View source for Projects:2019s1-115 Deep Learning Based Specific Emitter Identification


== Team ==
===Students===
<ul>
<li>Vaibhav Sekhar</li>
<li>William Voss</li>
</ul>

===Supervisors===
<ul>
<li>Dr. Brian Ng</li>
<li>Dr. Joy Li (DST Group)</li>
<li>Dr. Tharaka Lamahewa (DST Group)</li>
</ul>

== Introduction ==
Specific Emitter Identification (SEI) is the association of a received signal with its transmitter. Such capability is highly desirable for many military and civilian applications.  Cognitive radio applications heavily rely on SEI aided emitter tracking to reinforce the Dynamic Spectrum Access (DSA) rules to improve networks reliability. In wireless communication, many networks often use SEI capability to support emitter authentication to enhance network security. 
<br><br>
Traditionally, the SEI process is achieved through precise measurement of the intercepted signals and carefully extracts some of the signal features which are caused by the unintentional impairments of the emitter, known as RF finger printing. Based on the signal feature extraction, the intercepted signals are then clustered by emitter before identification process is carried out. This identification process is usually database aided
<br><br>
IQ imbalance exists in most of the direct-up converters, and is one of the most commonly seen emitter impairments. It refers to the mismatch between the in-phase and quadrature paths of the transmitter and is introduced by the imperfection of analogue components within the transmitter. This mismatch can be amplitude or phase mismatch. But most commonly, it is found in a combination of both. It is also most likely non-linear and frequency dependent, which reflect the nature of analogue components. IQ imbalance is very often used for transmitter characterization as part of the SEI process.
<br><br>
The SEI process is challenging both from the hardware requirement perspective and from the processing perspective. As the SEI process is only made possible by identifying the naturally occurring and unintentional characteristics of the emitter, the receiver that is used for signal intercept must be highly linear to avoid ambiguities. Any hardware imperfection introduced by the intercept receiver such as frequency dependent amplitude or phase offsets must be calibrated out before the intercept process taken place. This calibration process can be very time consuming, and may require to be repeated just before the actual signal intercept and collection process begins. 
<br><br>
The computational cost associated with the SEI process is significant. Depending on the method used for the actual SEI process, signals are often transformed into one or multiple joint domains, which can be very computationally expensive. Dimensionality reduction is also used in some of the SEI techniques, which requires large matrix manipulation. This further increases the complexity and number of operations required for the SEI process. 
<br><br>
In recent years, a new machine learning methodology known as Deep Learning (DL) has found its own success in areas such as computer vision, speech recognition and image classification areas [1,2].  As the Deep Neural Network (DNN) has proven itself to be very effective at processing unstructured data, many innovative technology companies have started to experiment with DNN to replace signal processing tasks for many communication applications [6].  In particular, authors in [3,4,5] have successfully applied DL algorithms as a computational efficient way to estimate IQ imbalance and characterize transmitters for SEI purposes.

== Project Aims and Objectives ==
This research aims to investigate the use of CNNs for IQ imbalance estimation. The key objectives of this research are:
<ul>
<li> To determine the feasibility of CNNs for IQ imbalance estimation</li>
<li> To compare the IQ imbalance estimation performance of CNNs to existing Digital Signal Processing techniques </li>
</ul>

== Background ==
===Overview of digital signal transmission===
Complex-valued signals can be represented digitally as streams of binary data. Transmitters first process the digital data through a digital signal processor (DSP). The DSP divides the binary stream into smaller segments known as symbols, with the length of the symbols (in bits) depending on the desired modulation alphabet. The modulation alphabet maps the symbols to complex numbers. The complex numbers are then separated into two branches of real-valued data, as shown in Figure 2.2. These are known as the In-Phase (I) component, representing the real part of the complex numbers, and the Quadrature (Q) component, for the imaginary part of the complex numbers. The I and Q components are then each multiplied by Local Oscillator (LO) signals such that they are 90 degrees out-of-phase with each other. The resulting LO multiplied signals are then summed and processed for transmission.   

<gallery class="center" widths="300px" heights="300px" >
[[File:Screen Shot 2019-11-08 at 1.09.12 pm.png|frame|Conversion from binary to complex numbers using QAM16 modulation alphabet]]
</gallery>


===Direct conversion transmitters: issues with I/Q imbalance===
Transceivers used in communications systems transmit/receive signals at RF frequencies. Direct conversion transmitters are commonly used in wireless communications. Unlike traditional (superheterodyne) transmitters, direct-conversion transmitters do not step-up to an Intermediate Frequency when converting from (input) baseband frequency to Radio Frequency (RF). Instead, they upconvert directly from the baseband frequency to RF, hence the name. This reduces the number and size of components required to transmit at RF, which enables direct conversion transmitters to have smaller packaging and a simpler design than superheterodyne transmitters. 

However, due to imperfections in the analogue components of the direct conversion transmitter, there can be a mismatch in the paths of the I and Q components. The distortion of I and Q data is referred to as I/Q imbalance. I/Q imbalance can cause undesirable gain and phase offsets on the signal.  Since I/Q imbalance is unique to the type of components used in a transmitter, it can be used as a feature by SEI systems to characterise an emitter.


===Artifical Neural Networks===
Artificial Neural Networks (ANNs) are a machine learning framework loosely modelled on the structure of biological neural networks of animal brains. ANNs consist of layers of neurons, linked to each other through a series of weighted connections. Neurons are mathematical computation units mapping inputs to outputs. ANNs consist of at least two layers, an input layer and output layer, although they often also consist of multiple hidden layers. Each neuron processes its input and connection weight through an activation function, with the result passed onto the next layer. During training, neurons also add an adjustment term, known as bias, that improves the accuracy of predictions.

Common types of ANNs include Recurrent Neural Networks (RNNs), Radial Basis Function Networks (RBFNs), and Multilayer Perceptron (MLP) networks. Each of these forms of ANNs differs in the structure of their neurons. This project involved the use of Convolutional Neural Networks (CNNs), which is a special type of MLP network. 

====Convolutional Neural Networks====
Unlike other forms of neural networks, all CNNs contain at least one hidden convolutional layer. Convolutional layers consist of filters used to recognise features or patterns in the input data. These layers perform a convolution operation by sliding the filter over the input (hence the name). Filters have two properties:
<ul>
<li> 	Size (M×N): the dimensions of the filter – M rows, N columns </li>
<li>	Stride (M, N): the length of each filter step, i.e. how far the filter slides between steps – move M points across and N points down for each step </li>
</ul>

CNNs learn features of the input data and correct for errors in estimation using a process known as backpropagation. The backpropagation algorithm is as follows:
<ul>
<li>1.	The network uses a cost function to compare the computed output with the expected output. </li>
<li>2.	The network then minimises the error by computing the partial derivatives of the cost function with respect to the network weights and readjusting the weights accordingly. The cost function in backpropagation is usually the mean squared error </li>
</ul>

== Method ==
This was split into three parts:<br>
1. Determine if CNNs can estimate IQ imbalances<br>
2. Compare the performance against traditional techniques, by testing for robustness to background noise and accuracy<br>
3. Determine the versatility of CNNs to different inputs

== Results ==

Part 1:<br>
The results show that CNNs can estimate IQ imbalance accurately. Moreover, they are more robust to background noise compared to existing techniques. We can see from the graphs below that the Mean Squared Error (plotted on the vertical axis), between the estimated IQ imbalance and the true value, is lower for the CNN than it is for the traditional technique regardless of how noisy the signal is. 

<gallery class="center" widths="450px" heights="450px" >
Screen Shot 2019-11-08 at 2.55.46 pm.png|Figure A: Gain imbalance estimation error vs SNR: comparing CNN (blue) against existing techniques (yellow) for three different signal lengths of QAM16
Screen Shot 2019-11-08 at 2.57.55 pm.png|Figure B: Phase imbalance estimation error vs SNR: comparing CNN (blue) against existing techniques (yellow) for three different signal lengths of QAM16
</gallery>

Part 2:<br>
Previous research mainly focussed on Mean Squared Error, however, we were also able to look at the average bias. The best estimators should have low MSE and bias. We can see that the CNN estimates of IQ imbalance (gain and phase imbalances) are less biased than the traditional estimator, regardless of SNR.

<gallery class="center" widths="450px" heights="450px" >
Screen Shot 2019-11-08 at 3.02.18 pm.png|Figure C: Average bias of gain imbalance estimates over SNR
Screen Shot 2019-11-08 at 3.02.33 pm.png|Figure D: Average bias of phase imbalance estimates over SNR
</gallery>


Part 3:<br>
The results in Figures E and F indicate that CNNs are versatile/adaptable to different formats of inputs. Despite magnitude/phase being different quantities to I/Q values, CNNs were able to estimate IQ imbalance as they did with IQ data. Generally, MSE decreased when signal length increase, and when modulation order decreased. The accuracy of gain offset estimates deteriorated when magnitude and phase values were used as input compared to raw I/Q data. Regardless of modulation type, the MSE of gain offsets with magnitude and phase were approximately twice that of the MSE with raw I/Q data. This indicates that the CNN found it difficult to establish a relationship between the magnitude/phase and gain offset, than with I/Q data. The MSE of phase offset estimates using magnitude/phase were also higher than with I/Q data. This could be due to the difficulty in distinguishing between π and -π radians.

<gallery class="center" widths="450px" heights="450px" >
Screen Shot 2019-11-08 at 3.52.19 pm.png|Figure E: Comparing Gain imbalance estimation error when the CNN is trained on Magnitude/Phase versus I/Q data
Screen Shot 2019-11-08 at 3.52.27 pm.png|Figure F: Comparing Phase imbalance estimation error when the CNN is trained on Magnitude/Phase versus I/Q data
</gallery>

== Conclusions and Future Work ==
===Conclusion===
The primary aim of this thesis was to understand if CNNs could be used for IQ imbalance estimation. The results from showed that CNNs were capable of estimating IQ imbalance. It was also shown that the accuracy of estimates was proportional to the signal length, with longer signals estimated with greater accuracy than shorter signals. The estimation accuracy also depended on the modulation scheme and order used. PSK signals were estimated with lower error compared to QAM signals, as the constellations of PSK signals were more robust to IQ imbalance than those of QAM signals. Signals with lower modulation order were estimated with greater accuracy than the higher modulation orders due to their simplicity.  

Further experimentation demonstrated CNNs could outperform existing receiver-side and transmitter-side IQ imbalance estimation techniques and were more robust to noise. CNNs were also significantly more robust to frequency-offset compared to traditional techniques, which were unable to handle frequency offset. Unlike CNNs, which were trained on signals with frequency offset, frequency offset detection was not built-in to the transmitter-side estimation algorithm.

Moreover, CNNs proved to be more versatile at adapting to different formats of signal input compared to existing techniques, which can only handle I/Q values. 

===Future Work===
SEI consists of feature extraction/estimation and emitter classification using clustering. This research only covers the feature estimation aspect of SEI. Future work could include covering the remaining aspects of SEI using CNNs.  One area of research that could be considered in the future is the use of CNNs for clustering. The network would be designed as a classifier, rather than as an estimator. The classification accuracy of CNNs could then be compared to traditional clustering techniques.

== References ==
[1]	A. Krizhevsky, I. Sutskever and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 807-814.<br>
[2]	I. Goodfellow, Y. Bengio and A. Courville, “Deep learning,” MIT Press, 2016.<br>
[3]	L. J. Wong, “On the use of convolutional neural networks for specific emitter identification”, Masters Thesis, Virginia Polytechnic Institute and State University, Blacksburg, Virginia, Apr. 2018.<br>
[4]	L. J. Wong, W. C. Headley and A. J. Michaels, “Emitter identification using CNN IQ imbalance estimators”, https://arxiv.org/pdf/1808.02369.pdf<br>
[5]	L. J. Wong, W. C. Headley and A. J. Michaels, “Estimation of transmitter I/Q imbalance using convolutional neural networks,”, in proc. IEEE 8th Annual Computing and Communication Workshop (CCWC), Las Vegas, NV, Jan. 2018, pp 948-955.<br>
[6]	T.J. O’Shea, J. Corgan and T.C. Clancy, “ Convolutional Radio Modulation Recognition Networks”, Engineering Applications of Neural Networks, EANN 2016. Communications in Computer and Information Science, Srpinger, 2016, 629