RFAL: Adversarial Learning for RF Transmitter ...

2332-7731 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCCN.2019.2948919, IEEETransactions on Cognitive Communications and Networking

1

RFAL: Adversarial Learning for RF TransmitterIdentification and Classification

Debashri Roy∗, Tathagata Mukherjee†, Mainak Chatterjee∗, Erik Blasch‡, Eduardo Pasiliao§

∗ Computer ScienceUniversity of Central Florida

Orlando, FL 32826{debashri, mainak}@cs.ucf.edu

† Computer ScienceUniversity of AlabamaHuntsville, AL 35899{tm0130}@uah.edu

‡ Fellow IEEE{erik.blasch}@gmail.com

§ Munitions DirectorateAir Force Research Laboratory

Eglin AFB, FL, 32542{eduardo.pasiliao}@us.af.mil

Abstract—Recent advances in wireless technologies have ledto several autonomous deployments of such networks. As nodesacross distributed networks must co-exist, it is important that alltransmitters and receivers are aware of their radio frequency(RF) surroundings so that they can adapt their transmissionand reception parameters to best suit their needs. To thisend, machine learning techniques have become popular as theycan learn, analyze and predict the RF signals and associatedparameters that characterize the RF environment. However, inthe presence of adversaries, malicious activities such as jammingand spoofing are inevitable, making most machine learningtechniques ineffective in such environments.

In this paper we propose the Radio Frequency AdversarialLearning (RFAL) framework for building a robust system toidentify rogue RF transmitters by designing and implementing agenerative adversarial net (GAN). We hope to exploit transmitterspecific “signatures” like the the in-phase (I) and quadrature (Q)imbalance (i.e., the I/Q imbalance) present in all transmittersfor this task, by learning feature representations using a deepneural network that uses the I/Q data from received signalsas input. After detection and elimination of the adversarialtransmitters RFAL further uses this learned feature embeddingas “fingerprints” for categorizing the trusted transmitters. Morespecifically, we implement a generative model that learns thesample space of the I/Q values of known transmitters and usesthe learned representation to generate signals that imitate thetransmissions of these transmitters. We program 8 universalsoftware radio peripheral (USRP) software defined radios (SDRs)as trusted transmitters and collect “over-the-air” raw I/Q datafrom them using a Realtek Software Defined Radio (RTL-SDR),in a laboratory setting. We also implement a discriminatormodel that discriminates between the trusted transmitters andthe counterfeit ones with 99.9% accuracy and is trained inthe GAN framework using data from the generator. Finally,after elimination of the adversarial transmitters, the trustedtransmitters are classified using a convolutional neural network(CNN), a fully connected deep neural network (DNN) and arecurrent neural network (RNN) to demonstrate building of anend-to-end robust transmitter identification system with RFAL.Experimental results reveal that the CNN, DNN, and RNN areable to correctly distinguish between the 8 trusted transmitterswith 81.6%, 94.6% and 97% accuracy respectively. We alsoshow that better “trusted transmission” classification accuracy isachieved for all three types of neural networks when data fromtwo different types of transmitters (different manufacturers) areused rather than when using the same type of transmitter (samemanufacturer).

Keywords: RF fingerprinting, GAN, machine learning,convolutional neural network, deep neural network, recurrentneural network, I/Q imbalance, software defined radios.

I. INTRODUCTION

The advent of Internet-of-Things (IoT) [1], large scalesensor networks and the possibility of having large scalevehicular networks in the near future, have ushered in anew era of industrial scale deployment of radio frequency(RF) signal sources (aka RF transmitters). Different automaticmethods for optimized communication that involve localiza-tion as well as identification and characterization of suchdevices, are becoming indispensable for many applicationssuch as locating a cell phone, identifying jammers and roguetransmitters, detecting the presence or absence of a signalsource, tracking objects, etc. With large scale autonomousdeployments of wireless networks, identification of trans-mitters has become an important problem. For example, awireless sensor network (WSN) relies on trustworthy signals;however, malicious transmitters can contaminate the signalsand jeopardize the utility of the sensor network. Existence ofsuch threats underscore the need for techniques that recognizeand authenticate the identity of transmitters, irrespective ofthe network protocols and communication technologies beingused. One of the ways that this can be done is through theexchange of keys and identity information. However, suchmethods are prone to malicious attacks by adversaries aswell. Hence techniques for robust transmitter identificationand authentication in adversarial settings is important formaintaining the integrity of large scale RF deployments. Arobust alternative to using a system based on key exchangesis to use machine learning (ML) for automatic identificationand characterization of communicating entities.

In recent years, there has been a proliferation of autonomoussystems that use ML algorithms on large scale sensor data.These systems assume that the data and signals being usedare trustworthy even though they are received from several het-erogeneous sources (e.g., cameras, microphones, etc.). Whenusing ML techniques for communication networks, maliciousentities, such as rogue transmitters can alter the signal (andhence the data) by exploiting different targeted data generationtechniques based on generative ML models. Such threatsand their ease of implementation in communication networksnecessitates the use of robust learning algorithms, that areagnostic to the network and radio parameters. In this workwe demonstrate the use of generative adversarial learning for



2

the task of robust transmitter identification.Unlike the image or speech processing domain, where ML

techniques have been widely successful, learning in the RFdomain is in a nascent state. Traditional ML techniques cannotbe easily extended to the RF domain because of the unpre-dictable and varied nature of RF signals. Furthermore, thepresence of adversaries make it even more difficult to learn andcharacterize RF signals due to unreliability of the underlyingdata. For example, in order for ML techniques to be effectivefor the task of emitter identification, one must choose anattribute or feature that is unique to a transmitter, irrespectiveof the signals it transmits. A feature that is commonly usedfor this task is the “I/Q imbalance” that is generated by therandom noise introduced into the transmitter manufacturingprocess due to the use of uncharacterized mixers, oscillatorsand unbalanced low pass filters [2]. However, building robustlearning systems using traditional methods that only uses the“I/Q imbalance” is hard due to the underlying characteristicsof the RF channel. This when coupled with the presenceof active adversaries, renders the use of traditional learningalgorithms in RF channels moot.

Inspired by the possibility of using “I/Q imbalance” for thetask of transmitter identification, we explore the idea of auto-matic feature learning using multi-layer neural networks (NNs)to learn deep feature representations that are able to implicitlyexploit this imbalance for transmitter “fingerprinting”. In thispaper, we first demonstrate the use of generative adversar-ial nets (GAN) to disambiguate trusted transmitters fromrogue (fake) ones. After eliminating the rogue transmitters,we use a standard NN model to identify (classify) the trustedtransmitters based on their radio fingerprints. Among the var-ious approaches that can be used to discern this feature space,deep learning (DL [3]) based methods provide an efficient andautomatic way of learning and characterizing the same. Thesemethods can learn and analyze the inherent properties of largedeployments and characterize the associated parameters forautomatic feature learning for the purpose of classification (orregression). Deep neural networks (DNNs) have been shown tobe effective for automatically learning discriminating featuresfrom the data [4] and with proper choice of the NN architectureand associated parameters, arbitrarily good approximations tothe decision boundaries can be computed [5]. Since the task ofclassification is equivalent to learning the decision boundary,we use NNs as the obvious choice for a learning machine.

The main contributions of this paper are:• We propose and implement RFAL using a GAN with

two primary components: i) a generative model thatuses a deep neural network (DNN) for generating fake(aka, counterfeited) signals that closely resembles the realsignals, by deducing the parameter space and replicatingthe time-invariant features and ii) a discriminative modelthat also uses a neural network (DNN) to distinguishtrusted transmitters from rogue ones. During the learningphase, the outcome of the decision process is fed to thegenerative model, allowing the adversary (generator) toupdate its model. The generator thus serves as a compactfront-end for mimicking a transmitter (rogue transmitterin this case).

• Once RFAL detects the trusted transmitters from therogue ones, RFAL uses standard NN models to differenti-ate between the different trusted transmitters. RFAL firstuses a convolutional neural network (CNN) which lever-ages the correlation between the complex-valued I/Q dataconstellations. We also test RFAL using a fully connecteddeep neural network (DNN) that improves upon theaccuracy of the CNN. Finally, a recurrent neural network(RNN) with both long short term memory (LSTM) andgated recurrent units (GRU) is used with RFAL thatexploits the time series properties of the I/Q data.

• Our models have been validated on a laboratory testbedconsisting of several Universal Software Radio Periph-eral (USRP) B210s [6] transmitters and a RTL-softwaredefined radio (RTL-SDR) receiver [7]. The USRPs trans-mitted signals on a particular frequency which werereceived by the RTL-SDR. The generative and discrim-inative models were trained and tested on the collecteddataset which has 1024 complex I/Q values per times-tamp, generating 2048 features. The unique pattern ofvariation of the I/Q imbalances for each radio is capturedby the deep feature embedding learned by multiple layersof the DNN.

• We collect I/Q data from a SDR made by another man-ufacturer, namely ADALM-PLUTO [8]. We show thatthe I/Q imbalance is more pronounced (and thus easierto exploit both explicitly as well as implicitly) whendifferent types of SDRs (from different manufacturers)are used as transmitters.

• We also collect three more datasets of I/Q values from8 USRP B210s with varying signal-to-noise-ratio (SNR).We use distance and multi-path as the defining factors forSNR variation during data-collection.

• We train the proposed models and present a competitiveanalysis of the performance of our models against thetraditional techniques and state-of-the art techniques fortransmitter identification (classification). Results revealthat the proposed methods out-perform the existing onesthus establishing the superiority and usefulness of theproposed models, more so considering the fact that theproposed models do not require any pre-processing of theraw I/Q data that feeds into the NN models.

• The novelty of the proposed work lies in accuratelymodeling and implementing the proposed generative anddiscriminative models on real hardware using raw I/Qdata. To the best of our knowledge, this is the first paperthat uses GANs to identify adversarial RF signals forfingerprinting radio transmitters.

The rest of the paper is organized as follows: Section IIpresents some background on I/Q imbalance and discussessome prior work that are relevant to this paper. The GANarchitecture along with the generator and the discriminatormodels are proposed in Section III. The various NN modelsused are discussed in Section IV. We present the testbedsetup and evaluation framework in Section V. The results arepresented in Section VI and conclusions are drawn in the lastsection.



3

II. BACKGROUND AND RELATED WORK

In this section, we introduce the idea of I/Q imbalance anddiscuss its usefulness for the task of transmitter fingerprinting.We also present the machine learning (ML) techniques thatare relevant to this paper, specifically GAN and NNs. Wealso present existing work that uses these ML techniques fortransmitter identification and classification.

A. I/Q Imbalance

The low price of radio devices comes with a trade-off;namely the presence of “almost undetectable” hardware im-pairments in the radio units due to the use of inexpensivebulk produced commercial off-the-shelf (COTS) componentsduring manufacturing. One such impairment is the imbalancebetween the in-phase (I) and quadrature (Q) components ofthe transmitted signal, commonly known as the “I/Q imbal-ance,” that is unique to different radio hardware and are causedby imperfections in local oscillators and mixers. As a resultof this, the I and Q components of the modulator are notorthogonal. When a signal is transmitted using a particularradio transmitter having an I/Q imbalance, this is imposedover the complex-valued I/Q data that is being transmitted[9], as shown in Fig. 1. The presence of I/Q imbalanceleads to performance degradation for higher order modulationsbecause the symbol rotation becomes more sensitive withincreasing number of constellations, towards both the I andQ components [10].

(a) (b)

Quadra

ture

Inphase

Quadra

ture

Inphase2-2

2

-2

1

-1

0 1-1

0

-1

1

-1

-2-2

2

2

0

0 1

Fig. 1. I/Q Imbalance for QPSK: (a) Before (b) After 45° Phase Imbalance

The number of resources having information about the I/Qimbalance of real systems is limited. Some prior works [10]–[16] on I/Q imbalance estimation and compensation havereported amplitude imbalance test values ranging from 0.02to 0.82 and phase imbalance test values between 2° and11.42° [9]. These estimates can in turn be used to compensatefor the imbalance. In spite of these techniques for compen-sation of the imbalance, the fact remains that all transceiversexhibit this unique I/Q imbalance. Furthermore, the I/Q imbal-ance depends on the choice of the hardware components used,and is an unwanted byproduct of the manufacturing processthat is hard to imitate. Hence the I/Q imbalance can be usedas a basis for feature engineering (implicit or otherwise) fortransmitter identification and recognition.

B. Generative Adversarial Net

The idea of a generative adversarial net (GAN) [17] inmachine learning is based on synergistic application of ideasfrom game theory and unsupervised learning. It consists of

two competing systems: a generator (adversary) and a dis-criminator. The input from the “adversaries” is used to buildrobust discriminative models that can operate in the presenceof real adversaries. The overall training mechanism can beconceptualized as a min-max game with two players, namelythe generator and the discriminator. They help each otherto improve themselves through an iterative training process.Though in theory, the generator and discriminator should playthe game indefinitely, in reality, depending on the ratio of dataand model density, the discriminator overpowers the generatorin a finite amount of time, due to the vanishing gradientof the generator. In practice, this results in generating moreaccurate and robust ML models. Note that the discriminatorimplementation is made deeper than the generator in order toget a purposeful implementation of the GAN framework.

C. Relevant Neural NetworksNeural networks (NNs) are a biologically-inspired learning

paradigm that uses supervised training for building robustlearning models for a variety of tasks and have been shown tobe successful across several application areas. For the task oflearning from RF signal data, we consider the following threelearning paradigms using NNs.

1) CNN: Convolutional neural networks (CNN) are a classof NNs where the architecture operates around the idea oflearning local features from the input (like spatial correlation)through the use of specially designed structures called “filters,”which can be considered as masks that are designed to elicitspecial responses from the input data. Fully connected net-works can be augmented with convolutional layers for fastertraining through learning progressively finer features, that inturn help the network to learn more compact and meaningfulrepresentations. Convolutional layers are like ordinary hiddenlayers, but instead of being fully connected to the input (orthe nodes of the previous layer), they are only connectedto a subset of the input (or previous layer) nodes. CNNshave proven to be effective not only for image and videoprocessing [18], but also in the radio frequency (RF) domain.There have been quite a few attempts at using CNN forlearning different RF parameters [19], [20].

2) DNN: Fully connected Deep Neural Networks (DNNs)are standard NNs but with multiple hidden layers (usuallymore than one hidden layer). Though the idea of NN hasbeen around for quite sometime and there have been sev-eral studies as regards to their capability for approximatingunknown functions, it has been the success of deeper neuralarchitectures, which has brought the field of neural networkpowered Artificial Intelligence (AI) into renewed focus.

DNNs [3] (like other neural networks) compute an embed-ding of the raw inputs in an embedded feature space anduse this representation for learning a model for tasks such asclassification and regression. One of the advantages of usingsuch architectures for building learning systems is the fact thatthe feature space embedding is computed automatically.

Given enough training data to the DNN, it can learn therequired representations without the need of incorporatingprior knowledge from experts, through the use of featureengineering.



4

3) RNN: Automatic transmitter identification using charac-teristics of the transmitted signal is akin to learning the noiseof the transmitter and using it for creating a “fingerprint” thatis used for disambiguation. In order to estimate the noise of aRF system, the learning algorithm needs to “listen” to the un-derlying signal over time and “remember” the same, for noiseestimation. Previously, NNs lacked this capability when usedin the context of temporal data. Another issue was the problemof vanishing gradients, when trying to use back propagationwith temporal data. Both these problems were solved by theinvention of Recurrent Neural Networks (RNN) [3]. RNNs areused for learning and forecasting different features from timeseries data. Long Short Term Memory (LSTM) [21] and GatedRecurrent Unit (GRU) [22] are special types of RNN whichare designed to learn the long term inherent dependencies oftime series data.

D. Related Works

The problem of transmitter classification has been studied inthe past. Here we first discuss a few traditional learning basedapproaches for transmitter classification. Finally, we discussmore recent transmitter identification methods based on theidea of automatic feature detectors (like neural networks). Wealso discuss the advantages of using automatic feature learningtechniques over the traditional ones.

1) Traditional Learning based Techniques: Traditionaltransmitter classification methods are based on statistical learn-ing techniques that use expert engineered features which lever-age some unique characteristics of the transmitters. In [23],the authors proposed a genetic algorithm based solution fortransmitter classification based on transients. A transient signalis transmitted when a transmitter is powered up or powereddown. During this short period (typically few micro seconds),capacitive loads charge or discharge. A genetic algorithmgenerated the “transient times” from 5 different types oftransmitters, which were later classified using a NN modelyielding a 85% - 98% accuracy. It is to be emphasized thatthe experimental results were solely based on the syntheticallygenerated transient values. Though this work used NNs for thefinal classification, the features (transients) were empiricallydetermined and hence we categorize this as an example ofa traditional approach. A multifractal segmentation techniquewas proposed in [24] using the same concept of transients.The segmentation technique extracted significant features fromtransient signals and generated a compact multifactral model.Later a probabilistic NN classifier achieved 92.5% successrate over the extracted transient features in a simulated en-vironment. Another transient based transmitter classificationwas proposed in [25]. A k-nearest neighbor discriminatoryclassifier was used to create a classification engine whichleveraged transient signals for spectral feature selection. Theauthors achieved a 97% accuracy at 30 dB SNR and 66%accuracy at 0 dB SNR for classification of 8 transmitters.

A different approach for transmitter fingerprinting wasproposed in [26], where the authors classified FM radiotransmitters based on unique stray features extracted fromspurious modulation characteristics. The proposed approach

was able to classify samples (20 dB SNR) from 3 FMradio stations with 62%-71% accuracy. This method doesnot provide a competitive accuracy and is also constrainedby the need to have knowledge of modulation technique. Aparticle swarm optimization (PSO) technique was proposedin [27], where two radar transmitter models were classifiedbased on the radar pulse’s time-frequency representation. Anacceptable classification accuracy was reported with 20 dBSNR and relatively low component tolerances. In [28], theauthors proposed a location-based transmitter fingerprintingapproach by extracting signal characteristics (skewness andkurtosis) from wavelet transform. This transmitter locationfingerprint was performed for 4 stationary transmitters inan indoor office environment. In [29], the authors proposeda dimensionality reduction method for extracting intrinsicfeatures from bispectrum information of transmitters. Theyreported 99% accuracy for transmitter identification from thebispectrum matrices. However, deployment of such techniquesin real-time is challenging due to the additional overhead forgenerating the bispectrum before the classifier can be invokedfor identification.

As seen from the above discussion, there are some ad-vantages to using traditional fingerprinting techniques suchas classifying the transmitters based on their unique identi-fications in that we are able to leverage the expertise of the“human-in-the-loop” using such techniques through featureengineering. However, these methods have extra overhead dueto the feature extraction step and furthermore the quality ofthe solution is constrained by the knowledge of the expertmaking such techniques limited in scope. Moreover, as thefeatures are signal and protocol dependent, any change inthe nature of the transmission mandates a change in theunderlying model, thus making them hard to generalize acrossdifferent types of transmissions (having varying protocols,heterogeneous transmitters etc.).

2) Automatic Feature Learning based Techniques: In recentyears, there has been some effort at using automatic featurelearning techniques for fingerprinting RF transmitters. In [19]the authors presented a radio modulation classification methodusing naively learned features. They have shown that blindtemporal learning on densely encoded time series using CNNsis a viable approach. However, this method did not performwell in the low signal to noise ratio (SNR) regime. In [20],the authors have presented an unsupervised learning techniqueusing convolutional autoencoders, to learn the modulationbasis functions and then leverage that to recognize differentdigital modulation schemes. They also proposed and evaluatedquantitative metrics for evaluating the quality of the encodingusing domain relevant performance metrics.

In [30] the authors have demonstrated the use of NN formodulation detection. Apart from the results, an interestingaspect of the work is the way I/Q values were used as input tothe NN. More precisely, given N I/Q values, the authors useda vector of size 2N as an input to the NN, effectively usingthe I and Q components as a tuple representing a point in thecomplex plane. A method for modulation classification wasproposed in [31], for a distributed wireless spectrum sensingnetwork. The authors used a recurrent neural network using



5

long short term memory (LSTM) cells yielding 90% accuracyon a synthetic dataset [32].

An in-depth study on the performance of deep learningbased radio signal classification was presented in [33]. Theauthors considered 24 modulation schemes with a rigorousbaseline method that uses higher order moments and strongboosted gradient tree classification for detection. The authorsalso applied their method to real over-the-air data collected bySoftware Defined Radios (SDRs). In [34] an approach basedon the idea of adversarial learning was proposed for synthesiz-ing new physical layer modulation and coding schemes. Theadversarial approach is used to learn the channel responseapproximations in any arbitrary communication system, en-abling the design of a smarter channel autoencoder. All theseapproaches demonstrate the efficacy of using an “end-to-end”technique based on learning deep feature representations, fordifferent tasks in the RF domain.

There have been quite a few studies that have used CNNbased models for automatic feature learning [35]–[41] forthe task of transmitter classification (or identification). TheCNN models presented in [35], [36], [40] require some pre-processing on the raw signal data before it can be used as in-put. In [38], the authors compared several learning paradigmsfor the task of transmitter identification. More precisely, theylooked at conventional deep neural nets, convolutional neuralnets, support vector machines, and deep neural nets with multi-stage training. They showed that deep neural nets with multi-stage training worked best for the problem and achieved 100%accuracy with a novel dataset having 12 transmitters. On theother hand CNN models were proposed in [37], [39], [41] forexisting datasets (ACARS [42], ADS-B [43], FIT/CorteXlab[44]). However, none of these models directly take the the rawsignal data as input. Note that none of the methods that wehave discussed till now in regards to the task of transmitteridentification are resilient to the presence of active adversaries.This motivated us to propose a robust NN based model whichwould be resilient to the presence of active adversaries and atthe same time provide an end-to-end solution to the transmitterclassification problem.

3) Comparison of Traditional and Deep Learning basedMethods: All the traditional techniques that have been used forRF analysis lack flexibility and robustness. These approachesrequire an expert’s involvement for determining which features(e.g., transients, spurious modulation, etc.) to extract andhow to design an algorithm tuned to that feature. Even ifa feature is identified, it is not necessary that this featurewill be applicable for all scenarios. For example, locationbased fingerprinting [28] will work well for indoor environ-ments but will fail for non-stationary transmitters. However,deep learning (DL) based methods are capable of learningthe features automatically from the data and hence they donot require the feature engineering step. Furthermore it ispossible to use DL techniques in conjunction with adversariallearning to build robust transmitter classification models thatcan function in the presence of active adversaries. In this work,we use the raw I/Q data as an input to the learning model. Themodel automatically discerns the features that can encode theinformation required to disambiguate the transmitters. Note

that the features computed by the DL system can be implicitlybased on the “I/Q imbalance” or some other intrinsic featuresof the transmitter or a combination of these.

4) I/Q Imbalance based Fingerprinting: “I/Q imbalance”based fingerprinting approaches provide more significant dis-criminant information than transient based or modulation-metrics based approaches [45]. Though the use of RF signaldata in general (and I/Q data in particular) with machinelearning algorithms has been limited in the past, more recentlythere has been several applications (see [46] and referencestherein).

An “I/Q imbalance” based fingerprinting approach was pro-posed in [45], where a subclass discriminant analysis (SDA)ML method was used to estimate the distortion parametersfrom the I/Q constellations as features. The proposed methodwas tested on transmitted signals from 7 Time-division multi-ple access (TDMA) satellite terminals, relayed by a transparenttransponder, giving 97% accuracy over 15 dB SNR. However,this method is not guaranteed to capture the difference betweentransmitter’s “I/Q imbalances,” as it is aggregated by theimbalance of the transponder. Similarly, a classifier basedon Gaussian mixture models (GMM) was proposed in [47].Though, the study showed ∼100% accuracy, the experimentswere conducted on artificial data. A simulation-based trans-mitter authentication scheme was proposed in [48] using an“I/Q imbalance” matrix and multiple collaborating receivers.After analyzing these existing DL techniques, it is evident thatthere is still a lack of systematic “end-to-end” approaches, thatcan use the raw signal data from real transmitters and exploitthe I/Q imbalance for fingerprinting. It must be noted that DL-based RF methods will not only exploit the “I/Q imbalances”but also extract and use other intrinsic features related to thetransmitters that may or may not be directly related to I/Qimbalance. However, the end product will conceptually beable to differentiate between transmitters having different “I/Qimbalances.”

Motivated by the capacity of deep learning systems toautomatically learn deep discriminative features, we focus onthe problem of transmitter identification in the presence ofadversaries using a generative adversarial network (GAN).The idea of training discriminative models via an adversarialprocess was first proposed by Goodfellow [17]. Since then,GANs have been adopted for solving problems in varied fieldsof applications and particularly for image processing whereGANs have proved to be highly efficient for several differenttasks [49]–[51]. We take inspiration from [30] for using theraw I/Q data for input to our networks.

III. GAN FOR ROGUE TRANSMITTER DETECTION

As pointed out earlier, most of the traditional ML tech-niques are susceptible to malicious attacks. The suscepti-bility increases once the attacker knows the features usedby the learning algorithm. With the knowledge of the fea-ture space (and hence the underlying feature distribution)the attacker becomes smart enough to mislead the learningprocess. Thus, for the problem of transmitter identification,the target of the adversaries would be to learn the probability



6

distribution of the training data used for model creation, givena sample of the same. With this knowledge, the adversariescan use a generative model to generate signals so as to spoofthe transmission of known transmitters. A GAN [17] usesa generative model which enables the realistic creation ofsamples from a given distribution which can then be usedto train a discriminator for identifying real samples fromfalse/counterfeited ones obtained from the generator. The GANtraining makes the model resilient over the trained adversarialdata and thus intuitively helps it to be prepared for the “yet-to-be-seen” adversaries.

A. Proposed GAN Architecture

The proposed GAN framework, as shown in Fig. 2, has twoprimary components: a generative model (G) that generatesfalse data using a given data distribution and a discriminativemodel (D) that estimates the probability that a sample camefrom the training data (that is known transmitters) ratherthan G. The adversary generates random modulation scheme(m(t)), signal amplitude (r(t)) and phase (l(t)) and mixesadditive white Gaussian noise (AWGN) (n(t)) with the falsesignal. The generated signal (g(t)) which is initially randomin nature improves over time as the generator learns from thediscriminator and improves on its ability to imitate real data.On the other hand, the discriminator (D) gets input from boththe generator (G) and Trusted transmitters. This helps it tolearn to differentiate between real and false inputs. The knowntransmitter data is collected and fed to the discriminator (D)as raw I/Q values.

Overall, our objective is to train G in such a way that willmaximize the probability of D making a mistake. G tunesits hyper parameters with the feedback from D. We arguethat GAN is an efficient way to generate correlated data sam-ples and thereby approximate an accurate generative model,something the adversary aims to achieve. Once the modelis trained, RFAL synthesizes signals using the generator tomimic adversarial transmitters based on the learned probabilitydistribution on the sample space of I/Q signal data from theknown transmitters.

Modulation Schemes

Signal Amplitude

Signal Phase

Generator

Noise

D

G

Counterfeited?

Trusted?

I/Q Data(Through SDR)

Feed

backm(t)

r(t)

l(t)

GAN Implementation

n(t)

g(t)

Feature Extractor

Adversarial

Trusted Transmitter

(I/Q Generator)

s(t)

Fig. 2. Proposed RFAL GAN architecture

B. The Generative Model

In order to build the generator, we treat the overall problemas an N -class decision problem where the input is a complexbase-band time series representation of the received signal.

Sample Space

Noise

GGenerator

RealData

DDiscriminator

Is Real?

Tuning Parameters

Fig. 3. A Simplified View of GAN Implementation

That is, the training data consists of the in-phase and quadra-ture components of a radio signal obtained at discrete timeintervals through analog to digital conversion with a carrierfrequency to obtain a 1×N complex valued vector. Classically,this is written as:

s(t) = c1m(t) + c2r(t) + c3l(t) (1)

where s(t) is a continuous time series signal modulated ontoa sinusoid with either varying frequency, phase, amplitude,trajectory, or some permutation of these parameters. Here,m(t), r(t), and l(t) are the time series continuous signalsfor modulation, amplitude, and phase respectively, selectedrandomly by the generator. The coefficients c1, c2, and c3 aresome path loss or constant gain terms associated with m(t),r(t), and l(t) respectively. The output g(t) is obtained as:

g(t) = s(t) + n(t) (2)

where n(t) is the AWGN. The output g(t) is then fed toa generator which is used as an unsupervised learning toolas a part of the GAN framework. The generator learns theprobability distribution pg(x) over sample space (x) of theI/Q input.

C. The Discriminative Model

The discriminative model learns by minimizing a costfunction during training. The cost function, C(G;D), dependson both the generator (G) and the discriminator (D). It isformulated as C(G;D) = Epdata(x) logD(x) + Epg(x) log(1−D(x)), where pg(x) is the generator’s distribution over x,pdata(x) is the data distribution over x, D(x) is the probabilitythat x came from pdata(x) than pg(x) [17]. The training isformulated as:

maxD

minGC(G;D) (3)

For the GAN framework, there is an unique optimal discrim-

inator for a fixed generator, D∗(x) =pdata(x)

pdata(x) + pg(x)[17].

It is also inferred that G is optimal when pg(x) = pdata(x),i.e., the generator is optimal when the discriminator cannotdistinguish real samples from false ones. Similarly, the D isoptimal when the discriminator can recognize each real samplefrom the false ones.

D. GAN Implementation

For implementing the GAN, we use “over-the-air” signaldata collected from the trusted transmitters. (The testbedsetup is discussed in Section V.) The generator (G) generatescounterfeit data from the same sample space to impersonate



7

(None, 2048)

Inp

ut

(None, 1024)

Dense

tanh

(None, 512)

Dense

tanh

(None, 2)

Dense

2

Outp

ut

Discriminative Net

Signal Processing andData Collection

Noise

(None, 100)

(None, 512) (None, 1024) (#samples,2*sample size)

Generative Net

OR

Tru

sted

Data

Fake Data

[0,1] OR [1,0]

Rand

om

Dense

tanh

Dense

tanh

Genera

ted

Data

Softmax

Dro

pout

= 0

.5

Dro

pout

= 0

.5

Fig. 4. GAN Implementation for Rogue Transmitter Detection

a trusted transmitter. True and counterfeit I/Q data are fed tothe discriminator (D) with equal probability. We design thediscriminator and generator separately, as shown in Fig. 3.The overall GAN implementation is shown in Fig. 4.

The generator starts by generating random data within(−∞,+∞). As the training evolves, the generator learns thatthe sample space of real data is [−1, 1]. Thus, the generatorwill gradually generate I/Q values within [−1, 1] in turn re-ducing the parameter space. The time-invariant features beingautomatically learned by the discriminator should implicitlycapture the inherent imbalance within the I/Q data. The gen-erator gradually learns the real data distribution over multipletraining epochs and starts to replicate the I/Q imbalances inthe generated I/Q values.

Once the initial random values are generated, they arepassed through two dense layers of size 512 and 1024 re-spectively with tanh [18] activation. Then a single dense layerof twice the sample size (2048 in our case) is invoked withthe sigmoid activation function [18]. G continues to learn thedata distribution (pg) and generates counterfeit samples of size2048 within the sample space of the I/Q values.D consists of one input layer of 2048 nodes, two hidden

layers of 1024 and 512 nodes respectively, and finally asoftmax output layer of 2 nodes to classify an input as eitherCounterfeited or Trusted. We used tanh as activationfunction at the hidden layers and added Dropout [52] of0.5 in between these layers for regularization. We train boththe generator and discriminator through iterative sequentiallearning to strengthen the generative model over time.

Once the discriminator recognizes the trusted transmittersfrom the counterfeit ones, we feed the trusted transmitterdata into another DNN (either convolutional or fully con-nected (dense)) for further classification into a number ofclasses (as determined by the number of trusted transmitters).Next we discuss the NN architectures we use for this purpose.

IV. PROPOSED NEURAL NETWORKS FOR TRANSMITTERCLASSIFICATION

Recent advances in NNs have made it possible to obtainrobust models with low generalization errors by training“deep” neural architectures efficiently. The “depth” signifiesthe number of iterative operations performed on the input datausing each layer’s transfer function and deeper architecturesallow the network to learn more robust feature representationsfrom the input data. Though such techniques require highercomputation power and involve complicated layer-by-layerback-propagation, nevertheless, most DL systems are able to

efficiently learn deep feature representations from the trainingdata using back-propagation and some variation of gradientdecent, with adaptive learning rates (e.g., Adam [53]) andregularization to avoid overfitting (e.g., Dropout [52]). Next,we present our proposed NN models for classification ofTrusted transmitters. We use a total of 8 transmitters, thedetails of which are given in Section V.

A. CNN Model

In order to capture the correlation between I/Q values,we started by implementing a convolutional neural net-work (CNN). We implement the CNN with three conv2Dlayers with 1024, 512 and 256 filters respectively, a Flattenoperation and three fully connected (FC) layers of size 512,256 and 8 nodes [54] respectively, as shown in Fig. 5. We useDropout [52] of 0.25 and 0.5 after each conv2D and denselayer respectively. We also use kernel size of (2,3) and strideof (2,2) at each of the Conv2D layers. We apply a poolinglayer MaxPooling2D after each conv2D layer with pool sizeof (2,2) and stride of (2,2). We use ReLU [55] activation forall convolution and fully connected layers, other than the last,where we use softmax. Note that, the number of nodes inthe last layer is changed based on the number of classes thedataset has. We use Adam [53] (with learning rate of 10−3)based optimization with categorical cross-entropy training.

It is to be noted that we design the CNN with only 3convolution layers and 4 fully connected layers for fastertraining [56], since no significant increase in the testingaccuracy was observed after increasing the number of layers.

(None, 1024, 2)

Inp

ut

Flatten

8 8

Outp

ut

Convolution Neural Net

Signal Processing and

Data Collection

Conv

Layer

1

Conv L

ayer

2

Conv L

ayer

3

FC FC FC

1024 512 256

512 256

Data

Fig. 5. CNN Implementation for Transmitter Classification

B. DNN Model

Apart from the CNN discussed above, we also use a fullyconnected (dense) DNN for the task of trusted transmitterclassification. The implementation of the DNN is similar tothe discriminator model of GAN and is shown in Fig. 6. Theonly difference is that the softmax output layer has 8 nodes torecognize the 8 transmitters (or more generally k nodes if there



8

are k transmitters). We implement a DNN with 5-layers with1 input layer and 4 dense layers. We use tanh [18] activationfunction for the Dense layers in this model. We apply biasesand regularization to avoid under- and over-fitting. In this casealso, use Adam [53] based optimization with learning rate of10−3, for categorical cross-entropy training.

(None, 2048)

Input

(None, 1024)

Dense

tanh

(None, 512)D

ense

tanh

(None, 8)

Dense

8

Outp

ut

Deep Neural Net


Data Collection

Softmax

DataD

rop

out

= 0

.5

Dro

pout

= 0

.5Fig. 6. DNN Implementation for Transmitter Classification

C. Recurrent Neural Network (RNN)

In order to exploit the the time-series property of thecollected data, we use RNNs with LSTM and Gated Recur-rent Unit (GRU) cells, as both avoid the “vanishing” or the“exploding” gradient problems [57].

1) Long Short Term Memory (LSTM) Cell Model: ThoughLSTM cells can be modeled and designed in various waysdepending on the need, we implement the cells as shown inFig. 7. In each LSTM cell, there are (i) three types of gates:input (i), forget (f ) and output (o); and (ii) a state update ofinternal cell memory. The most interesting part of the LSTMcell is the “forget” gate, which at time t is denoted by ft. Theforget gates decide whether to keep a cell state memory (ct)or not. The forget gates are designed as per the equation (4)on the input value of xt at time t and output (ht−1) at time(t− 1).

ft = σ(Wxfxt +Whfht−1 + bf ) (4)

where, σ denotes the sigmoid activation function, Wxf and bfrepresent the associated weight and bias respectively, betweenthe input (x) and the forget gate (f ). Once ft determines whichmemories to forget, the input gates (it) decide which cell states(ct) to update as per equations (5) and (6).

it = σ(Wxixt +Whiht−1 + bi) (5)

ct = tanh(Wxcxt +Whcht−1 + bct−1) (6)

In equation (7), the old cell state (ct−1) is updated to the newcell state (ct) using forget gates (ft) and input gates (it):

ct = ft ◦ ct−1 + it ◦ ct (7)

where, ◦ is the Hadamard product. Finally, RFAL filters theoutput values through output gates (ot) based on the cell states(ct) as per equations (8) and (9).

ot = σ(Wxoxt +Whoht + bo) (8)

ht = ot ◦ tanh(ct) (9)

We design and implement a RNN with LSTM, the structureof which is shown in Fig. 8. We use 2 LSTM layers with 1024and 256 units sequentially. Next we add 2 fully connected

LSTM Cell

σ tanh

σ σ

(xt,ht-1)

it

ft ot htct-1

ĉttanhct

σtanh

sigmoid activation

tanh activation

Hadamard product

sum over all elements

Fig. 7. LSTM Cell Architecture Used in the RNN Model

layers with 512 and 256 nodes respectively. We use batchnormalization on the output and pass it through a dense layerof 8 nodes. We use ReLU [55] as activation function for theLSTM layers and tanh [18] as activation for the dense layers.Lastly, we use stochastic gradient descent (SGD) [18] (withlearning rate of 10−3) based optimization with categoricalcross-entropy training.

(None, 2048)In

put

8 8

Outp

ut

Recurrent Neural Network


Data Collection

LSTM

Layer1

LSTM

Layer

2

FC FC FC

1024 256512 256

Data

Dro

pout=

0.5

Dro

pout=

0.2

Dro

pout=

0.1

Batc

hN

orm

aliz

ati

on

Fig. 8. RNN Implementation for Transmitter Classification

D. Gated Recurrent Unit (GRU) ModelThe main drawback of using LSTM cells is the need for

additional memory. GRUs [22] have one less gate than LSTMsfor the same purpose, thus having a reduced memory and CPUfootprint. The GRU cells control the flow of information justlike the LSTM cells, but without the need for a memory unit.It just exposes the full hidden content without any control. Ithas a “reset gate” (zt), an “update gate” (rt), and a cell statememory (ct) as shown in Fig. 9. The reset gates determinewhether to combine the new input with a cell state memory(ct) or not. The update gate decides how much of ct to retain.The equations (10), (11), (12), and (13) related to differentgates and states of GRU are given below.

zt = σ(Wxzxt +Whzht−1 + bz) (10)

rt = σ(Wxrxt +Whrht−1 + br) (11)

ct = tanh(Wxcxt +Whc(rt ◦ ht−1)) (12)

ht = (1− zt) ◦ ct + zt ◦ ht−1 (13)

We implemented the recurrent model with GRU cells andused the same architecture as the LSTM implementation (GRUcells instead of LSTM cells at the first two layers). The pro-posed GRU network needs fewer parameters than the LSTMmodel. In this case, we also use SGD [18] based optimizerwith a learning rate of 10−3. An quantitative comparison ofthe results is discussed in Section VI-F.



9

GRU Cell

σ σ

(xt,ht-1)

zt

httanh ct

σtanh

sigmoid activation

tanh activation

Hadamard product

sum over all elements

W weight multiplication

rt

W

W

1-

xt

ht-

1 ht-1

Fig. 9. GRU Cell Architecture Used in the RNN Model

V. TESTBED SETUP AND EVALUATION

In order to validate the proposed models, we wanted touse our own “over-the-air” RF data collected from differenttransmitters, instead of synthetic or publicly available data. Weimplemented the generator and discriminator (of the GAN) todetect an unknown (adversarial) transmitter and then used theNN models to classify the known ones, the details of whichare discussed next.

A. Signal Generation and Data Collection

In order to learn the discriminating fingerprints (features)of similar transmitters, we used 8 universal software radioperipheral (USRP) radios of the same kind, namely B210 fromEttus Research [6]. The overall setup for signal generation andreception is shown in Fig. 10. The B210s were programmedto transmit random data on 904 MHz using Quadrature PhaseShift Keying (QPSK) modulation. We used GNURadio [58]for signal processing and data transmission. The flow graphis presented in Fig. 11. The modulated signal was transmittedthrough the USRP sink block. For the receiver, we used aRTL-SDR [7] which captured “over-the-air” raw I/Q data andwrote it onto a file.

B. Analysis of Data Collection Environment

We set up the data collection testbed in an indoor labenvironment with a direct line of sight between the transmitterand the receiver with a distance of 10 ft. Thus, the underlyingchannel can be modeled as a Rician fading channel. There wasalso multi-path effects due to the reflections from the walls. Wemeasured the signal to noise ratio (SNR) using a RTL-SDR [7]dongle and Spekrtum [59] which is an open source spectrumanalyzer available for both Windows and Linux. We decidedto calibrate the SNR using the Spekrtum software (rather thanusing a spectrum analyzer) due to cost and portability issues.We found that the noise floor in the lab was between -20 dBand -30 dB. The signal strength for the 200 KHz (from 903.9MHz to 904.1 MHz) channel was between 0 dB and 10 dB. Weset the transmitter gain to 45 dB and calculated the SNR as thedifference between the noise floor and the signal strength. Ourcalculated SNR was 5 dB - (-25 dB) = 30 dB, with a 45 dBtransmitter gain. It is to be noted that the signal strengths (in

dB) of noise and signal measured by the Spektrum is relative,but the difference between them is absolute.

We also collected data for different SNR values by varyingthe transmitter-receiver distance and hindering the line-of-sightin the laboratory. As mentioned earlier, we obtained a SNR of30 dB by keeping the transmitter and receiver at 10 feet fromeach other. Similarly we collect 3 more datasets with SNRof 20 dB, 10 dB and 0 dB at 20 feet, 30 feet, and 45 feetrespectively. It is to be noted that the SNR of the transmitterfrequency was measured at the receiver with Spektrum.

RandomSignal

QPSKMod Transmitters Receiver

DatasetGeneration

GNURadio USRP B210/ PLUTO

RTL-SDR rtlsdrPythonLibrary

Data

(#samples,2*sample size)

Signal Processing and Data Collection

Fig. 10. Signal Generation and Data Collection Setup

Fig. 11. GNU Radio Flow Graph for Data Collection for USRPs

C. I/Q Datasets

We collected raw I/Q signal data with a sample size of 1024,i.e., each sample consists of 1024 I and 1024 Q values. Thechoice of 1024 as the sample size was sufficient to capturethe unique pattern of I/Q imbalances and at the same timeit was not computationally expensive. We experimented withother sample sizes as well: smaller sample size yields degradedperformance whereas larger sample size does not improve theaccuracy. We collected 40,000 training samples from eachtransmitter to avoid the data skewness problem observed inML. The configuration parameters used are given in Table I.We collected two different datasets at 30 dB SNR and threedatasets with three different SNRs, as discussed below.

1) Homogeneous Dataset: For the “homogeneous” dataset,we use only one type of radio, namely, the USRP B210 fromEttus Research. We collected two sets of data: (i) using 4USRP B210 transmitters: 6.8 GB size, 160K rows and 2048columns and (ii) using 8 USRP B210 transmitters: 13.45 GBsize, 320K rows and 2048 columns. Note that the SNR was30 db.



10

Parameters ValuesTransmitter Gain 45 dB

Transmitter Frequency 904 MHz (ISM)Bandwidth 200 KHz

Sample Size 1024Samples/Transmitter 40,000

# Transmitters 2 to 8TABLE I

TRANSMISSION CONFIGURATION PARAMETERS

2) Heterogeneous Dataset: In order to investigate theperformance of the proposed classification methods, whendifferent types of transmitters (from different manufacturers)are present, we use PLUTO SDR [8] apart from the B210swhen collecting the data. The GNURadio flow graph for signalgeneration is similar to Fig. 11 with a different sink block forthe PLUTO SDR. Note that the SNR in this case was also30 db. The ‘heterogeneous’ datasets were obtained using (i) 2USRP transmitters: 3.31 GB size, 80K rows and 2048 columnsand (ii) 1 USRP B210 and 1 PLUTO transmitter: 2.85 GB size,80K rows and 2048 columns.

3) Varying SNR Datasets: We collected 3 more datasetswith 8 USRP B210 transmitters and SNRs of 20 dB, 10 dB,and 0 dB respectively. Each dataset is of size ∼13 GB with320K rows and 2048 columns.

D. Correlation in Dataset

Correlation between each data sample plays a crucial rolein transmitter classification. Given T training samples (for Ttimestamps) and a sample size of M for each time stamp,where each sample is a vector (I,Q) ∈ C representinga number in the complex plane, we create a new vectorX(t) = [Ii, Qi; i = 1, 2, · · · ,M ]t ∈ C2M ; t = 1, 2, · · · , T fortimestamp t, and use it as an input to the NN. As mentionedbefore in our case (M = 1024). Thus we represent the Iand Q values of each training sample at timestamp (t) as:[I0Q0I1Q1I2Q2I3Q3I4Q4 · · · I1023Q1023]t. We used QPSKmodulation, which generates a constellation plot like Fig. 1.This signifies that the correlation should be between everyfourth value, i.e., between I0 and I4, and Q0 and Q4 and soon. Hence we calculate the correlation coefficient of betweenI0I1I2I3, I4I5I6I7 and similarly, between Q0Q1Q2Q3 andQ4Q5Q6Q7 and so on. We take the average of all thecorrelation coefficients for each sample.

We use numpy.corrcoef for this purpose which uses Pearsonproduct-moment correlation coefficients, denoted by r. ThePearson’s method for a sample is given by:

r =

(M−1)∑i=0

(Ii − I)(Qi − Q)√(M−1)∑i=1

(Ii − I)2

√(M−1)∑i=0

(Qi − Q)2

(14)

where, M is the sample size, Ii and Qi are the sample values

indexed with i. The sample mean is I =1

M

(M−1)∑i=0

Ii and

similarly for the Q values.The correlations for all the 40,000 samples for each trans-

mitter are shown in Fig. 12. We observed that around 75%

of the samples’ correlation coefficients are between −0.1 and0.1 and the remaining 25% are close to 0.9. However, fortransmitter ID 3, all the samples’ correlation coefficients arebetween −0.1 and 0.1. This behavior of transmitter ID 3 isan obvious example of manufacturing differences. However,this observation gives us intuition about the difficulty of CNNwith 2D convolutions for the task of transmitter classification.CNNs capture the correlation (local features) in the data.However, in this case, as the nature of the correlations arethe same for all but one transmitter, hence there is not enoughdiscriminative information in the correlations to disambiguatebetween all the transmitters, thus leading to the conclusion thatCNNs with simple two dimensional convolutions will not beeffective for the classification task. This was later corroboratedvia our testbed implementation (Section VI).

(a) Trans ID 1 (b) Trans ID 2

(c) Trans ID 3 (d) Trans ID 4

(e) Trans ID 5 (f) Trans ID 6

(g) Trans ID 7 (h) Trans ID 8

Fig. 12. Correlation Plot for Different Transmitters in the Collected Dataset



11

E. Machine Learning Libraries Used

There are several libraries and tools that implement learningframeworks with support for immensely concurrent GPUarchitectures, that reduce the burden of programming thetraditional GPU routines for training of larger NNs. Weuse Keras [54] as the frontend with Tensorflow [60] as thebackend. Keras is an overlay on neural network primitives withTensorflow [60] or Theano [61] that provides a customizableinterface for quick deployment of complex NNs. We also useNumpy, Scipy, and Matplotlib Python libraries.

F. Performance Metric

To measure the effectiveness of any NN architecture, “clas-sification accuracy” is used as the typical performance met-ric. However, “classification accuracy” as it is defined cansometimes be misleading and incomplete when the data isskewed. A confusion matrix overcomes this problem by show-ing how the classification model performs when it comes toerroneous detections (false alarms), and correct “counterfeit”classifications. It provides more insights on the performance byidentifying not only the number of errors, but more importantlythe types of errors. As a result we use confusion matrices todisplay and analyze the results of our experiments.

VI. IMPLEMENTATION RESULTS AND DISCUSSIONS

In this section, we present the results of i) adversarialtransmitter detection using GAN and ii) transmitter classi-fication using different NN architectures. We conducted theexperiments on a Ryzen 8 Core system with 64 GB RAMand a GTX 1080 Ti GPU unit with 11 GB memory. For thesake of being robust and statistically significant, we presentthe experimental results for each model after several runs ofeach implementation. We focused on four main aspects:• Implementing a GAN to distinguish adversarial transmit-

ters from trusted ones.• Implementing a CNN with 2D convolutions to exploit

the correlation in collected signal data of the trustedtransmitters for trusted transmitter identification.

• Implementing a DNN to classify the trusted transmitters.• Implementing RNN with both LSTM and GRU cells to

improve the accuracy of trusted transmitter classificationby exploiting the temporal aspect of the signal data.

A. GAN Results

In order to detect the adversarial transmitters, we imple-mented a GAN based model as described in Section III-D.We used categorical cross-entropy training and Adam [53] forgradient based optimization. We notice that the discriminatorwas able to detect the adversarial transmitters with 50%accuracy before the GAN based training. Once the GAN goesthrough several epochs (< 50) of adversarial training, the op-timal discriminator (D∗) is able to detect the Counterfeittransmissions with about 99.9% accuracy as shown by thereceiver operating characteristic (ROC) curve and confusionmatrix in Fig. 13. Note that one epoch consists of a forwardpass and a backward pass through the GAN over the entire

Dataset SNR #Trans #Parameters Acc (%)3.6 M (G)

6.8 GB 30 4 6.8 M (D) 99.910.4 M (GAN)

3.6 M (G)13.45 GB 30 8 6.8 M (D) 99.9

10.4 M (GAN)3.6 M (G)

13.45 GB 0 8 6.8 M (D) 99.910.4 M (GAN)

TABLE IIACCURACY OF GAN FOR 4 AND 8 TRANSMITTERS

dataset. It is clear from the confusion matrix that the number offalse negatives and false positives are very low and well withinan acceptable range [18]. The testing accuracy of the GANimplementation on datasets with different SNRs is presentedin Table II. The “parameters” represent the total number ofhyper-parameters required for the respective model.

(a) ROC Curve (b) Confusion Matrix

Fig. 13. ROC Curve and Confusion Matrix of Counterfeit TransmitterDetection from RFAL

In Fig. 14, we present three plots to represent how theproposed generator behaves. Note that we first show the plotsusing randomly selected 128 samples from both the real andgenerated datasets. We choose the number 128 as it is also usedas the batch size for the training. In Fig. 14(b), we see thatthe data generated before the GAN training does not representthe distribution of the real data as shown in Fig. 14(a). It isevident that initially the I/Q values are randomly generatedbetween 0 and 1. However, once the generator is trained overmultiple iterations, it starts generating more realistic data asshown in Fig. 14(b) which shows that the distribution of thegenerated data is starting to resemble the real data shown inFig. 14(a). Once the GAN training converges, the generator haslearned the data distribution of the I/Q values from the knowntransmitters and hence it starts to generate counterfeit data thatsuccinctly captures the actual distribution of the I/Q values.Thus now the generated I/Q values are distributed within therange of [-1,1] as is the case with the real data. Figure 15shows the plots for the real data and the generated data afterfull GAN training. These images are obtained by plotting 2000I/Q samples. Also we do not discuss the theory of why thegenerator is able to generate realistic data because though thereis some idea in the research community as to why GANswork, it is not fully understood yet. Furthermore, since thisis a more applied work we refrain from adding such theoryinto the paper. Rather we provide references which interestedreaders may consult to learn more about the GANs.



12

(a) Real Data (b) Generated Data be-fore GAN Training

(c) Generated Data afterGAN Training

Fig. 14. Output from the Proposed Generator (Plot for 128 samples)

(a) Real Data (b) Generated Data afterFull GAN Training

Fig. 15. Final Output from the Proposed Generator (2000 samples)

Note that though this paper is not about trusted transmitterclassification (but rather about using adversarial learning usingGANs for identifying rogue transmitters), we decided toexplore the capability of systems based on neural networksfor the task of trusted transmitter classification as well. Nextwe present the results of this endeavor.

B. CNN Results

Once the rogue transmitters are detected and eliminated,our system classifies the “trusted” transmitters using a neu-ral network. First we used a CNN built according to theimplementation details provided in Section IV-A for trustedtransmitter classification. We obtain 89.07% and 81.6% accu-racy for 4 and 8 transmitter classifications respectively. Theaccuracy plots and confusion matrices for both the cases arepresented in Figs. 16 and 17. We note that both the trainingand validation accuracy increases with the number of epochs.However for our CNN implementation the number of falsepositives and false negatives are somewhat high. Intuitively,this shows that the convolutional filters that were used with thenetwork were not able to identify and encode discriminativefeatures for this task. Since we know that there is at leastone discriminative characteristic that distinguishes betweenthe transmitters (namely the I/Q imbalance), we conclude thatthe input representation that we used with the CNN did noteffectively encode these characteristics and hence the systemcould not efficiently learn them.

In order to understand the inner workings of the CNN better,we present the feature maps obtained from the first convolutionlayer of the CNN, for the case of four transmitter classification,in Fig. 18. It can be seen that each of these feature mapsencodes a different pattern, one for each of the four differenttransmitters. However, we also notice that the convolutionsfail to capture the highly discriminative patterns from the I/Qdata. Thus for example, none of the feature maps resemble

(a) 4 Transmitters (b) 8 Transmitters

Fig. 16. Accuracy Plot for Transmitter Classification using CNN


Fig. 17. Confusion Matrix for Transmitter Classification using CNN

feature characteristic of the data shown in 15(a). Thus eventhough the features are different for each transmitter, they donot fully encode discriminative features present in the I/Qsamples. Intuitively this is because of the fact that the I/Qsamples does not have significant spatial correlation that canbe leveraged through the use of the convolution operation.Since further tuning of the CNN parameters was unsuccessful,we decided to build a fully connect DNN to try and achievebetter accuracy for this task.

C. DNN Results

To overcome the deficiencies of the CNN, we use a DNNas described in Section IV-B. The DNN yields an accuracy of96.49% for 4 transmitters and 94.60% for 8 transmitters. Theaccuracy plots and confusion matrices are shown in Figs. 19and 20 respectively. It is evident that the number of falsepositives and false negatives in the confusion matrices aresignificantly low for the DNN as compared to the CNN andthus intuitively the DNN can capture and learn better featuresfrom the I/Q samples than the CNN for the task of transmitterclassification. Note that from the perspective of the featureslearned by the DNN, it is hard to explain the types of featuresthat are learned by these systems. However since the task ofclassification boils down to learning a decision boundary andrecently it has been shown that DNNs are capable of learningapproximations of such boundaries efficiently [5], the possiblereason for the better performance of the DNN might be tiedto learning a better approximation to the underlying functionrepresenting the decision boundary for this task.



13

(a) Radio 1

(b) Radio 2

(c) Radio 3

(d) Radio 4

Fig. 18. Feature Maps for the First Convolution Layer of the Proposed CNNModel


Fig. 19. Accuracy Plots for Transmitter Classification using DNN


Fig. 20. Confusion Matrices for Transmitter Classification using DNN

D. RNN (with LSTM Cells) Results

For the RNN, we first implement the LSTM cells asdescribed in Section IV-C1. We achieved 97.40% and 95.78%testing accuracy for 4 and 8 transmitters respectively. Theaccuracy plots and confusion matrices are given in Figs. 21and 22 respectively.


Fig. 21. Convergence of Transmitter Classification using LSTM Cells


Fig. 22. Confusion Matrices for Transmitter Classification using LSTM Cells

E. RNN (with GRU Cells) Results

Finally, we implement RNN with GRU cells as describedin Section IV-D. We achieved 97.85% and 97.06% testingaccuracy for 4 and 8 transmitters respectively. The accuracyplots and confusion matrices are shown in Figs. 23 and 24respectively. It must be noted that the GRU implementationachieves better accuracy than the one using LSTM cells.


Fig. 23. Convergence of Transmitter Classification using GRU Cells

During the training phase, the hyper-parameters get adjusteddepending on the categorical cross-entropy loss. Sometimes,with such adjustments, the model tends to over-fit the training



14


Fig. 24. Confusion Matrices for Transmitter Classification using GRU Cells

data. To avoid such scenarios, we used Dropout [52] regular-ization in our models. Another way to monitor and possiblyavoid over-fitting is through the use of cross validation duringthe training phase. This way, we ensure that the proposedmodel gets trained fairly so that it generalizes well duringthe testing phase. Note that the fluctuations in the validationcurve show that for certain epochs with a particular set ofhyper-parameter values, the model tends to over-fit the data,but in later epochs the hyper-parameter values get adjustedso as to counter the effect of the overfitting. Furthermore, thetraining can be tweaked using the results of the validationphase. Note that since the I/Q samples represent a time seriesdata, it is natural to model them using a RNN. Since RNNslearn the temporal correlation between the time series data,they can encode the transmitter specific variations in the I/Qdata and intuitively this makes the RNNs better at the task oftrusted transmitter classification than the CNNs.

F. Classification Comparison of CNN/DNN/RNN

Once a transmitter is found to be Trusted via the pro-posed GAN, we used a CNN, DNN, and RNN respectively, touniquely identify it. We used 90%, 5%, and 5% to train, vali-date and test respectively. The overall accuracy of the differentNNs for the task of trusted transmitter classification is shownin Table III. We see that the CNN does not exhibit the bestperformance for transmitter classification, which intuitively isdue to the lack of distinguishing spatial correlation in thedata from each transmitter. In Fig. 26, we show a graphicalrepresentation of the classification accuracy obtained using thedifferent neural network architectures.

We present the observed empirical training time for eachmodel as the last column in Table III. It is evident that thetraining time for the CNN is almost double compared to therest of the proposed models, as convolution operations aresignificantly more complex than other neural network compu-tations. We also conducted experiments by varying the numberof transmitters from 2 to 8, using the proposed DNN. In Fig.25, we show how the training and testing accuracy varies as thenumber of transmitters are increased. We note that the trainingaccuracy decreases when there are more classes (more numberof transmitters). However the testing accuracy of the model isstable and does not change significantly with the increase inthe number of transmitters. This establishes the efficacy ofthese methods for the task of transmitter classification, as for

#Trans Models #Parameters Acc (%) Time (min)4 CNN 38 Million 89.07 ∼254 DNN 6.8 Million 96.49 ∼124 RNN - LSTM 14.2 Million 97.40 ∼124 RNN - GRU 10.7 Million 97.85 ∼128 CNN 38 Million 81.59 ∼308 DNN 6.8 Million 94.60 ∼158 RNN - LSTM 14.2 Million 95.78 ∼168 RNN - GRU 10.7 Million 97.06 ∼16

TABLE IIICOMPARISON OF THE VARIOUS IMPLEMENTATIONS

Models #Layers Learning Rate Batch Size Epochs Optimizer

CNN 7 10−4 128 45-50 Adam [53]DNN 5 10−3 128 35-40 Adam [53]

RNN-LSTM 6 10−3 128 30-35 SGD [18]RNN-GRU 6 10−3 128 30-35 SGD [18]

TABLE IVCOMPARISON OF CONFIGURATION SETTINGS FOR DIFFERENT MODELS

production systems one wants the test accuracy to be stableno matter the number of classes involved in the problem.

In our testbed evaluation one of the goals was to exploredifferent types of NNs to find the architecture (and model)having the best possible accuracy within the constraints ofthe training time (which was not more than 30 minutesin our test-bed setup for all models). The performance ofany NN architecture depends, among other things, on thevalues of the hyper-parameters. Furthermore, as a given hyper-parameter setting may be optimal for one network but notfor another, we used different hyper-parameter values to traindifferent networks. Intuitively, this dependence on the valuesof the hyper-parameters is due to the fact that the underlyingoptimization problem and hence the solution space, is differentfor different types of networks. We observed that decreasingthe learning rate of the CNN to 10−4 increases the accuracy by7-8%. However, decreasing the learning rate to less than 10−3

for RNN and DNN does not improve the testing accuracy bya significant amount (even by 0.5%). In Table IV we recordthe values which gave the best possible accuracy under theconstraints of the training time. During each training, we setthe maximum epoch to 50 with an early stopping condition,such as, if there is no improvement of validation loss for fiveconsecutive epochs, then the training is stopped. We observedthrough multiple runs of training, that each of the modelsconverged within a given range of the maximum number ofepochs, as presented in Table IV.

2 4 6 894

96

98

100

Number of Transmitters

Acc

urac

y(%

)

TrainTest

Fig. 25. Training and Accuracy with Increasing numbers of Transmitters



15

CNN

89.07

%

DNN

96.49

%

RNNLS

TM97.40

%

RNNGR

U97.85

%

(a) 4 Transmitters

CNN

81.59

%

DNN

94.6%

RNNLS

TM95.78

%

RNNGR

U97.06

%

(b) 8 Transmitters

Fig. 26. Detection Accuracies of the Neural Networks

It is clear from the the results presented above that theGAN based NN is effective for the task of rogue transmitteridentification whereas DNN and RNN are effective for thetask of trusted transmitter classification. In summary we haveestablished the following:

1) A GAN is able to distinguish between Trusted andCounterfeited RF transmitters.

2) CNN yields 81%-86% accuracy for trusted transmitterclassification proving the inefficacy of spatial correlationas a discriminative attribute for transmitter classification.

3) DNN yields 94-97% accuracy for known transmitterclassification.

4) RNN yields an accuracy of 97% for known transmitterclassification using GRU cells.

5) Comparing the accuracies, we can conclude that I/Q dataof radio signals exhibits more temporal correlation thanspatial correlation.

6) DNN or RNN can be used for transmitter fingerprintingfor identifying trusted transmitters and in conjunctionwith GAN this can be used to create an end-to-endrobust system for transmitter identification.

G. Computational Complexities

In this section we present the computational time com-plexity for the training phase only, as the trained modelgives the output within constant time (O(1)) during thedeployment phase. Understanding the time complexity oftraining a NN is still an evolving research area. In [62],the authors proved that a NN of depth δ can be learned inpoly(s2

δ

) time, where s is the dimension of the input, andpoly(.) takes a constant time depending on the configurationof the system. However, the convolution operations of CNNadd additional time complexity along with the forward andback-propagation operations. In [56], the authors mentionedthat the time complexity for training all the convolutionallayers is: O(

∑ζτ=1(ητ−1ν

2τ .ητρ

2τ ), where ζ is the number of

convolutional layers, τ is the index of a convolutional layer,ητ−1 is the number of input channels of the τ th layer, ντ isthe spatial size of the filters at the τ th layer, ητ is the numberof filters at the τ th layer and ρτ is the size of the output

#Trans Models Complexity

4 GAN (8) poly(0.95× 160e328)

4 CNN (7) poly(0.95× 160e323)

×O(∑3

τ=1(ητ−1.ν2τ .ητ .ρ2τ )

+ poly(0.95× 160e324)

4 DNN (5) poly(0.95× 160e325)

4 RNN - LSTM (6) poly(0.95× 160e326)

4 RNN - GRU (6) poly(0.95× 160e326)

8 GAN (8) poly(0.95× 320e328)

8 CNN (7) poly(0.95× 320e323)

×O(∑3

τ=1(ητ−1.ν2τ .ητ .ρ2τ )

+ poly(0.95× 320e324)

8 DNN (5) poly(0.95× 320e325)

8 RNN - LSTM (6) poly(0.95× 320e326)

8 RNN - GRU (6) poly(0.95× 320e326)

TABLE VTIME COMPLEXITIES FOR TRAINING OF THE VARIOUS IMPLEMENTATIONS

features of the τ th layer. In the proposed CNN model, wehave 3 convolutional layers, and 4 fully connected layers andhence we add in additional time complexity for training thethree convolutional layers.

The time complexities for each implemented NN modelis presented in Table V, using the aforementioned resultson time complexity of neural network training. The numberswithin the parenthesis in the second column represents thetotal number of layers for a particular model. Note that wehave two different datasets of dimensions 160K and 320Kand as mentioned earlier, we use 95% of data for trainingand validation purpose. For example, the complexity for GANwith 8 layers using 95% of 160e3 data samples for trainingand validation, is poly(0.95 × 160e32

8

). On another note weshould mention that for quick estimates of the time complexityof training neural networks, the number of hyper-parameters(as presented in Table III) can also be used as a measure of therequired training time. Thus, the more the number of hyper-parameters, the more the training time required.

H. Experiments with Heterogeneous Dataset

So far, we have used the proposed trusted transmitteridentification models on “homogeneous” datasets in that thetransmitters were implemented using the SDRs from the samemanufacturer. However, in reality the trusted transmitters canbe from several different manufacturers. Now we want toexplore how the accuracy of the trusted transmitter identifi-cation system would change if “heterogeneous” data obtainedfrom different types of transmitters (manufacturers) (as wasdiscussed in Section V-C) was used. From the testing accuracyas shown in Table VI we observe that all the NNs performbetter when the transmitters are from different manufacturersand hence are fundamentally of different types. This confirmsthe intuition that radios manufactured using different processes(from different manufacturers) contain easily exploitable char-acteristics in their I/Q samples, that can be implicitly learnedusing a NN. We also observe that the CNN can exploitthe spatial correlation better for the heterogeneous dataset,



16

Models USRP-USRP PLUTO-USRPCNN 89.91 (%) 99.91 (%)DNN 99.9 (%) 100 (%)RNN 99.95 (%) 100 (%)

TABLE VICOMPARISON OF TESTING ACCURACIES FOR DIFFERENT CLASSIFICATION

MODELS FOR HOMOGENEOUS AND HETEROGENEOUS DATASETS

yielding a 11.2% increase in the testing accuracy comparedto the homogeneous dataset.

I. Existing Transmitter Classification Techniques Compar-isons

Though RFAL demonstrates the use of adversarial learningfor rogue transmitter identification, a part of our work has beendevoted to the problem of “trusted transmitter classification”after elimination of the rogue transmitters. “Trusted transmit-ter” identification systems when augmented with adversariallearning systems, as was done in RFAL, results in robusttransmitter identification systems that are immune to presenceof adversarial transmissions. Though our focus was not onimproving or building novel “trusted transmitter” classificationsystems, we have explored the use of deep learning forbuilding the same using I/Q data from the received signal. Inthis section we present a comparative study of our approachto “trusted transmitter classification with I/Q data” againstsome existing techniques for “transmitter classification” andthe results of this endeavor is shown in Tables VII for tradi-tional approaches (low accuracy) and VIII for “state-of-the-art” (high accuracy) respectively. Note that here the “Inputs”column refers to the type of inputs used for the classificationalgorithms. It is to be noted that all the traditional methodsuse some form of extracted features (obtained through pre-processing of the data) as inputs ( [23]–[26]), or work withsynthetic dataset ( [31]). A few existing work on modulationrecognition [19], [30], [33] using NN based approaches alsowork on synthetic datasets [32], [63] and hence they donot yield to a fair comparison with our “trusted transmitterclassification” approach as well.

For the “state-of-the-art” approaches, we observe that thetest bed experiments have used various types of SDRs forobtaining over-the-air data, or used datasets from actual infras-tructure transmitters (ACARS etc.). We present comparisonwith: (i) [36], where same SDR (USRP B210) was used,(ii) [38] and [40], where different SDRs from the samemanufacturer (USRP N210, USRP X310) were used, (iii) [35],where different types of devices (Zigbee) were used, and (iv)[37], [39] and [41] where different datasets (ACARS [42],ADS-B [43], FIT/CorteXlab [44]) were used.

The RFAL “trusted transmitter identification” models out-perform [35] where the authors achieved 91.38% accuracyfor classifying 7 Zigbee devices. The CNN proposed in [36]achieved 98% accuracy for 5 USRP B210 with preprocesseddata (from MATLAB Communication Systems Toolbox). Sim-ilarly, in [40], Sankhe et al. presented a CNN classifier with16 X310 radios with 99.5% accuracy, but that method useddemodulated symbols rather than raw signal data. Youssef etal. presented a multi-stage training model to achieve 100%

Approach #Trans SNR (dB) Acc (%) InputsGenetic Algorithm [23] 5 25 85-98 Transients

Multifractal NotSegmentation [24] 8 mentioned 92.5 Transients

Orthogonal Component SpuriousReconstruction (OCR) [26] 3 20 62 - 71 Modulation

k-NN [25] 8 30 97 TransientsRNN [31] - 20 90 Synthetic Dataset [32]

RFAL (Ours) 8 30 97.04 Raw SignalTABLE VII

COMPARISON OF THE RFAL IMPLEMENTATION WITH THE TRADITIONALONES

accuracy to classify 12 USRP N210s. Multi-stage trainingis a complex procedure and needs a easily parallelizabletraining environment, whereas CNN and DNN use a first-order update rule (stochastic gradient) and are comparativelysimple procedures. For the sake of generality, we compare theirproposed DNN method (which has the best accuracy comparedto other implemented ML methods) in Table VIII. The CNNmodels presented in [37], [39] and [41] achieved between 96%and 99.9% accuracy for existing datasets (ACARS [42], ADS-B [43], FIT/CorteXlab [44]).

We need to point out that none of methods discussed aboveare robust enough to work in adversarial settings. Thus ifthere is an adversarial transmitter then all the aforementionedmethods will fail. However RFAL, due to its adversarialtraining will be able to identify the adversarial transmitter andeliminate it from consideration before classifying the “trustedtransmitters”, thus being more resilient and robust to suchinterference. Even if we just compare the “trusted transmitter”classification of RFAL with the methods discussed above,our method for “trusted transmitter classification” achievesthe same accuracy (of 97%) using just the raw I/Q data asinput, thereby paving the way for real-time deployment of“transmitter fingerprinting” systems.

Considering the state-of-art, and to the best of our knowl-edge, our work is the first to:

1) propose a GAN based model to detect rogue transmittersfrom authentic ones;

2) demonstrate a high accuracy (97%) to classify 8 USRPB210s using an RNN model considering the temporalproperty of RF data;

3) present a testbed evaluation for classifying transmittersfrom different manufacturers;

4) provide an end-to-end solution of RF transmitter classi-fication without any preprocessing of raw data. The rawdata can be captured through any SDR and is identifiedby the proposed models at once;

5) propose RNN based models which needed half the train-ing time than CNN models for the same experimentaltraining time. Our proposed RNN model entails to befaster than all state-of-the art CNN models for the sameexperimental environment.

J. Performance Comparison for Varying SNR

In this section, we present the results of RFAL “trustedtransmitter classification” for varying SNR values. We com-pare the accuracy for the proposed NN models having 8 USRPB210s with 30 dB SNR, with 3 other datasets collected at



17

Approach #Trans SNR (dB) Acc (%) InputCNN [35] 7 30 91.38 Preprocessed data

from MATLABCNN [36] 5 50 98 Preprocessed data

from MATLABCNN [37] - - 99.67 ACARS data [42]DNN [38] 12 - 84.4 Raw signalInception - - 98.1 & 96.3 ACARS [42] &

ResNet [39] ADS-B [43]CNN [40] 16 30 99.5 Demodulated symbolsCNN [41] 21 - 99.99 FIT/CorteXlab [44]

RNN (Ours) 8 30 97.04 Raw signalTABLE VIII

COMPARISON OF THE RFAL IMPLEMENTATION WITH STATE-OF-THE-ART

SNR(dB) Accuracy (%)CNN DNN RNN (GRU)

0 51.53 85.12 92.310 78.64 92.24 95.6420 81.3 94.60 97.0230 81.59 94.60 97.06

TABLE IXACCURACIES FOR DIFFERENT NEURAL NETWORK MODELS WITH

VARYING SNRS

0 dB, 10 dB, and 30 dB SNRs having the same number oftransmitters (8 B210s) as shown in Table IX. It is seen that weachieve better accuracy with all the models for higher SNRvalues, which is intuitive. It is to be noted that the proposedRNN (with GRU cell) model gives more than 92% accuracy at0 dB SNR too, whereas CNN and DNN models fail to achievethat.

It must be pointed out that the proposed NN models canbe a trained using raw signal data from any type of radiotransmitter. We would also like to point out that though ourdata was collected in a lab setting, we had no control over theRF environment: there were other transmissions, uncontrolledmovement of people, and multi-path effects due to the locationand layout of the lab. Moreover, the power of the transmitterswas low which compounded the problem further. Thus, thoughwe mention that the data was collected in a lab environment, inreality it was an uncontrolled RF environment reflective of oursurroundings. We can safely argue that the proposed methodswill perform equally well in any real world deployment oflarge scale radio networks.

VII. CONCLUSIONS

In this paper, we address the problem of building a robustand resilient model for identifying similar RF transmitters inthe presence of adversaries. We argue that non-adversarialmachine learning techniques would not be effective in adver-sarial settings and that breakthroughs in generative adversarialnets (GANs) can be instrumental in building such systemsfor detection of rogue transmitters and subsequent accurateidentification of known ones in such settings. We proposeand implement RF Adversarial Learning (RFAL) frameworkwhich includes a discriminative model for identifying roguetransmitters trained with data generated from a generativemodel. RFAL also contains a “trusted transmitter” identifica-tion system that for categorizing the known transmitters oncethe adversarial transmitters have been identified and eliminatedfrom consideration. We collected over-the-air raw I/Q data

using USRP B210s and used that to train the GAN. Thediscriminator was able to detect rogue transmitters with anaccuracy of ∼99.9%. As for the subsequent trusted transmitterclassification, we first implemented a CNN (accuracy ∼89%)for exploiting the spatial correlation between the I/Q data.Then we designed and implemented a fully connected DNNand RNNs both of which obtained an accuracy of around 97%for trusted transmitter identification. We also show how theproposed NN models for “trusted transmitter classification”(especially CNN) worked when radios from different manu-facturers were used. RFAL will be able to detect any activeattack that involves a secondary device to pose as an authenticemitter, such as replay attacks, but it will not be able to detectpassive attackers, such as traffic sniffers. Going forward wewould like to use these methods for identification of actualinfrastructure transmitters (for example FM, AM or GSM) incontested real world settings.

ACKNOWLEDGMENT

This work was partly supported by the Air Force ResearchLaboratory. The views and conclusions contained herein arethose of the authors and should not be interpreted as neces-sarily representing the official policies or endorsements, eitherexpressed or implied, of the United States Air Force.

REFERENCES

[1] R. H. Weber and R. Weber, Internet of Things. Springer, 2010, vol. 12.[2] M. Valkama, M. Renfors, and V. Koivunen, “Advanced Methods for I/Q

Imbalance Compensation in Communication Receivers,” IEEE Transac-tions on Signal Processing, vol. 49, no. 10, pp. 2335–2344, 2001.

[3] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,no. 7553, pp. 436–444, 2015.

[4] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press,2016.

[5] H. W. Lin, M. Tegmark, and D. Rolnick, “Why does deep and cheaplearning work so well?” Journal of Statistical Physics, vol. 168, no. 6,pp. 1223–1247, 2017.

[6] Ettus Research, “USRP B210,” https://www.ettus.com/product/details/UB210-KIT, 2018.

[7] NooElec, “USRP B210,” http://www.nooelec.com/store/sdr/sdr-receivers/nesdr-mini-rtl2832-r820t.html, 2018.

[8] A. Devices, “ADALM-PLUTO Overview,”https://wiki.analog.com/university/tools/pluto, 2018.

[9] M. D. L. Angrisani and M. Vadursi, “Clustering-based method fordetecting and evaluating I/Q impairments in radio-frequency digitaltransmitters,” IEEE Transactions on Instrumentation and Measurement,vol. 56, no. 6, pp. 2139–2146, 2007.

[10] C.-L. Liu, “Impacts of I/Q Imbalance on QPSK-OFDM-QAM Detec-tion,” IEEE Transactions on Consumer Electronics, vol. 44, no. 3, pp.984–989, 1998.

[11] C. Hsu and W. Sheen, “Joint Calibration of Transmitter and ReceiverImpairments in Direct-Conversion Radio Architecture,” IEEE Transac-tions on Wireless Communications, vol. 11, no. 2, pp. 832–841, 2012.

[12] F. E. Churchill, G. W. Ogar, and B. J. Thompson, “The Correction of Iand Q Errors in a Coherent Processor,” IEEE Transactions on Aerospaceand Electronic Systems, vol. AES-17, no. 1, pp. 131–137, 1981.

[13] S. A. Bassam, S. Boumaiza, and F. M. Ghannouchi, “Block-WiseEstimation of and Compensation for I/Q Imbalance in Direct-ConversionTransmitters,” IEEE Transactions on Signal Processing, vol. 57, no. 12,pp. 4970–4973, 2009.

[14] S. Burglechner, G. Hueber, and A. Springer, “On the Estimation andCompensation of IQ Impairments in Direct Conversion Transmitters,”in European Conference on Wireless Technology, 2008, pp. 69–72.

[15] L. Anttila, M. Valkama, and M. Renfors, “Blind Moment EstimationTechniques for I/Q Imbalance Compensation in Quadrature Receivers,”in IEEE 17th International Symposium on Personal, Indoor and MobileRadio Communications, 2006, pp. 1–5.



18

[16] J. Tubbax et al., “Compensation of IQ imbalance in OFDM systems,”in IEEE International Conference on Communications, 2003, pp. 3403–3407.

[17] I. Goodfellow et al., “Generative Adversarial Nets,” in Advances inNeural Information Processing Systems, 2014, pp. 2672–2680.

[18] C. M. Bishop, Pattern Recognition and Machine Learning (InformationScience and Statistics). Springer-Verlag, 2006.

[19] T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional RadioModulation Recognition Networks,” in Engineering Applications ofNeural Networks, 2016, pp. 213–226.

[20] T. J. O’Shea, J. Corgan and T. C. Clancy, “Unsupervised RepresentationLearning of Structured Radio Communication Signals,” in First Inter-national Workshop on Sensing, Processing and Learning for IntelligentMachines (SPLINE), 2016, pp. 1–5.

[21] S. Hochreiter and J. Schmidhuber, “Long Short-term Memory,” Neuralcomputation, vol. 9, no. 8, p. 17351780, November 1997.

[22] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical Evaluation ofGated Recurrent Neural Networks on Sequence Modeling,” ComputingResearch Repository (CoRR), vol. abs/1412.3555, 2014.

[23] J. Toonstra and W. Kinsner, “A radio transmitter fingerprinting systemODO-1,” in Canadian Conference on Electrical and Computer Engi-neering, vol. 1, 1996, pp. 60–63.

[24] D. Shaw and W. Kinsner, “Multifractal Modelling of Radio TransmitterTransients for Classification,” in IEEE WESCANEX, 1997, pp. 306–312.

[25] I. O. Kennedy, P. Scanlon, F. J. Mullany, M. M. Buddhikot, K. E.Nolan, and T. W. Rondeau, “Radio Transmitter Fingerprinting: A SteadyState Frequency Domain Approach,” in IEEE Vehicular TechnologyConference, 2008, pp. 1–5.

[26] S. Xu, L. Xu, Z. Xu, and B. Huang, “Individual Radio TransmitterIdentification based on Spurious Modulation Characteristics of SignalEnvelop,” in IEEE MILCOM, 2008, pp. 1–5.

[27] M. C. Du Plessis and J. C. Olivier, “Radar Transmitter Classificationusing a Non-stationary Signal Classifier,” in International Conferenceon Wavelet Analysis and Pattern Recognition, 2009, pp. 482–485.

[28] C. Zhao, L. Xie, X. Jiang, L. Huang, and Y. Yao, “A PHY-layerAuthentication Approach for Transmitter Identification in CognitiveRadio Networks,” in International Conference on Communications andMobile Computing, vol. 2, 2010, pp. 154–158.

[29] Y. Gong, G. Hu, and Z. Pan, “Structured Sparsity Preserving Projectionsfor Radio Transmitter Recognition,” in International Conference onMobile IT Convergence, 2011, pp. 68–73.

[30] T. O’Shea and J. Hoydis, “An Introduction to Deep Learning for thePhysical Layer,” IEEE Transactions on Cognitive Communications andNetworking, vol. 3, no. 4, pp. 563–575, 2017.

[31] S. Rajendran et al., “Deep Learning Models for Wireless SignalClassification With Distributed Low-Cost Spectrum Sensors,” IEEETransactions on Cognitive Communications and Networking, vol. 4,no. 3, pp. 433–445, 2018.

[32] radioML, “RFML 2016,” https://github.com/radioML/dataset, 2018.[33] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-Air Deep Learning

Based Radio Signal Classification,” IEEE Journal of Selected Topics inSignal Processing, vol. 12, no. 1, pp. 168–179, 2018.

[34] T. J. O’Shea et al., “Physical Layer Communications System DesignOver-the-Air Using Adversarial Networks,” Computing Research Repos-itory (CoRR), vol. abs/1803.03145, 2018.

[35] K. Merchant, S. Revay, G. Stantchev, and B. Nousain, “Deep learningfor rf device fingerprinting in cognitive communication networks,” IEEEJournal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 160–167, 2018.

[36] S. Riyaz, K. Sankhe, S. Ioannidis, and K. Chowdhury, “Deep LearningConvolutional Neural Networks for Radio Identification,” IEEE Com-munications Magazine, vol. 56, no. 9, pp. 146–152, 2018.

[37] S. Zheng, S. Chen, L. Yang, J. Zhu, Z. Luo, J. Hu, and X. Yang, “BigData Processing Architecture for Radio Signals Empowered by DeepLearning: Concept, Experiment, Applications and Challenges,” IEEEAccess, vol. 6, pp. 55 907–55 922, 2018.

[38] K. Youssef, L. Bouchard, K. Haigh, J. Silovsky, B. Thapa, and C. V.Valk, “Machine Learning Approach to RF Transmitter Identification,”IEEE Journal of Radio Frequency Identification, vol. 2, no. 4, pp. 197–205, 2018.

[39] S. Chen, S. Zheng, L. Yang, and X. Yang, “Deep learning for large-scalereal-world ACARS and ADS-B radio signal classification,” ComputingResearch Repository (CoRR), vol. abs/1904.09425, 2019.

[40] K. Sankhe, M. Belgiovine, F. Zhou, S. Riyaz, S. Ioannidis, andK. Chowdhury, “ORACLE: Optimized Radio clAssification throughConvolutional neuraL nEtworks,” in IEEE Conference on ComputerCommunications (INFOCOM), 2019, pp. 370–378.

[41] C. Morin, L. Cardoso, J. Hoydis, J.-M. Gorce, and T. Vial, “TransmitterClassification With Supervised Deep Learning,” Computing ResearchRepository (CoRR), vol. arXiv:1905.07923, 2019.

[42] ARINC618-5, “air/ground character-oriented protocol, specification,”Aeronautical Radio, Inc., August 2000.

[43] ——, “Automatic Dependent Surveillance-Broadcast (ADS-B),” https://www.faa.gov/nextgen/programs/adsb/, August 2018.

[44] A. Massouri, L. Cardoso, B. Guillon, F. Hutu, G. Villemaud, T. Risset,and J. Gorce, “CorteXlab: An open FPGA-based facility for testingSDR cognitive radio networks in a reproducible environment,” in IEEEConference on Computer Communications Workshops (INFOCOM WK-SHPS), 2014, pp. 103–104.

[45] Y. Huang and H. Zheng, “Radio Frequency Fingerprinting based on theConstellation Errors,” in IEEE Asia-Pacific Conference on Communica-tions, 2012, pp. 900–905.

[46] T. Mukherjee et al., “RSSI-Based Supervised Learning for Uncoopera-tive Direction-Finding,” in Proceedings of ECML-PKDD. Springer, 102017.

[47] M. Pospil, R. Marsalek, and J. Pomenkova, “Wireless Device Authenti-cation through Transmitter Imperfections Measurement and Classifica-tion,” in IEEE International Symposium on Personal, Indoor and MobileRadio Communications, 2013, pp. 497–501.

[48] P. Hao, X. Wang, and A. Behnad, “Performance Enhancement of I/QImbalance based Wireless Device Authentication through Collaborationof Multiple Receivers,” in IEEE International Conference on Commu-nications, 2014, pp. 939–944.

[49] E. L. Denton et al., “Deep Generative Image Models using a LaplacianPyramid of Adversarial Networks,” in Advances in Neural InformationProcessing Systems, 2015, pp. 1486–1494.

[50] J. Zhu et al., “Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks,” Computing Research Repository(CoRR), vol. abs/1703.10593, 2017.

[51] D. Berthelot, T. Schumm, and L. Metz, “BEGAN:Boundary EquilibriumGenerative Adversarial Networks,” Computing Research Repository(CoRR), vol. abs/1703.10717, 2017.

[52] N. Srivastava et al., “Dropout: A Simple Way to Prevent NeuralNetworks from Overfitting,” The Journal of Machine Learning Research,vol. 15, no. 1, pp. 1929–1958, 2014.

[53] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,”Computing Research Repository (CoRR), vol. abs/1412.6980, 2014.

[54] F. Chollet et al., “Keras: The Python Deep Learning library,”https://keras.io, 2015.

[55] V. Nair and G. E. Hinton, “Rectified Linear Units Improve RestrictedBoltzmann Machines,” in Proceedings of International Conference onInternational Conference on Machine Learning, 2010, pp. 807–814.

[56] K. He and J. Sun, “Convolutional Neural Networks at Constrained TimeCost,” in IEEE Conference on Computer Vision and Pattern Recognition,June 2015.

[57] R. Dey and F. M. Salemt, “Gate-variants of Gated Recurrent Unit (GRU)neural networks,” in 2017 IEEE 60th International Midwest Symposiumon Circuits and Systems (MWSCAS), 2017, pp. 1597–1600.

[58] GNURadio, “GNU Radio,” https://www.gnuradio.org, 2018.[59] Pavel Sorejs, “Spektrum,” https://github.com/pavels/spektrum, 2015.[60] M. Abadi et al., “Tensorflow: Large-scale machine learning on het-

erogeneous distributed systems,” CoRRComputing Research Repository(CoRR), 2016.

[61] R. Al-Rfou et al., “Theano: A python framework for fast computationof mathematical expressions,” Computing Research Repository (CoRR),2016.

[62] R. Livni, S. Shalev-Shwartz, and O. Shamir, “On the ComputationalEfficiency of Training Neural Networks,” in Advances in Neural Infor-mation Processing Systems, 2014, pp. 855–863.

[63] T. O’Shea and N. West, “Radio Machine Learning Dataset Generationwith GNU Radio,” Proceedings of the GNU Radio Conference, vol. 1,no. 1, 2016.

https://www.faa.gov/nextgen/programs/adsb/

https://www.faa.gov/nextgen/programs/adsb/



19

Debashri Roy received her M.S. degree in Com-puter Science at University of Central Florida, whereshe is currently a PhD Candidate. Her researchinterests are radio frequency machine learning Sys-tems (RFMLs), generative adversarial nets, real-timevideo transmission, vehicular networks, heteroge-neous networks, modeling and simulation, dynamicspectrum access.

Dr. Tathagata Mukherjee is an Assistant Professorof Computer Science in the University of Alabama inHuntsville. He obtained his M.S and Ph.D in Com-puter Science from The Florida State University. Hisinterests are in Cyber Security, Adversarial MachineLearning, Cognitive Radio Networks, Optimizationand Graph Theory. Prior to his current position hewas the Chief Scientist at Intelligent Robotics Inc.,a non-profit DoD research lab.

Dr. Mainak Chatterjee is an Associate Professorin the department of Computer Science at the Uni-versity of Central Florida, Orlando. He received theBSc degree in physics (Hons.) from the University ofCalcutta, the ME degree in electrical communicationengineering from the Indian Institute of Science,Bangalore, and the Ph.D. degree in computer sciencefrom the University of Texas at Arlington. His re-search interests include economic issues in wirelessnetworks, applied game theory, cognitive radio net-works, dynamic spectrum access, and mobile video

delivery.

Dr. Erik Blasch is an IEEE Fellow and principalscientist at the US Air Force Research Lab (AFRL)in the Information Directorate at Rome, NY. he isalso a program manager for the AFOSR DDDASprogram. He received his B.S. in Mechanical Engi-neering from the MIT in 1992, Masters Degrees inMechanical (94), Health Science (95), and IndustrialEngineering (Human Factors) (95) from GeorgiaTech. He served in Active Duty from 1996 in theUnited States Air Force. He completed an MBA(98), MSEE (98), MS Econ (99), and a PhD (99)

in Electrical Engineering from Wright State University and is a graduate ofAir War College (08).

Dr. Eduardo Pasiliao received B.S. degree inMechanical Engineering from Columbia University,New York, NY, USA, in 1992, and M.S. degree inCoastal and Oceanographic Engineering and Ph.D.degree in Industrial and Systems Engineering fromthe University of Florida, Gainesville, FL, USA,in 1995 and 2003, respectively. He is currentlya Senior Research Engineer with the Air ForceResearch Laboratory. He has co-authored over 100peer-reviewed journal and conference publications.His research interests are in the areas of Graph

Theory, Combinatorial Optimization, Discrete Mathematics, Network Science,and Machine Learning. He has been a member of the American Institute forAeronautics and Astronautics and the Institute for Operations Research andManagement Sciences.

RFAL: Adversarial Learning for RF Transmitter ...

Documents

Transcript of RFAL: Adversarial Learning for RF Transmitter ...