Heriot-Watt University Research Gateway

Power Amplifier enabled RF Fingerprint Identification

Citation for published version:Li, Y, Ding, Y, Goussetis, G & Zhang, J 2021, Power Amplifier enabled RF Fingerprint Identification. in 2021IEEE Texas Symposium on Wireless and Microwave Circuits and Systems (WMCS)., 9493272, IEEE, 2021IEEE Texas Symposium on Wireless and Microwave Circuits and Systems, Texas, United States, 18/05/21.https://doi.org/10.1109/WMCS52222.2021.9493272

Digital Object Identifier (DOI):10.1109/WMCS52222.2021.9493272

Link:Link to publication record in Heriot-Watt Research Portal

Document Version:Peer reviewed version

Published In:2021 IEEE Texas Symposium on Wireless and Microwave Circuits and Systems (WMCS)

Publisher Rights Statement:© 2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all otheruses, in any current or future media, including reprinting/republishing this material for advertising or promotionalpurposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of anycopyrighted component of this work in other works.

General rightsCopyright for the publications made accessible via Heriot-Watt Research Portal is retained by the author(s) and /or other copyright owners and it is a condition of accessing these publications that users recognise and abide bythe legal requirements associated with these rights.

Take down policyHeriot-Watt University has made every reasonable effort to ensure that the content in Heriot-Watt ResearchPortal complies with UK legislation. If you believe that the public display of this file breaches copyright pleasecontact open.access@hw.ac.uk providing details, and we will remove access to the work immediately andinvestigate your claim.

Download date: 09. May. 2022

978-0-7381-1299-2/21/$31.00 ©2021 IEEE

Power Amplifier enabled RF Fingerprint Identification

Yuepei Li*#1, Yuan Ding #2, George Goussetis #3, Junqing Zhang**4 # Institute of Sensors, Signals and Systems (ISSS), Heriot-Watt University, UK

* Institute of Digital Communications, School of Engineering, University of Edinburgh, UK **Department of Electrical Engineering and Electronic, University of Liverpool, UK

1yl12@hw.ac.uk, 2 yuan.ding@hw.ac.uk, 3g.goussetis@hw.ac.uk, 4Junqing.Zhang@liverpool.ac.uk

Abstract—The development of radio frequency fingerprint (RFF) identification technique aims to identify and classify wireless devices by exploiting their unique radio frequency (RF) features. This paper proposes a wireless device RFF identification scheme, which relies on the non-linear memory effect resulting from the concatenation of two matched root-raised cosine (RRC) filters with a non-linear power amplifier (PA) inserted in-between. This unique feature can be extracted from the distorted constellation diagrams, which is processed into density trace figure. Classification algorithm based on image recognition method is developed to exploit the non-linear memory effect from the density trace figure. Simulation setup is carefully designed, and the results validated the effectiveness of our proposed RFF feature classification approach. In the range of SNR 1 dB to 25 dB, the overall identification accuracy rate exceeds 99%.

Index Terms— Convolution neural network (CNN), radio frequency fingerprint (RFF), power amplifier non-linearity.

I. INTRODUCTION Wireless technologies have been evolving rapidly and

been widely adopted in many applications such as cellular communications, navigation systems, wireless control, and unmanned aerial vehicles (UAVs) sensing, etc. In the booming Internet of Things (IoT) era, wireless connectivity is especially important - IoT devices will be three times as high as the global population by 2021 [1]. On the other hand, the explosive growth of wireless connectivity is accompanied with great threats, especially cyber-attacks like Sybil attacks [2], masquerade attacks and resource depletion [3]. To initiate these attacks, the malicious users imitate legitimate transmitters or receivers to invade their system in order for information interception. Hence, a robust and reliable security system becomes a necessity.

The conventional device identification methods commonly use IP and/or MAC addresses, electronic serial number (ESN), international mobile station equipment identity (IMEI) number, and mobile identification number (MIN) as unique identifiers. However, they are modifiable via software [4] raising significant security concerns. Once the device identity is compromised, multiple attackers can masquerade as legitimate users. To address this issue, the radio frequency fingerprint (RFF) identification technique becomes a promising candidate, which identifies wireless devices through features intrinsically existed in radio frequency (RF) stages [5] [6]. The hardware characteristics

are the results from the uncontrollable factors in manufacturing process for any wireless device, e.g., the slight differences of semiconductor doping density. Thus, they are difficult to be modified, and their operation is not limited to different vendors.

Fig. 1. Illustration of RFF identification system blocks.

To analyze the reliability and uniqueness of RFF, the origins of various features need to be point out first. The model of the RFF identification is illustrated in Fig. 1. The receiver acts as an identifier, which captures the signals radiated from the transmitters (to be identified) and extracts features that originated from the RF stages at the transmitter end. Polak et al. exploited the unique non-linear I/O characteristics of power amplifiers (PAs), where Volterra series behavior model was used [7]. The result showed that this approach works effectively when the signal to noise ratio (SNR) is greater than 35 dB, unrealistic in most wireless links. In [8], the frequency representation (magnitude, cartesian and polar representation) of the received symbols was used as the feature, which is also based on the PA non-linearity. The results showed that the classification of this work performs well for different modulation system types. But the impact of PA non-linearity on received symbols was not fully discussed. In [9], the imperfection of RF oscillators was analyzed, for example the phase noise of phased-locked-loops, was found useful for device identification. In [10], RF front-end non-linearity feature in frequency domain is implemented, the demonstration of RFF is influenced by the type of modulation system and digital shaping filter at the transmitter side. In [11], a method called differential constellation trace figure (DCTF) was developed. The DCTF is able to convert I/Q phase mismatch, I/Q DC offset, I/Q gain imbalance and carrier frequency offset into a 2-D image for further classification. It was validated on 54 ZigBee devices, showing an accuracy of over 99% when SNRs are greater than 15 dB. In [12] [13], a method with

silicon physical unclonable function (PUF) and RFF was exploited in wireless device authentication applications, the PUF’s unique manufacture process variations make the classification system extremely robust.

The proposed RFF identification approach in this paper explores the non-linear memory effect which generates by the cascade of two matched raised root cosine (RRC) filters and a non-linear PA inserted in-between. In addition, the use of Bessel- Fourier model provides an opportunity to accurately analysis the characteristic of the PA. Motivated by the success of convolution neural network (CNN) image recognition, the non-linear memory effect of different PAs can be identified. In Section II, the non-linear memory effect is discussed, and the feature is presented as a density trace figure (DTF) in the form of distorted constellation diagram. In Section III, the CNN algorithm is then employed to classify the DTF, with simulation procedures and results presented in Section IV. Finally, Section V concludes the paper.

II. UNIQUE PHY LAYER RFF FEATURE

A. RF Link Model Considering the simplified link model of our proposed

RFF system shown in Fig. 2, a bit sequence b is first mapped to complex symbols through the modulator with the resulting symbols denoted as: sm =𝑠!" +j𝑠!

# , where the subscript m refers to the symbol index, 𝑠!" and 𝑠!

# are the in-phase and quadrature components. In most wireless communication systems digital pulse-shaping raised-cosine (RC) filters are widely used in order to reduce crosstalk in adjacent frequency bands. In practice, the RC filter is split into a pair of matched RRC filters, with one implemented at the transmitter side, and the other at the receiver side. The sm goes through a transmitter RRC, resulting the baseband sequence u[n], ready for up-conversion, amplification, and radiation. Here n is the index of digital samples.

Fig. 2. Simplified link model of the proposed RFF identification system.

B. Gaussian Distributed Inter-symbol-interference In practical systems, two cascaded RRC filters are

unable to represent an ideal RC filter since the RRC filters are implemented as finite impulse response (FIR) filters instead of infinite impulse response (IIR) filters. Consequently, inter-symbol interference (ISI) exists. In a

linear system, namely if the PA in-between two RRCs is linear, this ISI approximately follows zero-mean Gaussian distribution, see our simulation results for an example 16-QAM transmission shown in Fig. 3. To facilitate our discussion, the RRC filter time-domain response hRRC is shown in (1),

(1)

where β (0≤β≤1) and T are the roll-off factor and the symbol period of the RRC filter, respectively. The number of RRC filter taps is given by Ntap= D·S [14], where D represents the filter span, which determines the duration of filter response in number of symbols, and S refers to the up-sampling rate. T and β determine the filter bandwidth BRRC, i.e., BRRC = (1+β)/T. In Fig. 3(b), the variance of the ISI is affected by the RRC filter parameters, especially of β. When β = 0.1, the received constellation diagram suffers great ISI which is Gaussian distributed. In contrast, when the β = 0.5, the ISI is greatly suppressed.

(a)

(b) (c)

Fig. 3. (a) Cascaded RRC filters link model, (b) received 16-QAM signals (no PA insert) without channel noise, (RRC parameters: β = 0.1, D = 8, T = 100 ms), and (c) received 16-QAM signals (no PA insert) without channel noise, (RRC parameters: β = 0.5, D = 8, T = 100 ms).

C. Power Amplifier Behavior Model From the above discussion of the two cascaded matched

RRCs, it is shown that two RRCs can generate zero-mean Gaussian distributed ISI. Before introducing the memory effect of non-linear PA insertion, the PA behavior models are first presented here. Generally speaking, the PA behavior model is used to describe PA characteristics, in this case the amplitude-to-amplitude modulation (AM/AM) curve, in an analytical expression to fit the measurement data. There exist several different behavior models, such as

2

(1 ) sin(1 ) /cos[ ]2 4 /( )

1 (4 / )RRC

t t TT t Th t

t TT

b p b pb b

bp

+ × - ×+

× × ×=- × ××

Saleh’s [15], Ghorbani’s [16], Rapp’s [17], and Bessel-Fourier [18], etc. Among them, the Bessel-Fourier model reported in [18] [19] has a better fitting at the cost of computation complexity, which can be expressed as

(2)

where ρ is the magnitude of the PA input signal, and J1(·) refers to the Bessel function of the first kind with L being the number of Bessel series. bl and α are the coefficients to be determined to fit the measured AM/AM curve, and l is the index of the coefficient. In particular, the AM/AM curve of a PA (model: ZJL-4HG+ from mini-circuits) was measured and fitted using the Bessel-Fourier behavior model, seen in Fig. 4. To obtain an accurate fitting in this example, L of 21 was used.

Fig. 4. Measured and Bessel-Fourier fitted AM/AM curves of a PA (model: ZJL-4HG+).

D. Non-Gaussian Distributed Inter-symbol-interference At the receiver (identifier) side, the received signal is

down-converted to baseband first, and then matched RRC filtering and demodulation are performed. We utilize the received signals after the RRC filtering for RFF feature extraction, see constellation diagram ‘Plot 2’ in Fig. 2. For transmitter i, the received signal yi can be expressed as

yi = Hi xi + vi , (3)

where xi represents the amplified transmit signal at transmitter i, Hi is the corresponding channel coefficient, and vi is additive white Gaussian noise (AWGN). For an example, when the PA ZJL-4HG+, whose AM/AM curve is presented in Fig. 4, is used at the transmitter side. In Fig.5, the received constellation diagram of the example 16-QAM transmission is plotted for noiseless and noisy (SNR = 20 dB) cases. Here the average transmit signal power is kept identical as 15.2 dBm (seen in Fig. 4) for comparison purpose. The signal attenuation due to the channel is normalized out. In Fig. 5, it can be seen that the received signals with a non-linear PA inserted (i.e., blue clouds) suffer much greater ISI as compared to the linear counterpart (i.e., red dots). And crucially, the distorted constellation symbols have irregular shapes, which are

affected by PA characteristics (see the resulting constellations associated with two different PAs in Fig. 6). These simulations were implemented with the RRC parameters β = 0.5, D = 8 and T = 100 ms. In addition, the system added with non-linear PA creates much greater ISI, and this non-Gaussian ISI is dependent on the PA characteristics. In addition, inserting the RRC filter can effectively reduce the ISI, so that the non-linearity effect on the constellation diagram is more obvious. Overall, these results provide us with the opportunity to identify different PAs by distinguishing the subtle differences in the shape of symbol constellations, thereby identifying different devices.

(a) (b)

Fig. 5. Received signal constellations with (see blue clouds) and without (see red dots) the non-linear PA inserted at transmitter end in (a) noiseless case, (b) noisy case with SNR of 20 dB. The channel attenuation is normalized out in both systems.

(a)

(b) (c)

Fig. 6. (a) Bessel-Fourier fitted AM/AM curves of two different PAs (Model: ZX60-V62 and TRF37). The output power of both systems are kept identical as 15.2 dBm, and the attenuation is normalized out (b) Received constellation diagrams of the PA model ZX60-V62 with no AWGN. (c) Received constellation diagrams of the PA model TRF37 with no AWGN.

/ 11

( ) ( )L

AM AM ll

B b J lr a r=

= × ×å

E. Density Trace Figure The received signal constellations are further converted

to DTF, which is used to illustrate the statistical density of each constellation symbol in the diagram. The density level, and hence the color, of each constellation symbol is determined by the probability of data appearing in the pre-defined grids of fine resolution. The normalize point density P(n) is calculated as follow:

(4)

where K is the total number of received symbols, d(n) and q(n) are the functions to calculate real number of horizontal axis and imaginary number of vertical axis of the nth symbol. r is the radius of the selected circular region when determining the normalized point density, and g(·) is a scale function with output normalized from 0 to 1. Therefore, the function of density scale is defined as the distance between the nth and the 1st symbol of the received sample, and it is compared with the r. In other words, if that distance matches the condition (less than r), the density scale is closed to 1, otherwise, the density scale is closed to 0. The resulting DTFs (see Fig. 7) are saved as JPEG file with the pixel size of [Height = 400, Width = 500], and the dots per inch (dpi) of each trace figure is set to 200. The file size of the DTF is 10 Kilobyte (KB).

(a) (b)

Fig. 7. Obtained example DTFs for systems with (a) the PA model ZX60-V62 and (b) the PA model TRF37.

III. RFF CLASSIFICATION APPROACH A CNN classification algorithm was employed to classify

RFF features that are extracted from wireless devices. In Table I, the proposed CNN architecture consists of 3 convolutional layers, 2 max pooling layers and 1 fully connected layer. The size of the input image was set to 400 × 500 × 3 (Height×Width×RGB). The image filter size in convolutional layers were set to [3 × 3] to capture the local details of RFF in constellation diagram. The number of image filters respectively were chosen as 16, 32 and 64, and these correspond to the number of neurons in the convolution layers which connects to the same region in the

input. In the convolutional layer also contains Rectified linear unit (ReLU) activation function, it is able to make negative values amened to 0. The max pooling layers can help reduce the dimension of the input image size by combining the outputs of neuron clusters into a single neuron at the prior layer. The fully connected layer uses the Softmax activation function to perform the classification among the target devices. The total trainable parameter is up to 1616 + 51232 + 204864 + 800000k, where k refers to the index of target devices participate in CNN training. In addition, CNN efficiency and accuracy are affected by batch sizes, maximal epochs and initial learning rates. In our classification design, batch sizes of 64 were used, maximal epochs and initial learning rate were 20 and 0.1, respectively.

TABLE I. LAYERS OF THE CNN CLASSIFICATION

Layer Matrix dimension Trainable parameters

Input Image size: [400×500×3] -

Convolution 2- D + ReLU function

Filter size: [10 × 10], No. filter: 16 1616

Max pooling [2×2] - Convolution 2- D + ReLU function

Filter size: [10×10], No. filter: 32 51232

Max pooling [2×2] - Convolution 2- D + ReLU function

Filter size: [10×10], No. filter: 64 204864

Fully connected Final image size: [100 × 125 × 3] 800000

Generally, the supervised CNN algorithm includes training and classification stages. See the operation diagram of CNN classification system shown in Fig.8. Once the distorted diagram extracted from the transmitter, it will be processed into the DTF. A full dataset includes DTFs of valid transmitter, and the SNR range of the DTF is 1 dB to 25 dB. In the training stage, the training dataset is processed by CNN in order to find key patterns. In other words, the proposed CNN model learns network parameters (neurons with weights and biases) of the training DTFs, and uses each new training example to continuously update and store them in the database. In addition,1900 units of DTF in the dataset were randomly divided into 80% training dataset (1520 units of DTF) and 20% (380 units of DTF) validation dataset, after this the feature of the training dataset will analyze and store into the database with its label. In classification stage, we simulated another set of DTF as test dataset and imported those into CNN for identity verification. The SNR of the test dataset is the same as the SNR of the training dataset, ranging from 1 dB to 25 dB. Therefore, 7500 units of DTF as test data in total. Identification accuracy η is the classifications that have been processed, and the number of

1[| ( ) ( ) | & | ( ) ( ) | ]

( ) =

K

zg d n d z r q n q z r

P nK

=

- < - <å

times the classification has been corrected. Finally, accessed devices will be identified.

Fig. 8. CNN training and classification process.

IV. NUMERICAL RESULTS The proposed model is simulated in MATLAB and is

evaluated using the identification accuracy rate η for different SNR scenarios. 16-QAM data transmission was assumed in our simulation. In this paper, 5 different PAs were selected, and their AM/AM characteristics were measured and fitted using the Bessel-Fourier model, with the length of Bessel series L = 21, seen in Fig. 9. From the plotted results we can observe that they experience different gain and non-linear properties. In order to exploit their non-linear properties for the RFF, in our simulations the average transmit signal power is kept identical as 15.2 dBm for the PA models ZX60-V62, ZJL-4HG+, SKY67023 and TRF37, while the average transmit signal power for the PA ZJL-4G+ was set to 12 dBm (since it saturates at 13 dBm).

Fig. 9. PA AM/AM characteristics fitted using the Bessel-Fourier model. (Mini-Circuits: ZJL-4G+, ZX60-V62, ZJL-4HG+; SKYWORKS: SKY67023; Texas Instruments: TRF37).

In the identification stage, we assume 5 individual transmitter systems with the 5 PAs (seen in Fig. 9). These transmitters send 16-QAM modulated signals to the receiver as the identifier for classification. In addition, the channel attenuation was normalized out. After the distorted constellation diagram was extracted, it was processed into DTF. The DTFs of each PA with SNR range from 1 dB to 25 dB (2 dB per step) were used as the test dataset, thus, 7500 DTFs were tested in total. The purpose of this experiment is to verify the reliability of the classification system. In addition, the CNN system was trained with 1520 units of DTF. In Fig. 10, the identification accuracy versus

with the SNR, the result was averaged over 20 trials, and then calculated and plotted. In this classification result, even in the case of low SNR, 5 transmitters can be identified with an overall identification accuracy rate exceeds 99%. Since the SNR of the test data set is similar to the SNR of the trained data set, a higher accuracy rate is achieved.

(a)

(b)

Fig. 10. CNN classification results vs. SNR in the range of 1 dB to 25 on 7500 units of test data, the overall identification accuracy rate is 99%. (a) plot zoom out. (b) plot zoom in.

Fig.11 and Fig.12 respectively shows the CNN confusion matrix among 5 transmitter systems at SNR of 9 dB and 5 dB. As mentioned early, the test dataset has 7500 units of DTF, therefore, 500 units of DTF per SNR value (100 units of DTF for each transmitter). In the figures, where 100 is represented 100 units of DTF were accurately identified. Even under this relatively low SNR condition, the extracted feature is identifiable with accuracy 100% and no misidentification in this case. It demonstrates the good reliability of the proposed classification system.

Fig. 11. CNN confusion matrix for identification of 5 transmitters at SNR of 9 dB on 500 units of test data. Overall accuracy is 100%.

Fig. 12. CNN confusion matrix for identification of 5 transmitters at SNR of 5 dB on 500 units of test data. Overall accuracy is 100%.

V. CONCLUSION AND FUTURE WORK In this paper, we elaborated the non-linear memory

effect generated by the cascade of two matched RRC filters with a non-linear PA in-between. The PA non-linear effect at the transmitter is significant cause of the discrepancy in RFF. The CNN-based algorithm was adopted for classification. The simulated classification results validated the effectiveness of our proposed RFF system. In the range of SNR 1 dB to 25 dB, the overall identification accuracy rate exceeds 99%. The proposed system will be experimentally demonstrated in the future.

ACKNOWLEDGMENT Mr Yuepei Li’s work was supported by Heriot-Watt

University DTP scholarship. Dr Yuan Ding’s work was supported by EPSRC (UK) under Grant EP/L020009/1 and EP/V002635/1.

REFERENCES [1] (2021) Cybercrime magezone. [Online]. Available:

https://cybersecurityventures.com/. [2] J. Douceur, “The sybil attack,” in Proc. 1st IPTPS.,

Cambridge, MA, pp. 251-260, Mar. 2002. [3] D. B. Faria and D. R. Cheriton, “Detecting identity based

attacks in wireless networks using signalprints,” in Proc. 5th ACM Workshop Wireless Security., Los Angeles, California, USA, 2006.

[4] J. Yang, Y. Chen, and W. Trappe, “Detecting sybil attacks in wireless and sensor networks using cluster analysis,” in Proc. 5th IEEE Int. Conf. Mobile Ad Hoc Sensor Syst., Atlanta, GA, USA, pp. 834–839, Sep. 2008.

[5] Q. Xu, R. Zheng, W. Saad, and Z. Han, “Device fingerprinting in wireless networks: challenges and opportunities,” IEEE Commun. Surveys & Tutorials., vol. 18, no. 1, pp. 94–104, Firstquarter 2016.

[6] P. Scanlon, I. Kennedy, and Y. Liu, “Feature extraction approaches to RF fingerprinting for device identification in femtocells,” Bell Labs Tech. J., vol. 15, no. 3, pp. 141–151, Dec. 2010.

[7] A. Polak, S. Dolatshahi, and D. Goeckel, “Identifying wireless users via transmitter imperfections,” IEEE J. Sel. Areas Commun., vol. 29, no. 7, pp. 1469–1479, Aug. 2011.

[8] S. Hanna and D. Cabric, “Deep learning based transmitter identification using power amplifier nonlinearity,” in Int.

Conf. Computing, Netwroking & Commun., Honolulu, HI, USA, 2019, pp. 674–680.

[9] A. Polak and D. Goeckel, “Wireless device identification based on RF oscillator imperfections,” IEEE Trans. Forensics Security., vol. 10, no. 12, pp. 2492–2501, Dec. 2015.

[10] W. Wang, Z. Sun, S. Piao, B. Zhu and K. Ren, “Wireless physical-layer identification: modeling and validation,” IEEE Trans. Forensics Security., vol. 11, no. 9, pp. 2091–2106, Sept. 2016.

[11] L. Peng, J. Zhang, M. Liu, and A. Hu, “Deep learning based RF fingerprint identification using differential constellation trace figure,” IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 1091–1095, Jan. 2020.

[12] B. Chatterjee, D. Das, S. Maity and S. Sen, “RF-PUF: Enhancing IoT security through authentication of wireless nodes using in-situ machine learning,” in IEEE Internet. Things J., vol. 6, no. 1, pp. 388-398, Feb. 2019.

[13] Q. Zhou, Y. He, K. Yang and T. Chi, “Exploring PUF-controlled PA spectral regrowth for physical-layer identification of IoT nodes,” 2021 IEEE Int. Solid-State Circuit Conf., San Francisco, CA, USA, 2021, pp. 204-206.

[14] M. Cañavate Sánchez, A. Segneri, S. Kosmopoulos, Q. Zhu, T. Tsiftsis, A. Georgiadis, and G. Goussetis, “Novel data pre-distorter for APSK signals in solid-state power amplifiers,” IEEE Trans. Circuits & Syst., vol. 66, no. 10, pp. 4044–4054, Oct. 2019.

[15] A. Saleh, “Frequency-independent and frequency dependent nonlinear models of TWT amplifiers,” IEEE Trans. Commun., vol. 29, no. 11, pp. 1715–1720, 1981.

[16] A. Ghorbani and M. Sheikhan, “The effect of solid state power amplifiers (SSPAs) nonlinearities on MPSK and M-QAM signal transmission,” in 6th Int.Conf. Digital Proc. Signals Commun., pp. 193–197, 1991.

[17] C. Rapp, “Effects of HPA-nonlinearity on a 4- DPSK/OFDM-signal for a digital sound broadcasting system,” in 2nd Eur.Conf. Satellite Commun., pp. 179–184, 1991.

[18] J. Fuenzalida, O. Shimbo, and W. Cook, “Time domain analysis of intermodulation effects caused by nonlinear amplifiers,” Comsat Technical Review., vol. 3, no. 1, pp. 88–138, 1973.

[19] M. Cañavate Sánchez, A. Segneri, A. Georgiadis, S. Kosmopoulos, G. Goussetis, and Y. Ding, “System performance evaluation of power amplifier behavioural models,” in IET Active Passive RF Dev. Seminar., London, UK, Apr. 30, 2018.