DeepWiFi Cognitive WiFi With Deep Learning

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 20, NO.
2, FEBRUARY 2021 429
DeepWiFi: Cognitive WiFi with Deep Learning

Kemal Davaslioglu , Sohraab Soltani, Tugba Erpek , and Yalin E. Sagduyu
Abstract—We present the DeepWiFi protocol, which hardens the baseline WiFi (IEEE 802.11ac) with deep learning and sustains high
throughput by mitigating out-of-network interference. DeepWiFi is interoperable with baseline WiFi and builds upon the existing WiFi’s
PHY transceiver chain without changing the MAC frame format. Users run DeepWiFi for: i) RF front end processing; ii) spectrum
sensing and signal classification; iii) signal authentication; iv) channel selection and access; v) power control; vi) modulation and coding
scheme (MCS) adaptation; and vii) routing. DeepWiFi mitigates the effects of probabilistic, sensing-based, and adaptive jammers. RF
front end processing applies a deep learning-based autoencoder to extract spectrum-representative features. Then a deep neural
network is trained to classify waveforms reliably as idle, WiFi, or jammer. Utilizing channel labels, users effectively access idle or
jammed channels, while avoiding interference with legitimate WiFi transmissions (authenticated by machine learning-based RF
fingerprinting) resulting in higher throughput. Users optimize their transmit power for low probability of intercept/detection and their
MCS to maximize link rates used by backpressure algorithm for routing. Supported by embedded platform implementation, DeepWiFi
provides major throughput gains compared to baseline WiFi and another jamming-resistant protocol, especially when channels are
likely to be jammed and the signal-to-interference-plus-noise-ratio is low.
Index Terms—WiFi, machine learning, deep learning, dynamic spectrum access, RF signal processing, signal classification,
signal authentication
1 INTRODUCTION
radios provide wireless communication sys- jammers. DeepWiFi leverages the advances in machine
C OGNITIVE
tems with the capability to perceive, learn and adapt to
spectrum dynamics. The existing communication systems
learning algorithms supported by emerging computational
capabilities. The scope of DeepWiFi is to show how these
such as WiFi can greatly benefit from design concepts of algorithms can be used to enhance the performance and
cognitive radio that cover various detection, classification, security in multi-hop mobile ad hoc networks (MANETs).
and prediction tasks. These cognitive radio tasks can be Potential application areas of DeepWiFi include internet-of-
potentially performed by machine learning that adapt with things (IoT), sensors, and first responder communications.
spectrum data without being explicitly programmed [1], [2], One key aspect of DeepWiFi is to support interoperability
[3], [4]. In particular, deep neural networks have strong poten- with baseline WiFi while refraining from any detrimental
tial to process and analyze rich spectrum data. Examples effect on baseline WiFi operations. For that purpose, Deep-
of deep learning applications for cognitive radio tasks WiFi uses the PHY transceiver chain of baseline WiFi and
include modulation classification with convolutional neural does not change its medium access control (MAC) frame for-
networks (CNNs) [5], spectrum sensing with CNNs [6] mat. The DeepWiFi protocol stack involves the physical
and generative adversarial networks (GANs) [7], and anti- (PHY), link/MAC, and network layer algorithms that oper-
jamming and power control with feedforward neural net- ate in a distributed manner. DeepWiFi supports multi-hop
works (FNNs) [8], [9]. routing compared to the standard use of IEEE 802.11ac that
In this paper, we present the DeepWiFi protocol that aims relies on access points to connect individual nodes. As
to harden the baseline IEEE 802.11ac WiFi with the appli- an extension, WiFi Direct can be utilized to support multi-
cation of machine learning (in particular, deep learning) hop communications [10] and the IEEE 802.11s amendment
to improve the throughput performance and enhance is specifically designed for mesh networking that incorpo-
its security in the presence of jammers causing out-of- rates multi-hop communications. While this paper presents
network interference. We consider three types of jammers that our findings using a WiFi setting, algorithms of DeepWiFi
are modeled as probabilistic, sensing-based, or adaptive can be extended to other frequency bands and waveforms
as well.
K. Davaslioglu, S. Soltani, and Y. E. Sagduyu are with Intelligent 1.1 Summary of Contributions
Automation Inc., Rockville, MD 20855 USA.
E-mail: {kdavaslioglu, ssoltani, ysagduyu}@i-a-i.com.
DeepWiFi consists of seven steps: 1) RF front end process-
T. Erpek is with Intelligent Automation Inc., Rockville, MD 20855 USA ing; 2) spectrum sensing and signal classification; 3) signal
and also with the Department of Electrical and Computer Engineering, Vir- authentication; 4) channel selection and access; 5) power
ginia Polytechnic Institute and State University, Arlington, VA 22203 USA. control for Low Probability of Intercept and low probability
E-mail: [email protected].
of detection (LPI/LPD); 6) modulation and coding scheme
Manuscript received 5 July 2018; revised 1 Oct. 2019; accepted 20 Oct. 2019. (MCS) adaptation; and 7) routing. The main contributions
Date of publication 28 Oct. 2019; date of current version 7 Jan. 2021.
(Corresponding author: Kemal Davaslioglu.) of this paper are the algorithms in steps 1-4, namely RF front
Digital Object Identifier no. 10.1109/TMC.2019.2949815 end processing; spectrum sensing and signal classification;
1536-1233 ß 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.
430 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 20, NO. 2, FEBRUARY 2021
signal authentication; and channel selection and access. quickly to zero, whereas DeepWiFi can reliably identify
Steps 5-7 on power control, adaptive modulation and cod- channels for transmission and sustain its throughput. For 9
ing, and routing are used in other similar protocols and users operating over 40 channels, we show that there is no
included to provide the remaining functionalities (preserv- throughput loss for DeepWiFi when jamming likelihood is
ing the layered architecture of the protocol stack), imple- less than 60 percent and there is only 14 percent throughput
ment and assess the full protocol stack, and compare its loss when jamming likelihood is 80 percent. Finally, we
performance with other protocols. While executing these show that DeepWiFi outperforms a jamming-resistant MAC
steps, DeepWiFi controls the selection of transmit power, protocol called Jamming Defense (JADE) [12] and is robust
MCS ID, authenticated signal, channel ID, neighbor ID and against adaptive jammers.
flow ID, and provides the selections to the IEEE 802.11ac
transceiver chain.
There are three tasks that are enhanced by the use of 1.2 Related Work
machine learning in the DeepWiFi protocol. Jamming and anti-jamming mechanisms with machine
learning are studied in [8], [9] in which an adversarial user
1) RF front end processing: Each user applies autoencoding builds a deep learning based classifier to predict if there
to the in-phase and quadrature (I/Q) components of will be a successful transmission (replied with ACK mes-
the data collected on sensed channels and extracts sages) or not (no ACK) using an exploratory (inference)
spectrum features. The I/Q data can be collected by attack to jam data transmissions. A transmitter in return can
using additional receivers or extracted by using firm- launch a causative attack to poison jammer’s classifier (by
ware dispatches on existing WiFi chips such as dis- adding controlled perturbations to transmit decisions) as a
cussed in [11]. We show that autoencoding achieves defense mechanism. Similar to [8], [9], DeepWiFi considers
low reconstruction loss (0.2 percent) in dimensionality users with jamming capabilities. The difference is that
reduction compared to other methods such as Prin- DeepWiFi identifies jammed channels and enables trans-
cipal Component Analysis (PCA) and t-distributed missions (by adapting the MCS) even under jamming
Stochastic Neighbor Embedding (t-SNE). The outputs instead of backing off (which would be the case in the base-
of the RF front end processing are input to the signal line WiFi). There is also a rich literature on modulation rec-
classification module. ognition studies that distinguish signals with different
2) Spectrum sensing and signal classification: Each user modulation schemes. These related studies include earlier
applies deep learning (FNN or CNN) to the features work in [13], [14] that use cyclic spectrum characteristics
obtained by RF front end processing in order to clas- extracted from raw signal and more recent efforts in [5], [15]
sify each channel as idle (I), jammed (J) or used by that use the raw I/Q samples as inputs using CNN. Since a
another WiFi device (W). We show that both FNN frame may consist of multiple modulations (in the preamble
and CNN achieve over 98 percent accuracy, whereas and data portions) and modulations may change across
the accuracy of Support Vector Machine (SVM) is short MAC frames (5.484 msec in IEEE 802.11ac [16]), the
limited to 66 percent. problem in DeepWiFi is different from [5], [15] and focused
3) Signal authentication: Each user applies machine on classifying waveforms (chunked into frames) within the
learning-based RF fingerprinting to authenticate sig- frame time rather than simple modulations that remain
nals. Radio hardware effects on RF signals are used as fixed over longer periods of time. Note that overall process-
features to detect outliers. We show that the detec- ing overhead for the front end processing, signal classifica-
tion accuracy by Minimum Covariance Determinant tion, and signal authentication of DeepWiFi is measured as
(MCD) is close to 90 percent, whereas SVM and Isola- 0.1546 msec (2.8 percent of 802.11ac frame) to process each
tion Forest can only achieve around 69 and 70 percent sample. Furthermore, we note that the conventional energy
accuracy, respectively. This signal authentication capa- detectors for spectrum sensing cannot be applied in our
bility of DeepWiFi protects WiFi against replay attacks. framework because they cannot distinguish busy and jam-
4) Channel access: As opposed to baseline WiFi (where a ming signals such that an adversary can easily create a jam-
user backs off regardless of the type of interference), ming signal with the same energy characteristics to fool an
each user in DeepWiFi runs the signal classification, energy detector.
backs off when the interference is from another user There are several characteristics that can be used for RF
participating in DeepWiFi protocol, or does not back fingerprinting. One can focus on transient [17] or the
off when the interference is from a jammer. This steady-state [18], [19] behavior of the transmitted signal. For
way, DeepWiFi users continue communications (in a example, [17] studied the unique transient characteristics of
degraded mode) and still achieve a non-zero through- a transmitter, which are often attributed to RF amplifier, fre-
put (instead of backing off) through adaptation of quency synthesizer modules, and modulator subsystems.
power, modulation, and coding scheme. The duration of the transient behavior changes depending
In this paper, a distributed network of users running on the type and model of the transmitter. Also, as a device
DeepWiFi is simulated in the presence of probabilistic and ages, its transient behavior can change [17]. In DeepWiFi,
sensing-based jammers. MATLAB WLAN Toolbox is used we opt not to use transient behavior as a feature since it is
to generate realistic WiFi signals and channels. Simulation highly susceptible to noise and interference effects and
results show that as the jamming effect (either intentional requires precise timing of when the signal transmission has
adversarial jamming or in the form of out-of-network inter- started. Instead, DeepWiFi uses steady-state characteristics
ference) increases, the throughput of baseline WiFi drops between a transmitter and a receiver pair such as frequency
DAVASLIOGLU ET AL.: DEEPWIFI: COGNITIVE WIFI WITH DEEP LEARNING 431
and timing synchronization offset, both of which are often The rest of the paper is organized as follows. Section 2
attributed to the radio’s oscillator, power amplifier, and describes the system model. The DeepWiFi protocol over-
digital-to-analog converter. There are two relevant studies view is given in Section 3. RF front end processing of Deep-
that use steady-state characteristics similar to this paper. WiFi is described in Section 4.1. This is followed by the
First, in [18], three instantaneous signal characteristics such description of spectrum sensing and signal classification
as the amplitude, phase, and frequency of the signal are of DeepWiFi in Section 4.2. RF fingerprinting is applied in
used. After a preprocessing step (removal of the mean and Section 4.3 to authenticate signals in DeepWiFi. Section 4.4
normalization), their features (such as the variance, skew- describes channel selection and channel access of DeepWiFi.
ness, and kurtosis) are extracted within predefined window It is followed by the descriptions of power control for LPI/
sizes [18]. Second, the frequency offset, magnitude and LPD and MCS adaptation of DeepWiFi in Sections 4.5 and
phase offset, distance vector, and I/Q origin offset are 4.6, respectively. Section 4.7 presents the routing extension
used in [19] as features for RF fingerprinting, and weighted of DeepWiFi. Simulation setting is described in Section 5.
voting classifiers and maximum likelihood classifier are Performance results of DeepWiFi and baseline WiFi are pre-
developed for fingerprinting. In DeepWiFi, we use machine sented in Section 6. The implementation, overhead, and
learning for signal authentication purposes. complexity aspects of DeepWiFi are discussed in Section 7.
Ad hoc networking over WiFi has two standards, Section 8 concludes the paper.
namely, the ad hoc mode in 802.11 and its successor
WiFi Direct. The ad hoc mode (IBSS) is one of the two
modes of operation in the 802.11 standard, the other 2 SYSTEM MODEL
mode being the commonly used mode of infrastructure- 2.1 Network Setting
based mode (IM-BSS). The IBSS mode enables direct A wireless multi-hop network of n users is considered. Each
communication between any devices without the need user may act as the source, the destination, or the relay for
for an access point (AP). This standard has been evolved packet traffic. Each source i generates unicast packets for traf-
to what is known as WiFi Direct that is considered to be fic flow sij , addressed to destination j with rate rij . While
the main standard of ad hoc networking over WiFi, par- DeepWiFi algorithms do not require synchronization, we con-
ticularly, using Android smartphones [20]. Google has sider slotted time in discrete time simulation and follow the
developed a peer-to-peer (P2P) framework, WiFi P2P, 802.11ac MAC’s synchronization that uses the synchronizing
that complies with WiFi Direct standard. Devices after timers in the preamble of a frame (Non-HT Training Field
Android 4.0 (API level 14) can discover other devices [16]). Each user i holds a separate queue (with length Qsi ðtÞ at
and connect directly to each other via WiFi without an time slot t) for every packet flow s. Packets in each queue are
intermediate AP, when both devices support WiFi P2P served in a first-come-first-served (FCFS) fashion. Packets are
[21]. The APIs include calls for peer discovery, request, allocated to 802.11ac MAC frames before transmission. If the
and connection, all of which are defined in the WifiP2P- queue length is smaller than the frame, zero padding is used
Manager class (see [21] for more details). A related work for the missing data part. If the queue length is greater than
that uses WiFi direct in ad hoc mode and implements the frame, packets in the queue are partitioned into multiple
multi-hop routing is [10], which improves multi-group frames before transmission. There are m channels available.
link connectivity with low overhead. Note that WiFi At any given time slot, a user may select one of these channels
Direct, as defined in the standard, does not support to transmit packets. Users share these channels for data trans-
multi-hop communications, but recent work such as [10] missions and control information exchanges. There is no cen-
has shown that multi-hop routing can be implemented tralized controller. Users make their own decisions in a
using WiFi Direct. With the four novel contributions of Deep- distributed setting. There are nJ jammers that aim to interfere
WiFi focused on PHY and MAC layers (as highlighted in with transmissions. Without loss of generality, each jammer is
Section 2), DeepWiFi is designed to support both single-hop assigned to one of the channels (i.e., nJ ¼ m). There are three
and multi-hop communications. The routing approach in types of jammers:
DeepWiFi differs from [10] such that routing based on the
backpressure algorithm uses both queue and channel Probabilistic (random) jammer: The jammer is turned
information. on with fixed jamming probability pJ at any given
The RF preprocessing step in DeepWiFi utilizes autoen- time slot.
coders [22], [23]. In the literature, there are several studies that Static sensing-based jammer: The jammer is turned on
use autoencoders to extract features in a high dimensional if it detects a signal on the channel using a constant
space and provide a feature representation of the data, and sensing threshold t, else it is turned on with the
then train a separate classifier using these features (e.g., [24]). probability of jamming pJ at any given time slot.
In this paper, we also use the same semi-supervised approach. Adaptive jammer: This jammer is a sensing-based jam-
Other prepossessing methods such as Short Time Fourier mer where channel sensing threshold is adaptively
Transformation, Choi-Williams Transformation, and Gabor adjusted. If the jammer senses a signal on the chan-
Wavelet Transformation have been used in [25] to preprocess nel, it is turned on.
I/Q data before running them through a CNN. In this paper, Both the static and adaptive jammers use channel sensing
our preprocessing step is data-driven and consists of a denois- measurements to make jamming decisions. The sensing
ing autoencoder that is shown to suppress noise in the input threshold is constant for the static sensing-based jammer,
to the deep learning classifier and results in a small recon- while the adaptive jammer changes the threshold by observ-
struction loss. ing the actions of the transmitter in a given time window.
TABLE 1
Delay Models
Model
Parameter A B C D E F
Breakpoint distance (m) 5 5 5 10 20 30
RMS delay spread (ns) 0 15 30 50 100 150
Maximum delay (ns) 0 80 200 390 730 1050
Rician K-factor (dB) 0 0 0 3 6 6
Number of taps 1 9 14 18 18 18
Number of clusters 1 2 2 3 4 6
IEEE 802.11ac standard is followed for the PHY imple-

mentation and MAC frame format. The center frequency is
5.25 GHz. The instantaneous bandwidth is 20 MHz and
each OFDM subcarrier occupies a bandwidth of 312.5 kHz.
The subcarriers can be used independently and incoming
data bits can be distributed among the subcarriers (possibly
depending on the channel frequency response). Thus, a Fig. 1. DeepWiFi protocol diagram with seven steps.
20 MHz channel consists of 64 subcarriers. We allocate 8
subcarriers for pilot, 16 subcarriers as null, and 40 subcar-
riers for data (i.e., m ¼ 40). 3 OVERVIEW OF THE DEEP WIFI PROTOCOL
Each user selects its MCS depending on its received On top of 802.11ac, each user individually and asynchro-
SINR. The link rate from user i to user j at time slot t is nously runs the DeepWiFi protocol stack with the following
denoted by cij ðtÞ and depends on the selected MCS, trans- seven main steps (shown in Fig. 1):
mit power, and consequently the observed SINR. On each
channel, the SINR at a user depends on the channel and 1) RF front end processing: Each user hops among WiFi
interference effects among users (legitimate transmitters channels (one by one), collects RF signal on each
and jammers). Next, we describe how the channels among sensed channel and processes the RF signal at the RF
users and their signals are generated. front end to build the I/Q data and extract features.
2) Spectrum sensing and signal classification: Each user
2.2 Channel and Waveform Data Generation applies deep learning to these features (the output of
We generate the physical signals and channels using the the RF front end) in order to classify (label) each
802.11ac library in MATLAB WLAN System Toolbox. Sev- channel as idle (I), jammed (J) or used by another
eral channel models are offered. Each channel model con- WiFi device (W).
siders a different breakpoint distance, root mean square 3) Signal authentication: Each user applies machine
(RMS) delay spread, maximum delay, Rician K-factor for learning-based RF fingerprinting to authenticate
non-light of sight (NLOS) conditions, number of taps, and legitimate WiFi signals at the physical layer.
number of clusters. Waveform related parameters are speci- 4) Channel selection and channel access: Each user backs off
fied to generate the signal such as the input data bits, num- on any busy channel used by legitimate WiFi signals
ber of packets, packet format (VHT, HT, etc.), idle time (W) (to resolve future conflicts), selects an idle channel
(added after each packet), scrambler initialization, the num- (I); if none, selects a jammed channel (J) (including
ber of transmit antennas, payload length, MCS, and band- channels used by the non-legitimated WiFi signals)
width using the wlanWaveformGenerator function of with the best SINR for data transmission. The use of
the MATLAB WLAN System Toolbox. This function gener- jammed channels (when no idle channel is available)
ates a time-domain I/Q waveform x that goes through the corresponds to the degraded mode, where a non-zero
channel, where small scale fading and other channel impair- throughput can be still achieved.
ments are applied. The MATLAB function used for this 5) Power control for LPI/LPD: Each user selects the trans-
operation is wlanTGacChannel and it performs an Hx x mit power below the jammer threshold level to avoid
operation, where H is the channel matrix. There are several detection by jammers and achieve LPI/LPD.
parameters that can be set during the channel generation 6) Adaptive modulation and coding: There are nine possible
such as delay profile (models A-F are shown in Table 1), MCS options to choose from in 802.11ac. Each user
sample rate, carrier frequency, transmission direction selects the best MCS based on the measured SINR to
(uplink or downlink), the number of receive and transmit maximize the achievable rate on the selected channel.
antennas, spacing between antennas, and large-scale fading 7) Routing: Each user makes the routing decision by
effects (‘None’, ‘Pathloss’, ‘Shadowing’, or ‘Pathloss and selecting the flow to serve and the next hop for trans-
Shadowing’). Next, additive white Gaussian noise (AWGN) mission by applying the backpressure algorithm, which
is added to the signal that is passed through the channel optimizes a spectrum utility depending on traffic
using the awgn command of the MATLAB WLAN System congestion and link rate (computed in Step 6). Lower
Toolbox. Thus, we obtain the channel effect Hx x þ n induced layers of DeepWiFi are transparent to the routing
on the I/Q waveform x . algorithm and can be combined with other efforts
classifies the input signal as WiFi (W), Jammer (J), or

Idle (I).
2) A logic takes the signal classification result as input.
If the signal is a WiFi signal, it passes the I/Q data to
the physical layer authentication (RF fingerprinting)
module, which constitutes the first step of the physi-
cal layer security.
Instead of using I/Q data directly as input to the signal clas-
sifier, we use a denoising autoencoder to extract the features of
the received signal that are then fed into the signal classifier. In
an RF environment, we typically have many unlabeled data
samples (that can be collected by sensing the spectrum) while
the number of labeled data samples (needed to train the classi-
Fig. 2. RF front end processing of DeepWiFi (Step 1) and its connections fier) is relatively small compared to the dimension of each data
to Steps 2 and 3.
sample (40K is the dimension per sample in our case). How-
ever, unlabeled data can be used to train an autoencoder that
for multi-hop networking such as the extension of reduces the dimension of the input data to the classifier to build
WiFi Direct for multi-group networking [26]. a reliable classifier from few labeled data samples. In addition,
The output of the DeepWiFi protocol is specified as: we use a denoising autoencoder that further suppresses the
noise in the input. Another benefit of the autoencoder may be
a. TX power specifies the transmit power.
observed when the environment changes (e.g., new channel
b. MCS ID specifies which MCS of IEEE 802.11ac is
conditions (distributions) or unknown signals passing through
used and its corresponding rate.
known or unknown channel distributions). An autoencoder
c. Authenticated signal specifies which signals belong to
trained with unlabeled data in the new environment may help
legitimate WiFi.
adapt to the effects of this new environment by robust feature
d. Channel ID specifies which channel to send the next
extraction. Then, a small number of labeled data samples in the
data packet.
new environment could be sufficient. These capabilities would
e. Neighbor ID specifies which neighbor to select for the
not be achievable if we were to use the I/Q data directly as
next hop transmission.
input to the deep neural network-based classifier as we cannot
f. Flow ID specifies which application traffic flow to
train a reliable classifier for a high-dimensional problem using
serve when the data packet is transmitted.
a small number of labeled training samples.
These outputs tune parameters of the 802.11ac network
To preprocess the data, we also tried different standard
protocol stack at each user. Specifically, outputs a-c tune the
techniques such as Principal Component Analysis (PCA)
physical layer, outputs d and e tune the link/MAC layer, and
[27] and t-distributed stochastic neighbor embedding (tSNE)
outputs f and g tune the network layer, as shown in Fig. 1.
[28]. However, they failed in separating signals of interest.
These results are reported in the Appendix, which is avail-
4 DEEPWIFI PROTOCOL STEPS able in the IEEE Computer Society Digital Library at http://
doi.ieeecomputersociety.org/10.1109/TMC.2019.2949815.
4.1 Step 1: RF Front End Processing Instead, DeepWiFi uses an autoencoder to extract features
Input: The RF signal. from the I/Q data. An autoencoder is a deep neural network
Output: The set of extracted features given to Step 2 and that is trained to reconstruct its input and consists of two
the I/Q data given to Step 3. neural networks, namely, an encoder h ¼ fu ðx xÞ and a decoder
As shown in Fig. 2, the following steps are pursued to that produces a reconstruction r ¼ gF ðh hÞ, where u is the set
process the RF signal. of weights and biases of the neural network corresponding
1) 16-bit analog-to-digital converter (ADC) is used to to the encoder and F represents that of the decoder. The
sample the signal by allocating 8 bits for real and neural networks f and g can be constructed as FNN or
imaginary parts of the signal. CNN. Autoencoder is used for unsupervised learning of effi-
2) The digitized signal is bandpass filtered with 20 MHz cient codings. DeepWiFi uses the autoencoder to learn a
instantaneous bandwidth to remove interference representation (encoding) for a set of data, for the purpose
from adjacent bands. of feature learning and dimensionality reduction. In particu-
3) The digitized and bandpass-filtered signal is sam- lar, DeepWiFi uses a denoising autoencoder that adds noise to
pled at 40 MHz. its inputs and trains it to recover the original noise-free
4) The deep-learning based autoencoder takes the input inputs. This method prevents the autoencoder from trivially
samples and reduces them into latent features. copying its inputs to its outputs and the autoencoder finds
RF front end processing provides the I/Q data to Step 3 the patterns in the data, while avoiding overfitting.
(signal authentication) and the reduced set of features to The preprocessing of DeepWiFi (ADC, bandpass filtering,
Step 2 (spectrum sensing and signal classification). The and sampling) produces the I/Q data that has the dimension
details of this procedure (shown in Fig. 1) are given below: of 40000 (20000 for I and 20000 for Q components) for each
time instant in Step 1. DeepWiFi applies denoising autoen-
1) Spectrum sensing and signal classification (Step 2) coder to this I/Q data, and determines the latent features that
takes the outputs of the autoencoder in Step 1 and are further fed to the signal classifier of DeepWiFi.
Fig. 3. Denoising autoencoder in DeepWiFi.
The denoising autoencoder of DeepWiFi adds the Gauss-

ian noise to the I/Q data (to prevent overfitting) and then
applies four hidden layers (the first two layers for encoding
and the last two layers for decoding) after one initial nor-
malization layer. Hidden layers are trained through the Fig. 4. Training and test losses during denoising autoencoder training.
backpropagation algorithm to minimize the minimum squared
error (MSE) as the loss function. We used the hyperbolic thoroughly. Note that we use Hyperband to optimize the
tangent function (tanh) as the activation function that per- parameters of neural networks described in Sections 4.1 and
forms fðxÞ ¼ tanhðxÞ operation. In a neural network, the 4.2. We considered the maximum amount of resources that
activation function is used as an abstraction representing can be allocated to a single configuration as 81 iterations and
the rate of action potential firing in the cell. We performed a downsampling rate of three (h ¼ 3). The dataset used in the
hyperparameter optimization and observed that a Gaussian training of the denoising autoencoder consists of equal num-
noise N with zero mean and variance of 0.1 gives the best ber of WiFi, jammer, and noise signals. The dataset is divided
reconstruction loss and avoids overfitting. DeepWiFi uses into 80 and 20 percent splits for the training and test sets,
the following autoencoder structure: respectively. We use a batch size of 64 samples. Fig. 4 depicts
the loss in the training and test sets during the training pro-
Hidden layer 1: FNN with 534 neurons and tanh cess. We observe that the loss gradually decreases in both
activation. sets and no overfitting is observed.
Hidden layer 2: FNN with 66 neurons and tanh Fig. 5 illustrates the legitimate WiFi signal in the test set
activation. and its reconstruction. We observe the noise suppression on
Hidden layer 3: FNN with 534 neurons and tanh the sidebands up to 10 dB. This highlights the effect of
activation. denoising autoencoder on the noise reduction. Note that a
Hidden layer 4: FNN with 40000 neurons and tanh similar denoising effect using denoising autoencoders has
activation. been reported in the image processing [30] and Fig. 5 shows
The input and output layers have the same dimension. the same effect in the RF domain.
A denoising autoencoder adds a Gaussian noise N with The output of the autoencoder is the set of extracted fea-
0 mean and variance of s 2 to the input data samples X. tures that are given to the classifier in Step 2.
The resulting data Xnoisy ¼ X þ N is input to the neural
network. Denoising autoencoder solves the following loss
4.2 Step 2: Spectrum Sensing
function:
and Signal Classification
Input: The I/Q data on the sensed channel from Step 1.
ðF ; u Þ ¼ min EX ½jjgF ðfu ðXnoisy ÞÞ Xjj2 ; (1)
ðF;uÞ Output: The classification of the channel to idle, used by
WiFi or jammed, given as input to Step 3.
where EX ½ denotes the expectation over X and jj jj2 is the Each user applies a deep neural network-based classifier
‘2 -norm (the euclidean norm). The structure of denoising to the received I/Q data on the sensed channel and classifies
autoencoder is shown in Fig. 3.
We implemented a denoising autoencoder using the Ten-
sorFlow framework, which takes the input I/Q data and
reduces its dimensions. These features represent the latent
variables (extracted features). The reconstruction loss
between the original signal X and the reconstructed signal X ^
is computed by E½jjX Xjj^ 2 . To find the best parameters,
we performed a hyperparameter optimization using the
Hyperband framework [29]. This framework can provide
over an order-of-magnitude speedup over Bayesian hyper-
parameter optimization methods. The Hyperband randomly
samples a set of hyperparameter configurations, evaluates
the performances of all current configurations, sorts the score
of configurations, and removes the worst scoring configura-
tions (successive halving). The process is repeated progres-
sively with increasing number of iterations. Therefore, only
the configurations that yield good results are evaluated Fig. 5. Reconstruction of a WiFi signal in the test set.
the captured signal as noise (idle), WiFi signal, or jammer through the channel. Then, we add white Gaussian
signal. Features are the I/Q data received over time (output noise. The output is stored in the training data as the
of Step 1). There are three potential labels assigned to jammer signal samples with label (J).
each channel: idle (I), another WiFi device (W), and jammed 12000 samples are generated using different channel mod-
(J). Training data is collected offline and training is per- els. 80 percent of data is used for training and 20 percent is
formed offline. The trained classifier is loaded to each allocated to testing. Training data set includes signals gener-
user offline. Only neural network weights and biases are ated under six different channel conditions (2000 samples
stored in the memory. Each radio individually runs its per channel model).
classifier online.
4.2.2 Simple Machine Learning Classifiers
4.2.1 Signal Classes (I), (W), and (J)
For benchmark comparison, we evaluate the performance of
For training and testing phases, signals of each label are SVM classifier in signal classification. As its performance
generated as follows: depends on the kernel type and hyperparameters used, the
1) Noise: The background noise for WiFi is between hyperparameters are tuned for the best accuracy. Two of
80 dB and 100 dB. To emulate such a case, we the most common kernels, linear kernel and radial basis
choose a number uniformly at random between function (RBF) kernel, are used. For the RBF kernel, there
80 dB and 100 dB in the frequency domain and are two parameters that are considered, C and g. The
then take the inverse Fourier transform to obtain parameter C is common in all SVM kernel types and it
the time samples. The output is stored in the trades off misclassification of training examples against sim-
training data as the background noise samples with plicity of decision surface. When C is a small number, the
label (I). decision surface is smoother, and a high C puts more
2) WiFi Legitimate Signal: We use the MATLAB WLAN emphasis on classifying all training examples correctly. So
System Toolbox to generate the WiFi signal includ- C trades error penalty for stability. The second parameter g
ing the preamble and payload. We consider the fol- is the free parameter of Gaussian RBF, Kðxi ; xj Þ ¼ expðg
lowing parameters: jjxi xj jj2 ÞÞ where g > 0. It defines the influence of each
a) Center frequency: 5.25 GHz, training example. A larger g has smaller variance affecting
b) Channel Bandwidth: 20 MHz, only the closer neighbors. For hyperparameter tuning,
c) TX-RX distance: 5 m (default value; we vary the C ¼ ½0:0001; 0:001; 0:01; 0:1; 1:0; 10:0; 100:0; 1000:0 is used
SINR in simulations), for linear kernel, and g with the same C values for the
d) Normalize Path Gains: True, Gaussian RBF kernel. In the evaluation step, we implement
e) Transmit antenna spacing: 0.5 wavelength (default), k-fold cross-validation to calculate cross-validation accu-
f) Receive antenna spacing: 0.5 wavelength (default), racy and set k as 10. Using a grid search, we obtain the score
g) Packet format: VHT, of best performing model. The best performance is achieved
h) Scrambler initialization: 93 (default), using RBF kernel with C ¼ 1:0 and g ¼ 0:1, which achieves
i) Channel coding: BCC (binary convolutional only 66 percent accuracy.
coding),
j) APEP Length: 1, 4.2.3 Deep Learning Classifiers
k) PSDU Length: 36 (default; PSDU length is the We design two deep neural networks architectures, FNN
number of bytes carried in the user payload. For and CNN, for the signal classification task that achieves
a single user, the PSDU length is scalar integer small memory footprint, high accuracy, and low inference
from 1 to 220 1), time. These deep neural networks are implemented in Ten-
l) Long guard interval length: 800 ns, sorFlow using the Keras library [31]. Backpropagation is
m) Short guard interval length: 400 ns, applied to train the neural network using a cross entropy
n) A-MPDU Length: 256 bytes, P
loss function that is defined as L ¼ m i¼1 bi log ðyi Þ, where
o) Number of transmit antennas: 1, b ¼ fbi gm
i¼1 is a binary indicator of ground truth such that
p) Number of spatial streams transmitted: 1, and for a sample from label k, bk ¼ 1 whereas the other entries
q) MCS: Varying between 0-9. are all zeros. The neural network prediction is denoted by
y ¼ fyi gm
i¼1 . In both architectures, we used the ADAM opti-
We generate channel coefficients using the specific chan- mizer [32] with the learning rate of 105 .
nel model (as discussed in Section 2.2) and pass the WiFi
signal through the channel. Then, we add white Gaussian
noise. The output is stored in the training data as the WiFi 4.2.4 FNN
signal samples with label (W). We tune the hyperparameters of the FNN (such as the num-
3) Jammer Signal: In the time domain, we first generate ber of layers, number of neurons per layer and activation
normally distributed random numbers with zero functions). We use rectified linear unit (ReLU) as the activa-
mean and variance of one. Then we up-sample these tion function on neural network layer outputs that performs
samples to a selected carrier with a bandwidth that the fðxÞ ¼ maxð0; xÞ operation. The advantages of using
is the same as the WiFi signal. We generate channel ReLU activation function is that it is a bit faster to compute
coefficients using the same channel models dis- than other activation functions in hardware and its does not
cussed in Section 2.2 and pass the jammer signal suffer from vanishing gradient problem since it does not
TABLE 2 TABLE 3
Confusion Matrix for the FNN Confusion Matrix for the CNN
Predicted Label Predicted Label

(I) (W) (J) (I) (W) (J)
(I) 35.6% 0.0% 0.0% (I) 35.6% 0% 0%
True Label (W) 0.0% 32.4% 0.5% True Label (W) 0.1 % 31.6% 1.0%
(J) 0.1% 0.75% 30.75% (J) 0% 0.9 % 30.8%
4.3 Step 3: Signal Authentication

saturate for positive values, as opposed to the logistic func- Input: I/Q data from RF front end processing (Step 1) and
tion and hyperbolic tangent function [33]. The resulting channel label from spectrum sensing and signal classifica-
FNN consists of the following layers: tion (Step 2).
Fully connected layer with 15 neurons (ReLU activation). Output: Set of authorized signals (to be given to the WiFi
Dropout layer with 50 percent dropout probability. receiver chain).
Fully connected layer with 3 neurons (Softmax The goal of physical layer authentication (illustrated in
activation). Fig. 2) is to augment the standard WiFi security (at Layer 2)
It uses the dropout layer to drop out some neuron outputs at the physical layer by providing physical layer fingerprinting
from the previous layer of a neural network and serves the capability for processing I/Q data and authenticate legiti-
purpose of regularization for reducing overfitting by prevent- mate users. RF fingerprinting is motivated by mitigating
ing complex co-adaptations on training data. replay attacks in wireless networks. Recent open-source
PFor the output firmware patches [11] enable the capability to store and
layer, softmax activation function fi ðx
xÞ ¼ exi =ð j exj Þ is used at
the final layer of the neural network. Softmax activation func- transmit I/Q data in buffer. A software-defined radio (SDR)
tion is generally a good choice for classification tasks for which can also perform replay attacks by listening to the legitimate
output classes are mutually exclusive. FNN achieves the aver- communication between two parties and replaying the I/Q
age accuracy of 98.75 percent in predicting the correct signal data with the legitimate WiFi characteristics. In such a case,
labels. The confusion matrix for FNN is shown in Table 2. signal classification may not be enough for signal authenti-
cation, and RF characteristics in the signal transmitted by an
adversary can be used to authenticate the signal in the PHY
4.2.5 CNN layer. Note that this step uses the I/Q data directly, not the
We tune the hyperparameters of the CNN. The resulting features extracted by the autoencoder, as hardware impair-
CNN consists of the following layers: ments used in the authentication may be lost at the autoen-
coder. Therefore, I/Q data is directly used in this step. The
Five cascades of the following layers concatenated: details of this authentication procedure (shown in Fig. 2)
– A 2-D convolutional layer with 32 filters and ker- are given below:
nel sizes of (2,5).
– A 2-D convolutional layer with 32 filters and ker- 1) Spectrum sensing and signal classification (Step 2) takes
nel sizes of (2,5). the outputs of the autoencoder in Step 1 and classifies
– Max pooling layer with kernel size of (2,2) and the input signal as WiFi (W), Jammer (J), or Idle (I).
stride of (2,2). 2) A logic takes the signal classification result as input.
– Batch normalization layer. If the signal is a WiFi signal, it passes the I/Q data to
Fully connected layer with 18 neurons. the physical layer authentication (RF fingerprinting),
Dropout layer with a 50 percent dropout probability. which constitutes the first step of the physical layer
Fully connected layer with 3 neurons. security.
The 2-D convolutional layer is used to apply sliding filters 3) The physical layer authentication uses the physical
to the input. This layer convolves the input by moving layer impairments that are inherent in each transmit-
the filters along the input vertically and horizontally, com- ter and authorizes if it detects that the received sig-
puting the dot product of the weights and the input, and nal comes from a legitimate transmitter.
then adding a bias term. The max pooling layer is used to 4) If the received signal is authorized, the signal is proc-
progressively reduce the spatial size of the representation essed by the WiFi receiver chain.
to reduce the number of parameters and amount of compu- The objective is to analyze RF signal characteristics to
tation in the network, and hence to control overfitting. The authenticate legitimate WiFi users. One way of RF finger-
max pooling layer performs down-sampling by dividing printing is based on location identification by capturing
the input into rectangular pooling regions, and computing channel specific features (such as the Received Signal
the maximum of each region. The batch normalization layer Strength Indicator (RSSI) levels). However, this approach
is used to normalize each input channel across a mini- does not effectively apply to mobile users with rapidly
batch. The purpose is to speed up the training of CNN changing RSSI levels. Another way of RF fingerprinting is
and reduce the sensitivity to network initialization. CNN based on identification using radio characteristics that are
achieves the average accuracy of 98.0 percent in predicting divided into waveform domain and hardware domain char-
the signal labels. The confusion matrix for CNN is shown acteristics (impairments). The waveform domain approach
in Table 3. identifies transient-based behavior that lasts for a very short
Fig. 6. Two-layer approach for RF fingerprinting.
period of time (microseconds) and is hard to model. The

transient characteristics are also prone to noise and interfer-
ence effects. The hardware domain approach is based on
capturing hardware impairments such as frequency, magni-
tude and phase errors, I/Q offset and synchronization off-
sets in the time and frequency domains. We pursue this Fig. 7. WiFi signal with and without hardware impairments.
second approach.
To authenticate a signature, we use a two-layer approach where txSig and rxSig are the transmitted and received sig-
(as shown in Fig. 6): nals, respectively, kI and kQ are the in-phase and quadra-
ture gains, respectively, imbI and imbQ are the in-phase and
1) Outlier detection determines if the signature belongs quadrature imbalances, respectively, the amplitude imbal-
to any signature that is authenticated or not. If the ance c is given in dB, and phase imbalance f is given in
signature is not authenticated, it is rejected. If the sig- degrees. An example is shown in Fig. 7, where SRO is
nature belongs to an authenticated user, then we denoted by v and CRO is the carrier frequency offset.
proceed to next step.
2) Classification validates the signature belongs to the
4.3.2 Extracting the Features From Signal Waveform
transmitter that it claims. The classifier returns a
user ID that is stored. When the data preamble is Next, we extract the signal features and obtain a fingerprint
decoded, we verify that the sender and the output of signature of each waveform. We first detect the I/Q imbal-
the classifier (user ID) match. ance using the MATLAB Communication Toolbox’s comm.
In what follows, we describe the PHY-layer impairment IQImbalanceCompensator and obtain a compensator
model, feature extraction, data generation, and authentica- coefficient IQComp. The I/Q coefficients are estimated using
tion steps. iqcoef2imbal, which computes the amplitude and phase
imbalances given a compensator coefficient. We also need
4.3.1 Modeling the PHY-Layer Impairments to extract the synchronization in frequency and time
domains. The frequency-domain synchronization has two
Training data is generated in MATLAB to account for dif- components called the coarse and fine offsets. The time
ferent hardware impairments. In particular, the sampling domain synchronization has only one component. We use
rate offset, carrier frequency offset, and I/Q balance offset wlanCoarseCFOEstimate, wlanFineCFOEstimate,
(amplitude and phase) are taken into consideration. To and wlanSymbolTimingEstimate functions of MATLAB
model the sample rate offset (SRO) between the transmit- to estimate the coarse and fine center frequency offset, and
ter and receiver, the transmitted waveform is resampled symbol timing synchronization, respectively. Through these
with a factor of p=q times the original sample rate where steps, we obtain five signatures: Coarse center frequency
p and q are the interpolation and decimation factors. offset, fine center frequency offset, symbol timing synchro-
These parameters are varied for each transmitter-receiver nization, amplitude imbalance, and phase imbalance. Note
pair. We use the resample function of MATLAB. A fre- that to extract these features, we have used domain knowl-
quency offset is introduced to the previous signal using edge rather than following a purely data-driven approach.
the helperFrequencyOffset function of MATLAB. This is due to the fact that hardware impairments introduce
This joint SRO and carrier frequency offset impairments very subtle changes in the received signal such that domain
are suggested in [34]. Finally, for each transmitter- knowledge is required to provide a robust algorithm.
receiver pair, a constant I/Q offset is added by an amount
of c for amplitude imbalance and f for phase imbalance
using the following formulation [35]:
4.3.3 Data Generation
As a comprehensive dataset, we generate 50 signal samples
kI ¼ 100:5c=20 (2) at SNR values from 5 to 25 dB in 5 dB increments for each
subband. The same experiment is repeated for 10 users. To
kQ ¼ 100:5c=20 (3) introduce the frequency offset impairments, we consider
the SRO between a transmitter and receiver pair j to be a
imbI ¼ IReftxSigÞ kI expð0:5i fp=180Þ (4) multiple of 100 parts per million (PPM), i.e., vj ¼ 100j PPM
as suggested in [34]. For this purpose, we take the interpola-
tion factor as p ¼ 104 and vary the decimation factor
imbQ ¼ IImðtxSigÞ kQ expðiðp=2 þ 0:5fp=180ÞÞ (5)
q ¼ p j for each user j corresponding to a SRO of user j as
vj ¼ ð1 p=ðp jÞÞ106 PPM. Note that if the offsets are
rxSig ¼ imbI þ imbQ ; (6) larger, RF fingerprint signatures will be easier to
TABLE 4 TABLE 6
Confusion Matrix for MCD Confusion Matrix for the Isolation Forest
Outlier Detection Method
Predicted Label
(A) (O) Predicted Label
(A) 29.3% 10.2% (A) (O)

True Label (O) 0% 60.5% (A) 15.1% 24.4%
True Label (O) 6.5% 54.0%
“A” stands for authenticated and “O” stands for outlier.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
differentiate. These offsets are kept constant throughout dif- xÞ ¼ ðxx m x ÞT S1
MDðx x ðx x mx Þ, where m x and
ferent SNR and channel realizations. To introduce ampli-
tude and phase imbalance, 1 dB and 10 degrees increments S1
x are the mean and covariance of data x .
are added between users such that jth user has j dB and Suppose there are ten signatures, six of which are autho-
10 j degrees imbalance. rized and four are unauthorized (unseen) signatures. Table 4
presents the confusion matrix of MCD outlier detection that
achieves the average accuracy of 89.8 percent. We also present
4.3.4 Outlier Detection
the confusion matrix of one-class SVM with n ¼ 0:2 and isola-
To authorize the signals, we first use a simple outlier detection forest methods in Tables 5 and 6, respectively. The two
tion method in which the objective is to identify which sig- classifiers achieve only 70.0 and 69.1 percent accuracy, respec-
natures are authorized and which are not. We test and tively, in the same order as before. In Step 2, we use supervised
evaluate the performance of three different methods: (i) one- classification and employ RBF-kernel based SVM with C ¼ 1
class SVM [36], (ii) isolation forest [37], and (iii) Minimum and g ¼ 0:1 parameters, which achieves 100 percent accuracy.
Covariance Determinant (MCD) [38], [39] methods. Our
results show that MCD, also known as the elliptic envelope
method, outperforms the other two methods. In what fol- 4.4 Step 4: Channel Selection and Channel Access
lows, we briefly describe all three methods and then present Input: The set of channel labels and SINR from Step 2.
the performance evaluation results. Output: The ID of the channel selected for transmission.
Each user individually starts scanning the set of m chan-
One-class SVM is an unsupervised outlier detection nels (with a random initialization) and classifies each chan-
method. It learns a frontier delimiting the contour of nel as (I), (W) or (J) with the deep learning-based classifier
initial observations. If a new observation lays within (as described in Step 3).
the frontier-delimited subspace, it is as coming from
the same population (an authorized signature). Oth- 1) If channel i is classified as (I), the user transmits and
erwise, if they lay outside the frontier, it is an outlier breaks the scanning loop.
(an unauthorized signature). The parameter n is the 2) Else if channel i is classified as (W), the user backs-
margin of the One-Class SVM that determines the off (with exponential timer) and counts down from
probability of finding a new, but regular, observation 2k 1 (where k is the timer window).
outside the frontier. 3) Else if channel i is classified as (J), it is added to a
Isolation forest isolates observations by randomly possible list for data transmission and the user scans
selecting a feature and then randomly selecting a the next channel.
split value between the maximum and minimum Note that the third step does not exist in the baseline
values of the selected feature. This recursive parti- WiFi that treats channels (W) and (J) the same way and
tioning can be represented by a tree structure and backs off on all of these channels. At the end of scanning
the number of splittings required to isolate a sample channels, each user performs the following:
is equivalent to the path length from the root user to
1) If there is no (I) channel (channels are either (W) and
the terminating user. This path length, averaged
(J)), the user selects the channel (J) with the best
over a forest of such random trees, is a measure of
SINR and transmits
normality and our decision function.
2) If there is no channel (I) and (J), i.e., all channels are
MCD fits an elliptic envelope such that any data
(W), while back-off counter is not zero, the user waits
point outside the ellipse is an outlier. It is a highly
for one time slot and reduces all counters by 1.
robust estimator of multivariate distributions [38].
3) If any back-off counter is zero, the user senses that
It uses the Mahalanobis distance such that
channel.
If the channel is (I), the user transmits on that
TABLE 5 channel.
Confusion Matrix for the One-Class SVM Method Else, the user resets the back-off counter and uni-
formly selects a random number between 0 and
Predicted Label
2kþ1 1.
(A) (O)
(A) 22.8% 16.7%
4.5 Step 5: Power Control for LPI/LPD
True Label (O) 13.3% 47.2%
Input: The ID and SINR of the selected channel from Step 4.
TABLE 7
MCSs and Corresponding Rates for IEEE 802.11ac With One
Spatial Stream (GI Stands for Guard Interval)
Data Rate (Mb/s)

MCS Modulation Coding 800 ns GI 400 ns GI
0 BPSK 1/2 6.5 7.2
1 QPSK 1/2 13.0 14.4
2 QPSK 3/4 19.5 21.7
3 16-QAM 1/2 26.0 28.9
4 16-QAM 3/4 39.0 43.3
5 64-QAM 2/3 52.0 57.8
6 64-QAM 3/4 58.5 65.0
7 64-QAM 5/6 65.0 72.2
8 256-QAM 3/4 78.0 86.7
Fig. 8. Best selection of MCS levels at different SINRs for different
payloads.
Output: The transmit power given to the WiFi transmitter

Output: The ID of the flow to serve and the ID of the
chain. Each DeepWiFi user reduces its power to the mini-
neighbor selected for next-hop transmission.
mum to meet the minimum SINR requirement to its
Each user applies the backpressure algorithm [40] in a dis-
intended receiver for LPI/LPD capability, assuming the
tributed setting without any centralized controller. Note
channel estimation from itself to its intended receiver is car-
that another wireless routing algorithm can be also applied
ried out periodically. Consider a user i with the chosen
here. We use the backpressure algorithm to reflect both
transmit power Pi and jammer j with sensing threshold t j ,
channel and queue information in routing decisions that
and let hij denote the pathloss from user i to jammer j. Then
make use of the optimized link rates from Step 6. Each user
jammer j cannot detect transmission of user i if Pi hij < t j .
exchanges local information on its spectrum utility with its
Note that t j and hij are unknown to user i.
neighbors [41]. Each user i keeps a separate queue for each
flow s with backlog Qsi ðtÞ at time t. For all links, the user
4.6 Step 6: Adaptive Modulation and Coding chooses the flow to transmit as the one with the maximum
Input: The ID and SINR of the selected channel from Step 4. difference of queue backlogs at the receiving and transmit-
Output: The MCS level and the corresponding rate for ting ends, i.e., for each link ði; jÞ a user i chooses the flow
this channel given to the WiFi transmitter chain.
Each user selects the modulation and coding rate for sij ¼ argmax½Qsi ðtÞ Qsj ðtÞþ ; (7)
802.11ac based on the SINR that is measured on the sensed s
channel. We consider the MCSs defined for the Very High where ½þ ¼ maxð; 0Þ. Then, we define the spectrum utility
Throughput (VHT) scheme of 802.11ac, as shown in Table 7.
We label the generated samples at the receiver with the þ
s s
MCS scheme that gives the best error rate and throughput Uij ðtÞ ¼ cij ðtÞ Qi ij ðtÞ Qj ij ðtÞ (8)
trade-off. By design, we want lower MCS level (index) for
lower SINR and higher MCS level for higher SINR. This for user i transmitting to user j, where cij ðtÞ is the rate on link
approach provides an effective link adaptation. We deter- ði; jÞ at time t (depending on transmit power from Step 5 and
mine a table that assigns the best MCS for a given SINR. MCS from Step 6). User i transmits to the selected neighbor
This table is pre-loaded to each user that adapts its MCS j ðtÞ that yields the maximum spectrum utility, i.e.,
according to the measured SINR. To build up this table, we
initially generate the transmitted signal using a high MCS j ðtÞ ¼ argmaxj2Ni Uij ðtÞ; (9)
level. We transmit the signal over the channel and add
noise. We estimate the channel at the receiver using the pre- where Ni is the set of the next hop candidates for user i to select
amble and equalize the signal. The equalized samples are from. Each user uses the data channel (in-band) to exchange
then used to demodulate the signal. Thus, we obtain the control information with its neighbors. There is no separate
packet error rate by comparing the transmitted and received control channel. Users asynchronously make decisions in a
bits. If the packet error rate is not zero, we reduce the MCS distributed setting by using the following four phases [41]:
level and repeat the same procedure. If there are still errone-
1) neighborhood discovery and channel estimation,
ous bits at the receiver for MCS 0, then we keep the MCS as
2) exchange of flow information updates and execution
0. Fig. 8 shows the best MCS levels (when 256, 512, and 1024
of the backpressure algorithm,
bytes of data payload are used, respectively) to maximize
3) transmission decision negotiation, and
the rate depending on the SINR.
4) data transmission.
4.7 Step 7: Routing
5 SIMULATION SETTING FOR NETWORK-LEVEL
Input: The set of link rates (given from Step 6) from a given
user to its neighbors (note that the set of queue lengths of PERFORMANCE EVALUATION
the user and its neighbors are obtained through information There are n ¼ 9 users, m ¼ 40 channels, and 5 flows gener-
exchange in Step 7). ated at each simulation time. There are nJ ¼ 40 jammers,
Fig. 9. Network topology for simulations. Fig. 10. Throughput of individual users (when pJ ¼ 0:7) for fixed SINR
(0 dB).
each assigned to one channel. Each random jammer is inde-
Consider the following metrics to measure the
pendently turned on with the probability of jamming pJ on
performance:
every simulation time, where pJ is varied between 0 to 1
with 0.05 increments. The received SINR (due to jammer (1) Average throughput (Mb/s) per user: The average
and noise) is also varied with 1 dB increments. Each user i throughput rate that each user can achieve during
generates traffic to a randomly selected destination j. The the simulation.
traffic rate is randomly selected from [0,1] Mb/s with aver- (2) Cumulative throughput: The average throughput of
age rate rij ¼ 500 kbps for a source-destination pair ði; jÞ. the network for all users during the simulation.
The simulator is implemented in MATLAB. Simulation
time is 100 seconds. The baseline WiFi features are imple- 6 PERFORMANCE RESULTS
mented using the MATLAB WLAN System Toolbox. The
deep learning code is implemented in TensorFlow. 6.1 Probabilistic (Random) Jammer
We compare the performance of DeepWiFi with the First, we fix the SINR to 0 dB and start with random jammers.
baseline WiFi (without deep learning or jamming resis- The throughput of individual users (when pJ ¼ 0:7) for fixed
tance) and jamming-resistant MAC protocol JADE [12]. SINR (0 dB) is shown in Fig. 10. The end-to-end network
JADE is asymptotically optimal in the presence of adap- throughput is shown in Fig. 11 as a function of pJ . For small
tive adversarial jamming. As the baseline WiFi is not pJ , both DeepWiFi and baseline WiFi achieve the same
designed to mitigate any jamming, its comparison with throughput, since all users can find idle channels without
DeepWiFi quantifies the effect of jamming defense. In internal or external interference. JADE is run with its default
DeepWiFi, we use the exponential backoff defined in the parameters according to [12]. Due to its small channel
IEEE 802.11ac standard. The comparison of DeepWiFi access probability initialization, JADE starts off worse than
with JADE quantifies the difference in the channel access the baseline WiFi and DeepWiFi. As pJ increases beyond 0.3,
and different solutions spaces. Channel access in JADE the throughput of baseline WiFi starts dropping sharply,
does not differentiate between a WiFi or a jamming sig- while DeepWiFi sustains its throughput and provides major
nal, whereas DeepWiFi does. Therefore, the solution throughput gains relative to baseline WiFi. For pJ 0:5,
spaces are different. We show that DeepWiFi achieves JADE resists to jamming better than the baseline WiFi by
major performance improvement over the baseline WiFi adjusting the channel access probability and outperforms it
and JADE. In all three algorithms, we use adaptive mod- while performing worse than DeepWiFi. Even when all
ulation and coding and backpressure routing algorithm
for a fair comparison.
The network topology is shown in Fig. 9 where we
deployed nine users uniformly at random over a given
area. Since friendly users do not interfere with each other
due to the backoff mechanism and signal classification,
their locations do not affect the system performance. The
friendly users (DeepWiFi or baseline WiFi) are labeled
with IDs 1-9. Each link represents the set of neighbors
(depending on the SINR threshold less than 0 dB). The 40
jammers are depicted as users without links on the right
and top of Fig. 9 (with user ID 10-49). Each jammer is
responsible to jam a single channel (out of 40 channels).
For instance, jammers with ID 10, 11, 12 would respec-
tively jam channels 1, 2, 3 and so on. When the jammer is
on (off), it is depicted as a red (blue) user. Fig. 11. Cumulative throughput as a function of pJ for fixed SINR (0 dB).
Fig. 14. The transmit power is adjusted for LPI/LPD.

Fig. 12. Cumulative throughput as a function of SINR for pj ¼ 0:8.
channels are jammed all the time (with pJ ¼ 1), DeepWiFi can channel greater than or equal to a threshold t, that is
achieve the network end-to-end throughput close to 70 Mbs, ðnÞ
ri t, else it is turned on with a fixed probability of jam-
while the throughputs of baseline WiFi and JADE are zero ming, pJ .
(note that fluctuations are due to randomness in channels and In our experiments, the power detection threshold of
interference effects). DeepWiFi activates more links than base- sensing-based jammer is varied from 2 dB to 10 dB with 1 dB
line WiFi and JADE, as users back off on the jammed channels increments. Fig. 13 shows the average individual throughput
in both schemes, whereas DeepWiFi allows users to transmit (Mb/s) of DeepWiFi, JADE, and baseline when sensing-based
on the jammed channels (if no idle channel available) in the jammers are on, pJ ¼ 0:7 and t J ¼ 5 dB. We observe that the
degraded mode. We start changing the SINR due to jammer sensing-based jammer reduces the throughput of baseline
and noise effects. The cumulative throughput is shown in WiFi to zero as users end up backing off indefinitely. Next,
Fig. 12 as a function of SINR at pJ ¼ 0:8. The cumulative we present the results for transmit power control to DeepWiFi
throughput slowly increases with SINR for all schemes and to demonstrate the LPI/LPD capability. DeepWiFi with trans-
DeepWiFi outperforms others over all SINRs. Also, note that mit power control for LPI/LPD can avoid jammers by operat-
the cumulative throughput term in this paper accounts for ing below t J . Fig. 14 shows the histogram of transmit power
multiple hops across the network and represents the end-to- per user. We observe that about half of the transmissions are
end network throughput. with low power to provide LPI/LPD capability, whereas the
other half are with high power to support communication
6.2 Static Sensing-Based Jammer over the jammed channels. As a result, DeepWiFi with LPI/
So far, we considered random (probabilistic) jammers that LPD can achieve higher rates higher compared to baseline
are turned on with some fixed probability. Next, we evalu- WiFi, JADE, and DeepWiFi without LPI/LPD. Fig. 15 shows
ate the effect of sensing-based jammers and the perfor- the average individual throughput (Mb/s) of DeepWiFi and
mance of power control for LPI/LPD. Here, we consider a baseline WiFi when sensing-based jammers are on, pJ ¼ 0:7,
static jammer that has a constant sensing threshold t, t J ¼ 2 dB, and the transmit power is adjusted for LPI/LPD.
whereas an adaptive jammer (discussed in Section 6.3 can We observe that as the jammer becomes more reactive, i.e.,
ðnÞ
dynamically adjust its sensing threshold. Let rk denote the the detection threshold of jammer t J decreases from 10 dB to
jammer’s received signal on channel n at time k. The static 2 dB, the throughput of DeepWiFi and JADE decrease while
sensing-based jammer turns on if it detects a signal on the the baseline WiFi’s throughput diminishes completely.
Fig. 15. Individual throughput (Mb/s) of DeepWiFi and baseline WiFi

Fig. 13. Individual throughput (Mb/s) of DeepWiFi and baseline WiFi when sensing-based jammers are on, pJ ¼ 0:7, SINR is 5 dB, tJ ¼ 2 dB,
when sensing-based jammers are on, pJ ¼ 0:7 and tJ ¼ 10 dB. and the transmit power is adjusted for LPI/LPD.
Fig. 16. Cumulative throughput (Mb/s) as a function of pJ when the num- Fig. 17. Cumulative throughput (Mb/s) as a function of pJ in the presence
ber of channels is less than the number of users in the presence of static of adaptive jammers.
sensing-based jammers.
Fig. 17 shows the performance of DeepWiFi, baseline,
We also evaluate DeepWiFi when there are more and JADE under adaptive jamming. We took the initial
users than available channels. We consider 9 users and 9 sensing threshold as t 0 ¼ 1, weighting constant as w ¼ 1,
traffic flows sharing 6 channels. Fig. 16 presents the and step size constant D ¼ 0:5. We observe that the cumu-
cumulative throughput as a function of jamming proba- lative throughput reduces significantly in the presence of
bility pJ for DeepWiFi, baseline WiFi, and JADE. We adaptive jammers in all cases, while DeepWiFi still pro-
observe a linear decrease in cumulative throughput as vides the best performance compared to the baseline WiFi
the jamming probability increases. Note that the cumula- and JADE.
tive throughput of DeepWiFi depends on the jamming Overall, the results evaluated in the presence of three dif-
probability, the number of users, number of channels, ferent types of jammers indicate that DeepWiFi provides
backoff mechanism, jamming power, and signal classifi- reliable and robust communication.
cation accuracy. Comparing Figs. 11 and 16, we observe
that the slope of cumulative throughput versus jamming
7 IMPLEMENTATION, COMPLEXITY,
probability shows logarithmic-like behavior when there
are fewer users than channels, whereas the slope is AND OVERHEAD ASPECTS
almost linear when there are more users than channels DeepWiFi can be implemented in the kernel or in the WiFi
(i.e., effect of congestion increases). card. Open-source firmwares such as Nexmon [11] can
extract the I/Q data out of the WiFi such that the I/Q data
6.3 Adaptive Jammer can be processed in the kernel. System-on-chip (SOC) solu-
Next, we consider the adaptive jammer that dynamically tions built for IoT such as Qualcomm QCS603 / QCS605
adjusts the sensing threshold t t at time t by observing the [42] can be also used to extract the I/Q data that can be
channel access patterns of WiFi users. The channel utiliza- processed in the kernel (or can be moved to an additional
tion on channel n (as the jammer perceives it) at time k is FPGA or ARM for further processing). These SOCs include
ðnÞ
1ðrk t k Þ, where 1ðxÞ denotes the indicator function, that 802.11ac WiFi and a neural processing engine (for deep
is, 1 when x is True and 0 otherwise. At time t, the adap- learning) integrated in the chip that can be used for RF com-
tive jammer’s utility function that includes the channel uti- munication and deep learning. The training is usually per-
lization and its own power consumption is defined as formed offline and the trained models are ported to the
kernel or to the other platforms for inference.
N
1 X t X ðnÞ DeepWiFi can get the training data in two possible ways.
gðrðtÞ; pðtÞÞ ¼ 1ðrk t k Þ þ w pk ;
T0 k¼tT þ1 n¼1 First, channels and signals can be generated by MATLAB
0
WLAN Toolbox (as discussed in Section 2.2), and they can
(10) be used to build the training data. Second, a real 802.11.ac
where pk is the jamming power at time k and w 0 is a WiFi card can be used to make over-the-air transmissions
weighting constant to balance the tradeoff between jammer that are captured by an SDR, and the channels and signals
power consumption and WiFi channel utilization. The vec- can be used for training.
tors rðtÞ and pðtÞ denote the received signal and transmit To evaluate the time complexity and overhead, we
power values of the jammer, respectively, over a time win- implemented each deep learning task in a low-cost embed-
dow of T0 time instances ending at time t and N channels, ded system, NVIDIA Jetson Nano Developer Kit [43]. For
ðnÞ
namely rðtÞ ¼ frk ; 1 n N; t T0 þ 1 k tg and an efficient deployment, we converted the Tensorflow code
pðtÞ ¼ fpk ; t T0 þ 1 k tg. At time t, the jammer trans- to TensorRT [44], inference optimizer for NVIDIA’s embed-
mits when its received signal on channel n is greater than or ded systems. We repeated the tests 1000 times to calculate
ðnÞ
equal to its sensing threshold, namely rt t t . Using a suf- the average inference time. For example, front end process-
ficiently small step size constant D > 0, the adaptive jam- ing of each data sample takes 0.09 msec. A typical frame in
D
mer updates its sensing threshold as t tþ1 ¼ tDtþ if the 802.11ac standard is 5.484 msec. [16]. Hence, the front
þ t
gðrðt 1Þ; pðt 1ÞÞ > gðrðtÞ; pðtÞÞ, else t tþ1 ¼ t t t . end processing time is measured as small as 1.6 percent of
the 802.11 frame. The reported processing time is for the features by machine learning and authenticates legitimate
case when there is already data available to be processed. WiFi signals. Using these signal labels on sensed channels,
As the performance of the off-the-shelf hardware can be DeepWiFi supports users to use idle channels, back-off on
degraded when there are other processes (such as other channels used by legitimate WiFi signals, and access (if
WiFi operations) running in the background, we used needed in the degraded mode) channels that are occupied
the stress ng software[45] to stress test the memory by signals other than legitimated WiFis. DeepWiFi users
use of the embedded GPU by generating background pro- optimize their transmit powers for LPI/LPD and their MCS
cesses emulating high load conditions. We observed that to maximize link rates, and their routing decisions by back-
the processing time for deep learning operations increased pressure algorithm. We simulated DeepWiFi in a distrib-
33 percent but remained small relative to MAC frame uted network using the channels and signals generated by
length. Note that the overhead and processing times on MATLAB WLAN Toolbox. We showed that DeepWiFi
embedded platforms can be further reduced by customized helps WiFi users sustain their throughput with major per-
ASIC chips and FPGAs. formance gain relative to baseline WiFi and another jam-
For the signal classification task, we observed that the ming-resistant MAC protocol especially when channels are
FNN model takes 0.009 msec on average to predict each likely to be jammed and the SINR is low.
sample point. On the other hand, the CNN model takes
0.035 msec to predict a sample over 1000 repetitions. As a
ACKNOWLEDGMENTS
comparison between the FNN and CNN models, both archi-
tectures have similar performance. Since the CNN architec- This effort is supported by the U.S. Army under contract
ture has more layers, it takes approximately four times W91CRB-17-P-0068. The content of the information does not
longer to process a single waveform in inference time. necessarily reflect the position or the policy of the U.S. Gov-
CNNs in return have a smaller footprint on the hardware. ernment, and no official endorsement should be inferred.
Note that the difference in hardware footprint is expected
since the FNNs use fully connected layers to reduce dimen- REFERENCES
sions from one layer to another, whereas the CNNs employ [1] C. Clancy, J. Hecker, E. Stuntebeck, and T. O’Shea, “Applications
convolutional layers that utilize feature maps in a sliding of machine learning to cognitive radio networks,” IEEE Wireless
window manner. For more discussion on the number of Commun., vol. 14, no. 4, pp. 47–52, Aug. 2007.
[2] K. Thilina, K. W. Choi, N. Saquib, and E. Hossain, “Machine learn-
parameters in FNN and CNN models, we refer the reader ing techniques for cooperative spectrum sensing in cognitive
to [33, p. 357]. radio networks,” IEEE J. Selected Areas Commun., vol. 31, no. 11,
For the signal authentication task, the MCD outlier detec- pp. 2209–2221, Nov. 2013.
tion takes 0.0556 msec, one-class SVM takes 52.813 msec, [3] M. Alsheikh, S. Lin, D. Niyato, and H. Tan, “Machine learning in
wireless sensor networks: Algorithms, strategies, and applications,”
and isolation forest takes 33.521 msec. IEEE Commun. Surveys Tutorials, vol. 16, no. 4, pp. 1996–2018,
Adding these components, we observe that the overall Apr. 2014.
processing overhead for the front end processing, signal [4] M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah, “Machine
learning for wireless networks with artificial intelligence: A
classification using FNN, and signal authentication using tutorial on neural networks,” 2017, arXiv:1710.02913.
MCD is 0.1546 msec, which is only 2.8 percent of the [5] T. O’Shea, J. Corgan, and C. Clancy, “Convolutional radio modu-
802.11ac frame. lation recognition networks,” in Proc. Int. Conf. Eng. Appl. Neural
For the routing overhead, the backpressure algorithm Netw., 2016, pp. 213–226.
[6] W. Lee, M. Kim, D. Cho, and R. Schober, “Deep sensing: Coopera-
exchanges typical messages as in the regular distance-vector tive spectrum sensing based on convolutional neural networks,”
routing algorithms. There is no closed form expression for 2017, arXiv:1705.08164.
the number of message exchanges and the exact number [7] K. Davaslioglu and Y. E. Sagduyu, “Generative adversarial learn-
depends on the network and traffic conditions. The backing for spectrum sensing,” in Proc. IEEE Int. Conf. Commun.,
May 2018, pp. 1–6.
pressure algorithm was implemented with SDRs in [46] that [8] Y. Shi, Y. E. Sagduyu, T. Erpek, K. Davaslioglu, Z. Lu, and J. Li,
empirically evaluated the number of message exchanges. “Adversarial deep learning for cognitive radio security: Jamming
attack and defense strategies,” in Proc. IEEE Int. Conf. Commun.
Workshops, 2018, pp. 1–6.
8 CONCLUSION [9] T. Erpek, Y. E. Sagduyu, and Y. Shi, “Deep learning for launching
and mitigating wireless jamming attacks,” IEEE Trans. Cogn.
We presented the DeepWiFi protocol that applies machine Commun. Netw., vol. 5, no. 1, pp. 2–14, Mar. 2019.
learning, and in particular deep learning to adapt WiFi to [10] C. Funai, C. Tapparello, and W. Heinzelman, “Enabling multi-hop
ad hoc networks through WiFi Direct multi-group networking,”
spectrum dynamics and provides major throughput gains in Proc. IEEE Conf. Comput. Netw. Commun., 2017, pp. 491–497.
compared to baseline WiFi and another jamming-resistant [11] M. Schulz, D. Wegemer, and M. Hollick, “Nexmon: Build your
MAC protocol. DeepWiFi is designed to mitigate out-of- own Wi-Fi testbeds with low-level MAC and PHY-Access using
firmware patches on off-the-shelf mobile devices,” in Proc. Proc.
network interference effects from probabilistic and sensing- 11th Workshop Wireless Netw. Testbeds Exp. Eval. Characterization,
based jammers. Built upon the PHY transceiver chain of 2017, pp. 59–66.
IEEE 802.11ac, DeepWiFi provides the decision parameters [12] A. Richa, C. Scheideler, S. Schmid, and J. Zhang, “A jamming-
to WiFi transceiver without changing the PHY and the resistant mac protocol for multi-hop wireless networks,” in Dis-
tributed Computing. Berlin, Germany: Springer, 2010, pp. 179–193.
MAC frame format. DeepWiFi applies deep learning-based [13] W. A. Gardner and C. M. Spooner, “Cyclic spectral analysis for
autoencoder to extract features and applies deep neural net- signal detection and modulation recognition,” in Proc. IEEE
works (FNN or CNN) to classify signals as WiFi, jammed, Military Commun. Conf., 1988, pp. 419–424.
or idle. The signals classified as WiFi are further processed [14] A. Nandi and E. Azzouz, “Modulation recognition using artificial
neural networks,” Signal Process., vol. 56, no. 2, pp. 165–175, 1997.
through RF fingerprinting that identifies hardware-based
[15] T. J. O’Shea and N. West, “Radio machine learning dataset [43] “Jetson Nano Developer Kit,” NVIDIA, Mar. 19, 2019, Accessed:
generation with GNU radio,” in Proc. GNU Radio Conf., 2016, pp. 1–6. Mar. 19, 2019. [Online]. Available: https://developer.nvidia.
[16] M. S. Gast, 802.11ac Survival Guide. Newton, MA, USA: O’Reilly, com/embedded/jetson-nano-developer-kit
2013. [44] “NVIDIA TensorRT Programmable Inference Accelerator,” NVI-
[17] O. Ureten and N. Serinken, “Wireless security through RF finger- DIA, Mar. 27, 2018, Accessed: May 1, 2018. [Online]. Available:
printing,” Can. J. Elect. Comput. Eng., vol. 32, pp. 27–33, 2007. https://developer.nvidia.com/tensorrt
[18] B. Ramsey, M. Temple, and B. Mullins, “PHY foundation for [45] “Stress-ng,” Ubuntu Wiki, Dec. 18, 2018, Accessed: Oct. 1, 2019.
multi-factor ZigBee node authentication,” in Proc. IEEE Global [Online]. Available: https://wiki.ubuntu.com/Kernel/Reference/
Commun. Conf., 2012, pp. 795–800. stress-ng
[19] A. Candore, O. Kocabas, and F. Koushanfar, “Robust stable radio- [46] S. Soltani, Y. Sagduyu, Y. Shi, J. Li, J. Feldman, and J. Matyjas,
metric fingerprinting for wireless devices,” in Proc. IEEE Int. “Distributed cognitive radio network architecture, SDR imple-
Workshop Hardware-Oriented Security Trust, 2009, pp. 43–49. mentation and emulation testbed,” in Proc. IEEE Military Commun.
[20] Wi-Fi Alliance, “P2P Technical Group, Wi-Fi Peer-to-Peer (P2P) Conf., Oct. 2015, pp. 438–443.
Technical Specication v1.7,” Accessed: Jul. 2016. [Online]. Avail-
able: https://www.wi-fi.org/discover-wi-fi/specifications Kemal Davaslioglu received the PhD degree in
[21] Android Developers, “Wi-Fi Peer-to-Peer,” Jul. 30, 2013. [Online]. electrical and computer engineering from the
Available: http://developer.android.com/guide/topics/ University of California, Irvine. He is a senior
connectivity/wifip2p.html research scientist at Intelligent Automation Inc.
[22] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, “Extracting (IAI). He held long-term internships at Broadcom
and composing robust features with denoising autoencoders,” in Proc. Inc., Irvine, CA, where he worked on beamform-
Int. Conf. Mach. Learn., 2008, Art. no. 10961103. ing algorithms and 60 GHz channel characteriza-
[23] P. Vincent et al., “Stacked denoising autoencoders: Learning use- tion. His research interests include resource
ful representations in a deep network with a local denoising allocation in wireless networks, machine learning,
criterion,” J. Mach. Learn. Res., vol. 11, pp. 3371–3408, 2010. adversarial learning, and security. He is the co-
[24] D. P. Kingma, D. J. Rezende, S. Mohamed, and M. Welling, “Semi- author of the IEEE HST 2018 Best Paper Award.
supervised learning with deep generative models,” in Proc.
Advances Neural Inf. Process. Syst., Oct. 2014, pp. 3581–3589. Sohraab Soltani received the PhD degree in
[25] O. A. Topal, S. Gecgel, E. M. Eksioglu, and G. K. Kurt, computer science from Michigan State Univer-
“Identification of smart jammers: Learning based approaches sity, East Lansing, Michigan. He is a lead scien-
using wavelet representation,” 2019, arXiv:1901.09424. tist at Intelligent Automation Inc. His research
[26] C. Funai, C. Tapparello, and W. Heinzelman, “Supporting multi- interests include wireless communications, cog-
hop device-to-device networks through WiFi direct multi-group nitive radio, ad hoc and sensor networks, and
networking,” 2015, arXiv:1601.00028. wireless system implementation. He imple-
[27] I. T. Jolliffe, Principal Component Analysis. Berlin, Germany: mented a broad set of wireless (PHY, MAC, and
Springer, 2002. network layer) technologies on various radio plat-
[28] L. J. P. van der Maaten and G. Hinton, “Visualizing high-dimensional forms and executed in-lab and field tests.
data using t-SNE,” J. Mach. Learn. Res., vol. 9, pp. 2579–2605, 2008.
[29] L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar,
“Hyperband: A novel bandit-based approach to hyperparameter Tugba Erpek received the MSc degree in electri-
optimization,” J. Mach. Learn. Res., vol. 18, pp. 1–52, 2018. cal and computer engineering from George
[30] J. Xie, L. Xu, and E. Chen, “Image denoising and inpainting with Mason University, Fairfax, Virginia, and the PhD
deep neural networks,” in Proc. 25th Int. Conf. Neural Inf. Process. degree in electrical engineering from the Virginia
Syst., 2012, pp. 341–349. Polytechnic Institute and State University,
[31] M. Abadi et al., “Tensorflow: A system for large-scale machine Blacksburg, Virginia. She is a senior research
learning,” in Proc. Symp. Operating Syst. Design Implementation, engineer at Intelligent Automation Inc. Her
2016, pp. 265–283. [Online]. Available: tensorflow.org research interests include wireless communica-
[32] D. P. Kingma and J. Ba, “Adam: A method for stochastic tions, machine learning, resource allocation,
optimization,” 2014, arXiv:1412.6980. dynamic spectrum access, and cognitive radio.
[33] A. Geron, Hands-on Machine Learning with Scikit-Learn and
Tensorflow. Newton, MA, USA: O’Reilly Media, Mar. 2017.
[34] Mathworks, “Joint sampling rate and carrier frequency offset Yalin E. Sagduyu received the PhD degree in
tracking,” Sep. 15, 2016, Accessed: Feb. 2, 2018. [Online]. Avail- electrical and computer engineering from the Univer-
able: https://www.mathworks.com/help/wlan/examples/joint- sity of Maryland, College Park. He is the director of
sampling-rate-and-carrier-frequency-offset-tracking.html networks and security at Intelligent Automation Inc.
[35] Mathworks, “iqcoef2imbal function,” Sep. 15, 2016, Accessed: Feb. 2,
His research interests include networks, communi-
2018. [Online]. Available: https://www.mathworks.com/help/
cations, and machine learning. He chaired work-
comm/ref/iqcoef2imbal.html shops at ACM MobiCom, ACM WiSec, IEEE CNS,
[36] B. Schlkopf et al., “Estimating the support of a high-dimensional and IEEE ICNP, served as track chair at IEEE
distribution,” Neural Comput., vol. 13, pp. 1443–1471, 2001. PIMRC, IEEE MILCOM, and IEEE GlobalSIP,
[37] F. T. Liu, K. M. Ting, and Z. Zhou, “Isolation forest,” in Proc. 8th and served on the organizing committee of IEEE
IEEE Int. Conf. Data Mining, 2008, pp. 413–422. GLOBECOM. He received the IEEE HST 2018 Best
[38] M. Hubert and M. Debruyne, “Minimum covariance determi-
Paper Award. He is a senior member of the IEEE.
nant,” Advanced Rev., vol. 2, pp. 36–43, 2010.
[39] P. Rousseeuw and K. V. Driessen, “A fast algorithm for the mini-
mum covariance determinant estimator,” Technometrics, vol. 41,
pp. 212–223, 1999. " For more information on this or any other computing topic,
[40] L. Tassiulas and A. Ephremides, “Stability properties of con- please visit our Digital Library at www.computer.org/csdl.
strained queueing systems and scheduling policies for maximum
throughput in multihop radio networks,” IEEE Trans. Autom.
Control, vol. 37, no. 12, pp. 1936–1948, Dec. 1992.
[41] S. Soltani et al., “Joint network coding and backpressure algo-
rithm for cognitive radio networks,” in Proc. IEEE Military Com-
mun. Conf., 2017, pp. 858–863.
[42] “Qualcomm QCS605 SoC,” Qualcomm, Apr. 11, 2018, Accessed:
Nov. 14, 2018. [Online]. Available: https://www.qualcomm.
com/products/qcs605

DeepWiFi Cognitive WiFi With Deep Learning

Uploaded by

Copyright:

Available Formats

DeepWiFi Cognitive WiFi With Deep Learning

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DeepWiFi Cognitive WiFi With Deep Learning

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 20, NO.

2, FEBRUARY 2021 429

DeepWiFi: Cognitive WiFi with Deep Learning

IEEE 802.11ac standard is followed for the PHY imple-

classifies the input signal as WiFi (W), Jammer (J), or

Fig. 3. Denoising autoencoder in DeepWiFi.

The denoising autoencoder of DeepWiFi adds the Gauss-

Predicted Label Predicted Label

4.3 Step 3: Signal Authentication

Fig. 6. Two-layer approach for RF fingerprinting.

period of time (microseconds) and is hard to model. The

(A) 29.3% 10.2% (A) (O)

Data Rate (Mb/s)

Output: The transmit power given to the WiFi transmitter

Fig. 14. The transmit power is adjusted for LPI/LPD.

Fig. 15. Individual throughput (Mb/s) of DeepWiFi and baseline WiFi

You might also like