able to learn signal fluctuations phase including training and testing data.
through time and environmental dynamicity because of its b. Data augmentation: Data augmentation is used to sup-
deeper functions that map the input to the output [37], [38]. plement data when it is too expensive and rare. Thus,
In order to achieve good localization accuracy using an opti- it aims to increase the size and the diversity of a dataset
mal trained DNN model, a large amount of expensively col- by generating new fake samples based on real ones.
lected data is required during the offline phase which makes • Data generation: Based on collected RSSI vectors,
the fingerprinting method labor intensive and time/cost con- extra fake data are generated using a GAN in order
suming. To solve this issue, a selective GAN indoor localiza- to expand the dataset and ensure its diversity.
tion framework combined with semi-supervised learning is • Data pseudo-labeling: Once fake RSSI vectors are
proposed and detailed in this paper, which performs location generated, an artificial pseudo-label is associated with
prediction based on collected labeled data and fake gener- each vector using semi-supervised learning.
ated pseudo-labeled data. This system takes advantage of c. Data selection: After data generation and pseudo-
generating fake data and mixing it with labeled data in the labeling, we apply a data selection method. We first
training process in order to reduce the reliance on expensively generate a very large number of fake samples to cover
collected data. However, generated data can be unrealistic the whole environment, and then we eliminate unneces-
and can create imbalanced classes when the generated data sary and inaccurate RSSI data in order to improve the
happens to be more dense for specific regions. Therefore, localization accuracy.

VOLUME 9, 2021 98339

 W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

d. Localization based on DNNs: To localize a target, discriminator model D. For each DNN model, we introduce
a trained DNN model is applied to estimate its coordi- an input vector i(0) ∈ RN0 ×1 and its associated output vector
nates based on the corresponding collected RSSI vector. o ∈ RNH +1 ×1 . Let Nh be the number of neurons for the hth
During the training process, all of the RSSI vectors layer, such that 0 ≤ h ≤ H + 1 where H is the number
(e.g, real collected vectors and selected fake gener- of hidden layers. bh ∈ RNh ×1 and Wh ∈ RNh ×Nh−1 denote
ated vectors) are fitted to the model as inputs and the the biases and the weights matrices, respectively. The output
corresponding coordinates are considered as outputs. vector of the hth layer can be expressed as
o(h) = gh (b(h) + W(h) i(h−1) ), i(h) = o(h) , (1)
III. PROPOSED SELECTIVE SEMI-SUPERVISED GAN
FOR LOCATION DATA AUGMENTATION where the input vector i(h) undergoes a linear transformation
In this section, we describe in detail different steps of the represented by W(h) , a bias vector b(h) , and then a nonlinear
proposed localization system. activation function g(h) is applied. During DNN training,
the loss function L(θ) = L(W , b) is calculated in order to
A. DATA GENERATION update iteratively the DNN parameters θ = (W , b).
1) INTUITION BEHIND GANs • Training the generator: Let Z ∈ RM ×mf be the input

Generative models aim at learning the true data distribution dataset of the generator, such that each column z(i) ∈
of a training set in order to generate realistic new sam- RM ×1 , i = 1, · · · , mf corresponds to a random noise
ples with some variations. Thus, we try to produce samples vector whose samples are uniformly distributed over
with a distribution that is as similar as possible to the true [−1, 1[. At the output of the DNN generator, each input
data distribution. GANs are generative models that include noise vector z(i) produces a fake RSSI vector G(z(i) ) ∈
two components [39]: the generator and the discriminator, RM ×1 . The fake output vector is generated following (1),
as shown in Fig. 2. The generator model takes random noise where i(0) = z(i) , N0 = M , o(H +1) = G(z(i) ) and
as input and learns how to produce a realistic output rep- NH +1 = M . Then, G(z(i) ) is passed to the DNN dis-
resentation similar to the real data, while the discriminator criminator D(·) which makes a prediction of how real-
learns how to distinguish between fake and real samples. istic G(z(i) ) is. The generator loss is then calculated by
These two models are trained together until the generator is minimizing the loss function:
able to generate realistic examples from input noise. GANs mf
h 1 X 
i
have achieved impressive performance across a multitude of LG = ∇θg − log D G(z(i) ) . (2)
tasks (e.g. face generation, 3D object generation, and image mf
i=1
translation from one domain to another) and many companies Once the loss function is minimized, and the parameters
are using them including Google for text generation, IBM for θg = (Wg , bg ) do not change significantly after further
data augmentation, Adobe for next generation photoshop and iterations, we can save the generator DNN parameters.
Snapchat-Tiktok for image filters. • Training the discriminator: The discriminator learns
GANs are used for data augmentation in our case, which to distinguish real from fake. At the beginning of the
means the generated data can be used to supplement real training, it does not know which vectors are real and
data to increase dataset size and diversity by using generated which ones are fake. However, it has access to real
samples. The goal is to augment the dataset when real data data in order to compare them with input vectors and
are too expensive to collect. classify the input vectors as real or fake. Thus, the dis-
criminator receives more and more realistic examples
2) GAN TRAINING at each round from the generator till the examples are
GANs consist of two different neural networks: the gener- good enough to fill the discriminator. The discriminator
ator and the discriminator. To train a GAN, we alternate acts as a binary classifier and computes a probability of
the training of the DNN generator model G and the DNN an example being fake D(G(z(i) )), following (1), giving

FIGURE 2. GAN network for fake RSSI data generation during training.

98340 VOLUME 9, 2021

W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

both fake generated samples G(Z ), and the real dataset process, these fake vectors are going to be pseudo-labeled
P ∈ RM ×mr of collected mr RSSIs vectors denoted as and selected. For this, collected RSSI vectors are used as
p(i) , i = 1, . . . , mr . This probability will be given to inputs to a supervised DNN model that takes as outputs
the generator to improve its performance as expressed the corresponding labels, e.g. coordinates during training.
in (2). The training of the discriminator is performed Once trained, fake generated RSSI vectors are given to the
by minimizing the loss function derived from the binary model to predict the associated artificial pseudo-labels. Then,
cross-entropy (BCE) between the real and generated a general model used for localization is trained on real and
data [40]: selected data. The selection procedure is described in detail
h 1 X
mr in subsection III-C.
log D(p(i) )


Ld = ∇θd −
mr
i=1 C. SELECTION CRITERIA OF GENERATED FAKE DATA
mf
1 X 
i Initially, the idea of this work was to generate RSSI vectors,
− log 1 − D G(z(i) ) , (3) to generate more data and cover the whole environment, and
mf
i=1 then to estimate their pseudo-labels to build a mixed and
where θd = (Wd , bd ) are the parameters of the discrim- rich dataset. We have generated a large amount of fake data
inator DNN model D(.). log(D(p(i) )) refers to the log and tested the resulting system. However, when plotting the
probability that the discriminator is correctly classifying pseudo labels of the generated RSSIs, i.e. positions, we have


the real examples while maximizing log 1−D G(z(i) ) noticed that (i) generated positions do not cover the whole
area of the considered environment, (ii) these positions are
would help the discriminator to correctly label the fake sometimes condensed in a specific area, and (iii) generating
sample that comes from the generator. a higher number of fake data does not necessarily lead to
a better localization accuracy. Therefore, we introduce here
3) FAKE DATA GENERATION BASED ON GANs
selection criteria in order to choose only a useful subset of
Given a small number of training labeled data samples, extra the generated fake data:
fake RSSI vectors are generated based on the GAN. The
generator takes as inputs noise vectors that are going to be • Criterion 1: Environment coverage: Here, we con-
fitted to the discriminator with the real RSSI vectors. Based sider that our environment E is uniformly divided into
on the discriminator’s outputs, the generator and the discrimi- zones ej each one covering (lj × wj ) m2 , such that
2
P
nator models are updated in order to enhance its performance j lj × wj = L × W m . After randomly generating
generation and classification. We generate mf fake RSSI a large amount of data (RSSI vectors and the pseudo
vectors ∈ R(M ×1) based on the dataset of collected RSSI positions), we build a new dataset by selecting for each
vectors, mf is fixed experimentally in order to choose the best zone ej a number Nj of data samples proportional to its
number of generated vectors that gives the best localization surface area (lj ×wj ) m2 , i.e. the pseudo positions fall into
improvement. the desired zone ej . Thus, redundant data are eliminated,
and the whole environment E is covered and each zone
B. SEMI-SUPERVISED FOR PSEUDO-LABELING ej is equally
P represented in the new dataset that contains
OF GENERATED DATA ms = j Nj selected sample. We have first tried to select
Pseudo-labeling aims to estimate the labels of an unlabeled randomly Nj samples per zone ej , but we have realized
dataset given a DNN model trained on a labeled one. In the that this random selection can be improved by selecting
localization context, the labels correspond to the users’ loca- only the most realistic ms fake data, which leads us to
tion information (a room ID, a floor ID, a zone identifier, Criterion 2.
2-D / 3-D coordinates, etc.). The 2-D coordinates are the used • Criterion 2: Most realistic fake data: The selection
labels in our work. The steps of the pseudo-labeling process of the Nj most realistic fake data samples per zone ej
are summarized as follows: is performed by comparing the score of each generated
• Step 1: The DNN model is trained on labeled RSSI RSSI G(z(i) ) at the output of the discriminator D(G(z(i) )).
fingerprints only, in a supervised way. Thus, in each zone ej , we select the Nj positions that are
• Step 2: Based on the trained DNN model, ’pseudo- most likely to be real, i.e. associated with the Nj lowest
labels’ are predicted for generated unlabeled RSSI vec- loss function values:
tors. mr
• Step 3: A mixed DNN model is trained combining (i)
h 1 X
log D(p(j) )


Ld = ∇θd −
labeled and selected pseudo-labeled data. To train such a mr
j=1
model, the collected and selected RSSI vectors are used 
i
as inputs, and the corresponding associated labels (real − log 1 − D G(z(i) ) , (4)
labels and artificial pseudo-labels) are used as outputs.
During the fake data generation step, we produce unlabeled where p(j) , j = 1, . . . , mr is the jth real collected RSSI
RSSI vectors. In order to be integrated into our localization vector.

VOLUME 9, 2021 98341

 W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

IV. RESULTS BASED ON SIMULATED DATA

In this section, we present the environmental setup and differ-
ent used DNN architectures and hyper-parameters. We also
present the localization performance of our proposed system
based on selective GANs and compare it with other methods
in order to showcase its usefulness in terms of localization
accuracy and data collection cost on a simulated environment.

A. PROPAGATION ENVIRONMENT AND

ENVIRONMENTAL SETUP
To model the propagation environment, we consider that
signals can be degraded and blocked by the obstacles:
• Blockage: 40% of RSSI measured data are unknown.
We assign the value of −110 dBm to non-detected APs
in order to have a weak signal that does not affect the
calculation process. The choice of this value is based
on several experiments in our simulated indoor environ- FIGURE 3. The location of training and test positions on the simulated
indoor environment covering 400 m2 .
ments.
• Degradation is considered combining pathloss and
shadowing effects. TABLE 1. Different parameters used during the simulation process.

Let pij be the RSSI measured at the i-th position of the

signal transmitted by the j-th AP. It can be expressed as
pij = pt − pLij + Bσ [dBm], (5)
where pt is a constant transmitted power, Bσ is a zero-mean
Gaussian distributed random variable with standard deviation
σ representing the shadowing effects, and pLij is the pathloss
calculated as follows:
d 
ij
pLij = pL0 + 20 log10 (f ) + 10µ log10 , (6)
d0
where pL0 denotes the pathloss value at a reference distance
d0 , f the carrier frequency, µ the pathloss exponent whose
value characterizes a specific environment and can be calcu-
lated empirically based on collected measurements and dij is
the distance between the j-th AP and the i-th position.
In this part, we consider a sensor network composed of
M = 10 APs placed randomly in an indoor environment
covering L × W = 400 m2 . In order to evaluate the local-
ization accuracy, we use a training database composed of the whole environment, we place randomly each two test
mr = 1000 RSSI vectors collected at 100 positions labeled positions in each zone of size 1 m2 as presented with red
by their coordinates. At each labeled position, we collect crosses in Fig. 3.
10 RSSI measurements which helps to minimize temporal
RSSI fluctuations caused by shadowing and fading effects. B. DATA GENERATION MODELS
These positions are distributed in a uniform way as illus- In this section, we detail the architectures and training param-
trated in Fig. 3. We divide the indoor environment covering eters used by different DNN models during the data genera-
L × W m2 into zones/classes ej of size lj × wj such that tion process.
∀j, lj = wj = l and place a labeled position in the center
of each zone in order to cover uniformly the whole area. 1) GAN ARCHITECTURES AND PARAMETERS
RSSI vectors are constructed as described above. We detail Generating useful fake data is not straightforward. For exam-
in Table 1 the parameters used for the propagation model and ple, when generating mf = 1000 fake RSSI vectors based on
the simulation parameters. The test database is built using mr = 1000 real RSSI vectors, and predicting their pseudo-
the same propagation model as the one used for training positions, we have realized that the generated data do not
data construction, but with different trajectories. We consider cover the whole environment as depicted in Fig.4. Thus,
mt = 8000 test labeled RSSI fingerprints collected at 800 test we generate a very large number of RSSI vectors in order
positions. In order to have a test dataset representative of to have pseudo-positions covering the whole area. Based on

98342 VOLUME 9, 2021

W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

for generated RSSI vectors. We use ADAM as the optimiza-

tion algorithm [41] for all DNN models used in this work
whether for pseudo-labeling or localization. The intensive
experiments have led to a DNN architecture of 2 hidden layers
with 30 neurons in the first layer and 20 neurons in the second
one using 200/250 epochs and a mini batch size equal to
50/100. Once generating and pseudo-labeling mf = 40000,
1000 are selected, by selecting samples per covered zone of
size 4m2 . For each covered zone, the selected samples are
associated with the most realistic fake RSSIs. Note that if we
skip Criterion 1, and we select directly the 1000 most realistic
fake data samples, we will end up with samples not uni-
formly distributed over the whole environment as presented
in Fig. 5b.

C. DNN MODELS USED FOR LOCALIZATION:

FIGURE 4. 1000 generated positions based on 100 training labeled ARCHITECTURES AND TRAINING PARAMETERS
positions.
For evaluation purposes, we compare the localization accu-
racy combining labeled and selected fake generated data.
We called each algorithm depending on the nature and the
extensive simulations, generating mf = 40000 fake samples
number of used samples as follows:
seems to cover our environment of 400 m2 . For example,
when generating mf = 10000 fake data samples, we cover • Supervised(mr ,mp ): when using a supervised method
only 362 zones of size 1 m2 each while with mf = 40000, based on mr labeled data collected at mp different
we cover 396 zones of size 1 m2 as can be seen in Fig. 5a. positions.
In order to improve the diversity of the GAN, we do not gen- • Selective-SS-GAN(mr ,mp ,ms ) is the localization
erate 40000 RSSI vectors at once, but we generate 10 times method where we combine mr labeled samples collected
4000 fake samples based on the same real samples. In this at mp different positions, and ms pseudo-labeled selected
paper, we introduce GANs based on a DNN optimized with fake generated samples for localization.
adaptative moment estimation (ADAM) using 0.01 as the To apply localization, we need to train a DNN model on
learning rate during 200 epochs. The activation function used labeled collected data and pseudo-labeled data. This model
by G is the rectified linear unit (ReLU) function used in one takes the RSSI vectors as inputs and the outputs are the
hidden layer having 10 neurons. The activation function of D corresponding labels which are the positions’ coordinates.
is the ReLU function, while the last layer uses the sigmoid A learning rate equal to 0.01 has been selected using 250
function. A-one-hidden- layer discriminator with 10 neurons epochs and a mini batch size equal to 100. The DNN archi-
is used. These choices are based on several experiments and tectures used are summarized in Table 2 where FCi (·) refers
tests. to the number of neurons in the ith fully connected layer.

2) PSEUDO-LABELING ARCHITECTURES AND PARAMETERS D. LOCALIZATION ACCURACY

Once all of the data is generated, a DNN is trained based on In this section, we present the localization accuracy of
labeled data. This model is then used to predict pseudo-labels our proposed system based on augmented dataset and we

FIGURE 5. Generated positions and selected positions based on defined selection criterions.

VOLUME 9, 2021 98343

 W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

TABLE 2. Obtained localization performance considering 1000 labeled data samples.

compare it with the scheme where only labeled collected data

samples are used for localization. Results corresponding to
these data are shown in Table 2.
In Table 2, we present the localization accuracy (i.e.,
localization mean error) for Selective-SS-GAN(mr , mp , ms )
trained over mr = 1000 labeled data samples and different
numbers of selected fake data samples vs. supervised learning
model trained only on mr = 1000 labeled samples. We select
different numbers of fake data samples [100 − 4000] from
the set of mf = 40000 generated samples while satisfy-
ing Criterion 1 and Criterion 2. We notice that for all aug-
mented datasets, the localization accuracy is improved com-
pared to a dataset limited only to real data. The localization
improvement varies between 17.92% and 21.96%. The best
localization accuracy is obtained with ms = 1000 selected
fake data samples, where we achieve 21.96% localization FIGURE 6. Localization performance comparison.
accuracy increase vs. the conventional supervised algorithm
without any additional cost in collecting additional data. This
improvement is explained by the consideration that the DNN
performance is close with only 2 cm of difference, which
has been trained over a larger dataset which covers new
means that we can achieve almost the same performance by
positions that are not observed in the limited dataset based
only collecting half of the labeled data and artificially gener-
on collected data. Moreover, we can see that generating more
ating the other half. Thus, the proposed data generation and
fake data does not necessarily improve the localization accu-
selection process provides an improvement of localization
racy, which can be explained by the GAN generation error due
accuracy without additional collection cost.
to generating less realistic RSSI vectors. Note that starting
from 2000 generated samples, the performance saturates and
no improvement can be achieved by generating additional V. RESULTS BASED ON EXPERIMENTAL DATA
fake data. This can be explained by the fact that based on 1000 In order to support the simulation results, we validate and
labeled vectors collected at 100 training positions, we cannot test our system on real RSSI measurements from the pub-
provide a higher measurement diversity to the GAN. lic UJIndoorLoc database. This database covers the area of
Fig. 6 illustrates the cumulative distribution functions almost 110000 m2 including three buildings with four or five
(CDFs) of three algorithms combining data subsets of dif- floors. For implementation simplicity and since we work on
ferent types (labeled and generated). We can easily notice one-floor area, we consider collected data corresponding to
that the supervised indoor localization system based on 1000 the second floor from building 1 only. There are training
labeled samples collected at 100 known positions corre- measurements and validation measurements collected four
sponds to the worst localization accuracy. For fair compar- months later than training ones. For the considered floor,
ison, we use the same dataset of labeled positions to which we have 1395 training fingerprints taken at 80 training posi-
we add (i) 1000 labeled measurements collected at 1000 tions and 40 validation positions taken at different valida-
different labeled positions placed randomly in the considered tion positions. For fair comparison and since we consider
area i.e. Supervised(2000,1100), and (ii) based on these data, only 1000 training fingerprints as system input data during
we generate and select 1000 fake positions i.e. Selective- simulations, we choose randomly 1000 fingerprints from
SS-GAN(1000,100,1000). We notice that the localization the training set and we add the 395 rest training positions

98344 VOLUME 9, 2021

W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

TABLE 3. Obtained localization performance considering 1000 real fingerprints corresponding to Building1-Floor2 from UJIndoorLoc database.

to the 40 validation positions to be used for test. During VI. CONCLUSION

UJIndoorLoc construction, 520 APs are used. However, only In this paper, we have seen that GANs can be used to produce
18 APs are detected at each position and 91% of collected synthetic data, outstandingly realistic, to complement a real
RSSI are unknown. Consequently, for implementation sim- dataset in order to enhance the training of a DNN used for
plicity, we consider only APs detected at least once during localization. Thus, this technique is very useful in situations
data collection which are equal to 190. where data collection is expensive and time consuming such
After data re-organisation and reduction, we are able to as indoor localization. In particular, we have presented a
apply our algorithm. At first, we generate mf = 28000 fake selective semi-supervised GAN system for indoor localiza-
RSSI vectors with GANs based on DNN, optimized using tion, where we have generated fake data based on real labeled
0.01 as learning rate during 200 epochs. A one-hidden layer collected data in order to boost localization performance. Our
discriminator with 200 neurons and a one-hidden layer gen- proposed solutions have been validated and tested based on
erator with 200 neurons are considered. For the optimization several simulations which show that the combination of col-
algorithm and the activation functions, we use the ones used lected data and selected generated data is beneficial in terms
during simulations. Secondly, a DNN is trained on labeled of localization performance and data collection cost. The
measurements during 200 epochs with 50 as a batch size selection-generation process improves the localization accu-
for pseudo-labels estimation. The best achievable DNN is racy by 21.69% compared to the standard supervised method
composed by three hidden layers containing 200 neurons, based on the same subset of labeled data when considering
100 neurons and 50 neurons, respectively. For data selection, 1000 labeled samples collected at 100 different positions. Our
we consider that we split the zone of interest into 625 classes method works also for real data, from the public UJIndoorLoc
with l = 6.4959 m and w = 5.9760 m. The data selection database, leading to a localization accuracy improvement of
process is followed by location estimation applying a DNN 15.36%. The promising results of this paper motivate further
trained offline. A learning rate equal to 0.01 has been selected extension of the developed model by exploring new methods
using 150/250 epochs and a mini batch size equal to 50/100. for data augmentation.
The DNN architectures used for localization are presented
in Table 3. REFERENCES
We evaluate the localization accuracy of different [1] F. Zafari, A. Gkelias, and K. K. Leung, ‘‘A survey of indoor localization
algorithms using both real and selected fake data i.e. systems and technologies,’’ IEEE Commun. Surveys Tuts., vol. 21, no. 3,
Selective-SS-GAN (mr ,ms ). Table 3 presents the localization pp. 2568–2599, 3rd Quart., 2019.
performance (e.g., mean localization error, min localiza- [2] A. Yassin, Y. Nasser, M. Awad, A. Al-Dubai, R. Liu, C. Yuen, R. Raulefs,
and E. Aboutanios, ‘‘Recent advances in indoor localization: A survey on
tion error and max localization error) and the localization theoretical approaches and applications,’’ IEEE Commun. Surveys Tuts.,
improvement compared of to the localization system using vol. 19, no. 2, pp. 1327–1346, 2nd Quart., 2016.
only 1000 labeled data i.e. Supervised (1000). We select [3] J. A. Del Peral-Rosado et al., ‘‘Whitepaper on new localization methods
for 5G wireless systems and the Internet-of-Things,’’ 2018.
different numbers of fake positions [700 - 2000] from the
[4] A. Bourdoux, A. N. Barreto, B. van Liempd, C. de Lima, D. Dardari,
set of mf = 28000 generated positions. We notice that the D. Belot, E.-S. Lohan, G. Seco-Granados, H. Sarieddeen, H. Wymeersch,
localization improvement varies between 6.58% and 15.36%. J. Suutala, J. Saloranta, M. Guillaud, M. Isomursu, M. Valkama,
The best localization accuracy is obtained with ms = 1000 M. R. K. Aziz, R. Berkvens, T. Sanguanpuak, T. Svensson, and Y. Miao,
‘‘6G white paper on localization and sensing,’’ 2020, arXiv:2006.01779.
selected generated positions, where we achieve 15.36% local- [Online]. Available: http://arxiv.org/abs/2006.01779
ization accuracy increase vs. the conventional supervised [5] S. Liu, Y. Jiang, and A. Striegel, ‘‘Face-to-face proximity estimationusing
algorithm. Starting from 1500 selected generated positions, Bluetooth on smartphones,’’ IEEE Trans. Mobile Comput., vol. 13, no. 4,
pp. 811–823, Apr. 2013.
we notice that the performance is saturated and we cannot
[6] S. He and S.-H. G. Chan, ‘‘Wi-Fi fingerprint-based indoor positioning:
provide better accuracy based on training positions. Thus, Recent advances and comparisons,’’ IEEE Commun. Surveys Tuts., vol. 18,
experimental results strengthen results based on simulated no. 1, pp. 466–490, Jan. 2015.
data. Even when working in a realistic environment with high [7] B. Kempke, P. Pannuto, and P. Dutta, ‘‘SurePoint: Exploiting ultra wide-
band flooding and diversity to provide robust, scalable, high-fidelity indoor
dynamic and heterogeneous devices, our proposed system localization: Demo abstract,’’ in Proc. 14th ACM Conf. Embedded Netw.
achieves good localization accuracy. Sensor Syst. CD-ROM, Nov. 2016, pp. 137–149.

VOLUME 9, 2021 98345

 W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

[8] A. B. M. M. Rahman, T. Li, and Y. Wang, ‘‘Recent advances in indoor [31] A. Belmonte-Hernández, G. Hernandez-Penaloza, D. M. Gutiérrez, and
localization via visible lights: A survey,’’ Sensors, vol. 20, no. 5, p. 1382, F. Alvarez, ‘‘Recurrent model for wireless indoor tracking and positioning
Mar. 2020. recovering using generative networks,’’ IEEE Sensors J., vol. 20, no. 6,
[9] W. Shao, H. Luo, F. Zhao, Y. Ma, Z. Zhao, and A. Crivello, ‘‘Indoor pp. 3356–3365, Mar. 2020.
positioning based on fingerprint-image and deep learning,’’ IEEE Access, [32] K. M. Chen and R. Y. Chang, ‘‘Semi-supervised learning with GANs
vol. 6, pp. 74699–74712, 2018. for device-free fingerprinting indoor localization,’’ in Proc. IEEE Global
[10] N. Tadayon, M. T. Rahman, S. Han, S. Valaee, and W. Yu, ‘‘Decimeter Commun. Conf. (GLOBECOM), Dec. 2020, pp. 1–6.
ranging with channel state information,’’ IEEE Trans. Wireless Commun., [33] H. Zou, C.-L. Chen, M. Li, J. Yang, Y. Zhou, L. Xie, and C. J. Spanos,
vol. 18, no. 7, pp. 3453–3468, Jul. 2019. ‘‘Adversarial learning-enabled automatic WiFi indoor radio map construc-
[11] L. Chen, I. Ahriz, and D. Le Ruyet, ‘‘AoA-aware probabilistic indoor tion and adaptation with mobile robot,’’ IEEE Internet Things J., vol. 7,
location fingerprinting using channel state information,’’ IEEE Internet no. 8, pp. 6946–6954, Aug. 2020.
Things J., vol. 7, no. 11, pp. 10868–10883, Nov. 2020. [34] Q. Li, H. Qu, Z. Liu, N. Zhou, W. Sun, S. Sigg, and J. Li, ‘‘AF-DCGAN:
[12] Z. Yang, Z. Zhou, and Y. Liu, ‘‘From RSSI to CSI: Indoor localization Amplitude feature deep convolutional GAN for fingerprint construction in
via channel response,’’ ACM Comput. Surv., vol. 46, no. 2, pp. 1–32, indoor localization systems,’’ IEEE Trans. Emerg. Topics Comput. Intell.,
2013. vol. 5, no. 3, pp. 468–480, Jun. 2021.
[13] P. Kumar, L. Reddy, and S. Varma, ‘‘Distance measurement and error esti- [35] Y. Lei, D. Li, H. Zhang, and X. Li, ‘‘Wavelet feature outdoor fingerprint
mation scheme for RSSI based localization in wireless sensor networks,’’ localization based on ResNet and deep convolution GAN,’’ Symmetry,
in Proc. 5th Int. Conf. Wireless Commun. Sensor Netw. (WCSN), Dec. 2009, vol. 12, no. 9, p. 1565, Sep. 2020.
pp. 1–4. [36] J. Torres-Sospedra, R. Montoliu, A. Martínez-Usó, J. P. Avariento,
[14] Z. Yang and Y. Liu, ‘‘Quality of trilateration: Confidence-based itera- T. J. Arnau, M. Benedito-Bordonau, and J. Huerta, ‘‘UJIIndoorLoc: A new
tive localization,’’ IEEE Trans. Parallel Distrib. Syst., vol. 21, no. 5, multi-building and multi-floor database for WLAN fingerprint-based
pp. 631–640, May 2010. indoor localization problems,’’ in Proc. Int. Conf. Indoor Positioning
Indoor Navigat. (IPIN), Oct. 2014, pp. 261–270.
[15] S. Yiu, M. Dashti, H. Claussen, and F. Perez-Cruz, ‘‘Wireless RSSI finger-
[37] K. S. Kim, R. Wang, Z. Zhong, Z. Tan, H. Song, J. Cha, and S. Lee,
printing localization,’’ Signal Process., vol. 131, pp. 235–244, Feb. 2017.
‘‘Large-scale location-aware services in access: Hierarchical building/floor
[16] M. I. AlHajri, N. T. Ali, and R. M. Shubair, ‘‘Indoor localization for IoT
classification and location estimation using Wi-Fi fingerprinting based on
using adaptive feature selection: A cascaded machine learning approach,’’
deep neural networks,’’ Fiber Integr. Opt., vol. 37, no. 5, pp. 277–289,
IEEE Antennas Wireless Propag. Lett., vol. 18, no. 11, pp. 2306–2310,
Sep. 2018.
Nov. 2019.
[38] K. S. Kim, S. Lee, and K. Huang, ‘‘A scalable deep neural network
[17] W. Njima, I. Ahriz, R. Zayani, M. Terre, and R. Bouallegue, ‘‘Deep CNN
architecture for multi-building and multi-floor indoor localization based on
for indoor localization in IoT-sensor systems,’’ Sensors, vol. 19, no. 14,
Wi-Fi fingerprinting,’’ Big Data Anal., vol. 3, no. 1, pp. 1–17, Dec. 2018.
p. 3127, Jul. 2019.
[39] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
[18] W. Njima, M. Chafii, A. Nimr, and G. Fettweis, ‘‘Deep learning based data S. Ozair, A. Courville, and Y. Bengio, ‘‘Generative adversarial net-
recovery for localization,’’ IEEE Access, vol. 8, pp. 175741–175752, 2020. works,’’ 2014, arXiv:1406.2661. [Online]. Available: http://arxiv.org/
[19] Z. Chen, M. I. AlHajri, M. Wu, N. T. Ali, and R. M. Shubair, ‘‘A novel abs/1406.2661
real-time deep learning approach for indoor localization based on RF [40] L. Mescheder, A. Geiger, and S. Nowozin, ‘‘Which training methods for
environment identification,’’ IEEE Sensors Lett., vol. 4, no. 6, pp. 1–4, GANs do actually converge?’’ in Proc. Int. Conf. Mach. Learn., 2018,
Jun. 2020. pp. 3481–3490.
[20] R. Wang, Z. Li, H. Luo, F. Zhao, W. Shao, and Q. Wang, ‘‘A robust Wi-Fi [41] D. P. Kingma and J. Ba, ‘‘Adam: A method for stochastic opti-
fingerprint positioning algorithm using stacked denoising autoencoder and mization,’’ 2014, arXiv:1412.6980. [Online]. Available: http://arxiv.org/
multi-layer perceptron,’’ Remote Sens., vol. 11, no. 11, p. 1293, May 2019. abs/1412.6980
[21] J. Fan and T. Chow, ‘‘Deep learning based matrix completion,’’ Neurocom-
puting, vol. 266, pp. 540–549, Nov. 2017.
[22] R. Ayyalasomayajula, A. Arun, C. Wu, S. Sharma, A. R. Sethi, D. Vasisht,
and D. Bharadia, ‘‘Deep learning based wireless localization for indoor
navigation,’’ in Proc. 26th Annu. Int. Conf. Mobile Comput. Netw.,
Apr. 2020, pp. 1–14.
[23] W. Li, C. Zhang, and Y. Tanaka, ‘‘Pseudo label-driven federated learning-
based decentralized indoor localization via mobile crowdsourcing,’’ IEEE
Sensors J., vol. 20, no. 19, pp. 11556–11565, Oct. 2020.
[24] Y. Xia, L. Ma, Z. Zhang, and Y. Wang, ‘‘Semi-supervised positioning algo-
rithm in indoor WLAN environment,’’ in Proc. IEEE 81st Veh. Technol.
Conf. (VTC Spring), May 2015, pp. 1–5.
[25] M. Zhou, Y. Tang, W. Nie, L. Xie, and X. Yang, ‘‘GrassMA: Graph-
based semi-supervised manifold alignment for indoor WLAN localiza-
tion,’’ IEEE Sensors J., vol. 17, no. 21, pp. 7086–7095, Nov. 2017.
[26] A. Antoniou, A. Storkey, and H. Edwards, ‘‘Data augmentation genera-
tive adversarial networks,’’ 2017, arXiv:1711.04340. [Online]. Available:
http://arxiv.org/abs/1711.04340
[27] C. Bowles, R. Gunn, A. Hammers, and D. Rueckert, ‘‘GANsfer WAFA NJIMA (Member, IEEE) was born in
learning: Combining labelled and unlabelled data for GAN based Tunisia, in 1991. She received the degree in engi-
data augmentation,’’ 2018, arXiv:1811.10669. [Online]. Available:
neering from the Institut National des Sciences
http://arxiv.org/abs/1811.10669
Appliquées et de Technologies, Tunis, in 2015, and
[28] A. Brock, J. Donahue, and K. Simonyan, ‘‘Large scale GAN training for
the Ph.D. degree in radio-communications from
high fidelity natural image synthesis,’’ 2018, arXiv:1809.11096. [Online].
Available: http://arxiv.org/abs/1809.11096 the Conservatoire National des Arts et Métiers,
[29] A. Odena, ‘‘Semi-supervised learning with generative adversarial Paris, in 2019, in collaboration with the Ecole
networks,’’ 2016, arXiv:1606.01583. [Online]. Available: http://arxiv. Supérieure des Communications de Tunis, under
org/abs/1606.01583 the supervision of Prof. Michel Terré and Prof.
[30] C. Bowles, L. Chen, R. Guerrero, P. Bentley, R. Gunn, A. Hammers, Ridha Bouallegue. She is currently a Postdoctoral
D. A. Dickie, M. V. Hernández, J. Wardlaw, and D. Rueckert, Researcher with CNRS, ETIS UMR 8051, ENSEA, CY Cergy Paris Uni-
‘‘GAN augmentation: Augmenting training data using generative versity, France. Her publications span several research areas, including sig-
adversarial networks,’’ 2018, arXiv:1810.10863. [Online]. Available: nal processing, wireless communications, sparse data, indoor localization,
http://arxiv.org/abs/1810.10863 the IoT, and machine learning for communications.

98346 VOLUME 9, 2021

W. Njima et al.: Indoor Localization Using Data Augmentation via Selective GANs

MARWA CHAFII (Member, IEEE) received the RAED M. SHUBAIR (Senior Member, IEEE)
master’s degree in advanced wireless communica- received the Ph.D. degree in electrical engineering
tion systems (SAR) and the Ph.D. degree in elec- from the University of Waterloo, Canada, in 1993.
trical engineering from CentraleSupélec, France, From 1993 to 2017, he was a Full Professor of elec-
in 2013 and 2016, respectively. From 2014 to trical engineering with Khalifa University, United
2016, she has been a Visiting Researcher with the Arab Emirates. He is currently a Full Professor
Poznan University of Technology, Poland; the Uni- of electrical engineering affiliated to New York
versity of York, U.K.; Yokohama National Univer- University (NYU) Abu Dhabi. His current and past
sity, Japan; and the University of Oxford, U.K. She academic and research appointments also include
joined the Vodafone Chair Mobile Communication the Massachusetts Institute of Technology (MIT),
Systems, Technical University of Dresden, Germany, in February 2018, as a Harvard University, and the University of Waterloo. His publication span
Research Group Leader. Since September 2018, she has been an Associate several research areas, including 6G and terahertz communications, mod-
Professor with ENSEA, France, where she holds the Chair of Excellence ern antennas and applied electromagnetics, signal and array processing,
on artificial intelligence from CY Initiative. Her research interests include machine learning, the IoT and sensor localization, medical sensing, and
advanced waveform design, machine learning for wireless communications, nano-biomedicine. He is a fellow of the MIT Electromagnetics Academy, and
and indoor localization. She received the prize of the Best Ph.D. in France a Founding Member of MIT Scholars of the Emirates and five IEEE society
in the fields of signal, image and vision, and she has been nominated in chapters, United Arab Emirates. He received the University of Waterloo
the top ten Rising Stars in Computer Networking and Communications by Distinguished Doctorate Dissertation Award for his Ph.D. degree. He was a
N2Women, in 2020. She is currently the Vice-Chair of the IEEE ComSoc ETI recipient of several international awards, including the Distinguished Service
on Machine Learning for Communications, leading the Education Working Award from ACES Society, USA, and the MIT Electromagnetics Academy,
Group, ETI on Integrated Sensing and Communications, a Research Lead USA. He is also a standing member of the editorial boards of several
with Women in AI, and managing the Gender Committee of the AI4EU international journals and serves regularly on the steering, organizing, and
Community. Since 2019, she has been serving as an Associate Editor for technical committees of IEEE flagship conferences in antennas, communi-
IEEE COMMUNICATIONS LETTERS, where she received the Best Editor Award, cations, and signal processing, including IEEE AP-S/URSI, EuCAP, IEEE
in 2020. GloablSIP, IEEE WCNC, and IEEE ICASSP. He is also a Board Member of
the European School of Antennas and the Regional Director for the IEEE
Signal Processing Society in IEEE Region 8 Middle East, and served as the
Founding Chair of the IEEE Antennas and Propagation Society Educational
Initiatives Program. He is also the Founder and the Chair of IEEE at NYUAD.
He is also an Officer of IEEE ComSoc Emerging Technical Initiative on
Machine Learning for Communications. He is also the Founding Director
of the IEEE UAE Distinguished Seminar Series Program for which he was
selected to receive, along with Mohamed AlHajri of MIT, the 2020 IEEE
UAE Award of the Year. He is also an Editor of the IEEE JOURNAL OF
ELECTROMAGNETICS, RF AND MICROWAVES IN MEDICINE AND BIOLOGY, and the
IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION.

H. VINCENT POOR (Life Fellow, IEEE)

received the Ph.D. degree in EECS from Princeton
University, in 1977. From 1977 to 1990, he was
with the Faculty of the University of Illinois at
Urbana–Champaign. Since 1990, he has been with
the Faculty at Princeton University, where he is
ARSENIA (ERSI) CHORTI (Senior Member, currently the Michael Henry Strater University
IEEE) is currently a Professor with the Ecole Professor. From 2006 to 2016, he served as the
Nationale Supérieure de l’Electronique et des ses Dean of the Princeton’s School of Engineering and
Applications (ENSEA); the Head of the Informa- Applied Science. He has also held visiting appoint-
tion, Communications and Imaging (ICI) Group, ments at several other universities, including most recently at Berkeley
ETIS Lab UMR 8051; and a Visiting Research and Cambridge. His research interests include information theory, machine
Fellow with Princeton University and the Univer- learning, and network science, and their applications in wireless networks,
sity of Essex. Her research interests include wire- energy systems, and related fields. Among his publications in these areas is
less communications and wireless system secu- the forthcoming book on machine learning and wireless communications
rity for 5G and 6G, with a particular focus on (Cambridge University Press, 2021). He is a member of the National
physical layer security, context aware security, multi-factor authentication Academy of Engineering and the National Academy of Sciences, and a
protocols, 5G/6G and the IoT, anomaly detection, machine learning for Foreign Member of the Chinese Academy of Sciences, the Royal Society,
communications, non-orthogonal multiple access (NOMA), and faster than and other national and international academies. Recent recognition of his
Nyquist signaling. She is a member of IEEE INGR on Security and the work includes the 2017 IEEE Alexander Graham Bell Medal and the D.Eng.
IEEE P1951.1 Standardization Workgroup (Smart Cities). She has been on Honoris Causa from the University of Waterloo awarded, in 2019.
a semi-sabbatical leave (CNRS delegation), since March 2021.

VOLUME 9, 2021 98347