channel availability prediction in cognitive radio ... · pdf filechannel availability...
TRANSCRIPT
Contemporary Engineering Sciences, Vol. 10, 2017, no. 12, 593 - 605
HIKARI Ltd, www.m-hikari.com
https://doi.org/10.12988/ces.2017.7758
Channel Availability Prediction in Cognitive Radio
Networks Using Naive Bayes
Hans Marquez
Universidad Distrital Francisco José de Caldas, Bogotá, Colombia
Cesar Hernández
Universidad Distrital Francisco José de Caldas, Bogotá, Colombia
Diego Giral
Universidad Distrital Francisco José de Caldas, Bogotá, Colombia
Copyright © 2017 Hans Marquez, Cesar Hernández and Diego Giral. This article is distributed
under the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Abstract
This work develops a channel availability predictive model that allows taking
advantage of opportunities in the cognitive radio network’s spectrum in the most
efficient way. The developed scheme creates an availability prediction matrix for
each available channel in the GSM band, which is used to determine the potential
bandwidth and availability time slots to optimize the channel allocation policies.
The allocation model contains a training process in charge of preparing the
prediction algorithm so it can make predictions that are more reliable.
Additionally, the prediction procedure uses the Naïve Bayes algorithm to estimate
the availability in each available frequency channel. This facilitates broadcasting
for secondary users. Measurements were computed for the average bandwidth, the
average delay and the prediction error. The obtained results were assessed with
real spectral occupation data in the GSM frequency band. The developed model
shows a low prediction error, which leads to establishing optimal channel
allocation mechanisms hence minimizing the failed handoffs due to channel
occupation by primary users.
Keywords: availability, cognitive radio, naïve Bayes, prediction
594 Hans Marquez et al.
1. Introduction
Currently, one of the most difficult tasks in the communications sector in general
is the fixed allocation of the spectrum, which has caused problems in the optimal
use of the spectrum going through excessively used areas to areas without
extensive use [1]. This has given birth to the study of cognitive radio networks
(CRN) that would allow passing from a rigid model for fixed allocation to one
that performs a flexible use of the spectrum hence optimizing the scarce resources
within wireless networks [2]. This flexibility would improve the service quality of
wireless technologies since an opportunistic use of the spectrum would be more
efficient. [3].
When the spectrum is not under use by a primary user (PU), secondary users
could take these spectral opportunities. However, such opportunities must be
relinquished once the PU requires them [4]. This restriction forces the SU to move
on to another spectral opportunity that is free. The described process is named
spectral handoff. [2].
To minimize the interference for the PU, the handoffs carried out by the SU must
be performed before the PU arrives so that the PU’s transmissions are not greatly
affected by the SU’s spectral opportunities [2]. To solve this problem, a predictive
model is proposed to foresee spectral occupation, the probability of a PU’s arrival
and optimizing the channel allocation process according to the previous metrics.
[5].
This article is structured as follows: section 2 presents related work, section 3
describes the development of the model, section 4 includes all the results obtained
and section 5 finishes off with a set of conclusions.
2. Related Work
In [6] a two-stage cognitive process is proposed aiming to learn from the abilities
of the physical layer under different channel conditions optimizing the package
delivery in a multi-radio antenna. In the first stage, it learns the features of the
available techniques. In the second stage, it defines the settings that most satisfy
the radio’s goals according to the channel’s conditions. A design based on the
Bayes rule will be the baseline for future comparisons. This work also studies the
Naïve Bayes, Semi-Naïve Bayes and binary search models that offer proper
learning and optimization alternatives for the reference design, requiring the
estimation of less parameters. Nevertheless, performance optimization is
sacrificed to increase speed and save memory.
In [7], cognitive engines in wireless networks are studied for environments that
require that the system dynamically selects a channel to transmit. Although the
different selection strategies focused towards a wide range of networks may be
Channel availability prediction in cognitive radio networks 595
organized in a smaller group of simple strategies, a universal DSA network
simulator is used to model the actions of a DSA network with these simplified
strategies. The channel selection is also described with the purpose of establishing
the behavior of secondary users, specifically those that have started transmitting.
This project centers in the estimation of the low-cost channels’ selection since the
adopted methods only require characteristics and monitoring tools already present
in a DSA network.
In [8], the dependency existing in cognitive radio networks in comparison to the
capabilities of artificial intelligence to perform a variety of tasks. An example of
these tasks is the self-organization and the mitigation of interference. One of the
objectives of the cognitive radio applications is to trying to decentralize and
distribute network functions between all participating nodes. Consequently, any
network node must be capable of coordinating functions between its peers and
unknown transmitters. Therefore, it is suitable that the individual nodes are able to
predict the future states of the unknown transmitters with the purpose of
optimizing its operation. Methods are studied in this article to identify the
cognitive behavior in an unknown transmitter and the prediction of future states.
The problem of the Universal DSA Network Simulation (UDNS) context is
discussed and two classification algorithms are presented to achieve this.
In [9] it is indicated that the detection of the spectrum is an effective method to
improve the performance of the availability detection. In spite of this
improvement, the authors indicate that the cooperation between cognitive radio
users may introduce a variety of general expenses such as additional detection
time, delays, energy consumption that affect or limit the reachable cooperative
gain. As a solution, they propose a learning algorithm based on a fusion algorithm
that can offer real time training. It is rooted under a learning machine that uses
statistical tests for energy detection to determine to the presence or absence of the
PU transmission and predict new statistical frames with this information. A
comparative simulation of four supervised machine-learning classifiers is detailed
over 1000 test frames: K-nearest neighbor (KNN), support vector machine
(SVM), Naïve Bayes (NB) and Decision Tree (DT). They are previously trained
with a set of 1000 frames. It is shown that KNN and DT surpass the other two
classifiers.
In [10], the problem of accessing the channeled dynamic spectrum is considered.
There are M secondary users with N ≥ M orthogonal channels. Therefore, each
user requires a channel to transmit so to guarantee that it will not conflict with the
channels assigned to other users. Nevertheless, due to geographic dispersion, each
secondary user can observe multiple occupation behaviors by primary users in
each channel. A new algorithm is presented with polynomial storage and
computation for this problem and it is discussed how the results offer a non-trivial
generalization of theoretical knowledge.
In [11], a stable trajectory selection scheme for cognitive radio sensor networks is
described. The routing protocol uses a priori delivery information and gives a
novel routing metric is derived from the Naïve Bayes interference.
596 Hans Marquez et al.
The metric is then used to establish the stability of the trajectory considering the
delivery probability. Extensive simulations show that the proposed routing
protocol is efficient to find the most likely delivery route and increase the
system’s performance.
In [12], a modulation classification algorithm is designed for MIMO systems
based on the pattern recognition focus. This is performed using higher-order
statistics (HOS) and a Bayesian network classifier. With the means of assessing
the efficacy of the Bayesian network methods, a comparative study is carried out
between these three classifiers: Naïve Bayes with discretization (NBD), the
augmented Naïve Bayes tree (TAN) and the decision tree (J48). The functional
characteristics of the receiver (ROC), the identification probability and the
training time show that the NBD and TAN classifiers reach a performance similar
to the J48 classifier. Hence, these classifiers can be used to distinguish between
different types of M-ary modulation allowing a better monitoring of the signals
intercepted in the wide band technologies with low complexity.
3. Proposed Prediction Model
The proposed prediction model can be seen in Figure 1. The first block called
“Occupation database of the spectrum” includes real occupation data of the
spectrum corresponding to the GSM band (824 MHz – 849 MHz). This block is
the entry point for spectrum data processing that is in charge of defining the
occupation or availability of each channel in the GSM band according to the false
alarm probability equation.
The rectangular area corresponds to the proposed model, which includes two
algorithms: (1) Naïve Bayes algorithm, (2) Channel allocation prediction. The
first one’s function is to train the algorithm for a 10-minute period by using
training variables or availability probability and the average time of availability.
The purpose is to compute the parameters cost and gradient, which are necessary
to adjust the predictor. The second algorithm assigns the channel occupation
through the assignment of “1” and “0” which outputs an availability prediction
matrix.
Figure 1. General structure of the proposed model
Channel availability prediction in cognitive radio networks 597
3.1 Defining the Simulation Time and the Training Occupation Data
The model assigns the resources according to the existing demand considering the
existing time intervals for simulation. Each time interval lasts 300 milliseconds
and the total simulation time is 10 minutes.
Additionally, the model has a spectral occupation data matrix with input data for
the GSM frequency bands that mark a significant difference in terms of the
measured traffic at different time intervals. Such differentiation refers to the
information and occupation volumes that are shown in each channel, i.e. low and
high traffic data. The assessment of the model is obtained through this
differentiation over two different scenarios with their own characteristics.
3.2 Naïve Bayes Algorithm
One of the main considerations for the selection of the prediction model is that it
has multiple characteristics or criteria that can improve prediction. This is
explained due to the training of the prediction model may take into account
information or criteria such as the availability probability (AP) and the average
availability time (AAT), among other metrics, that can enhance prediction.
According to the previous paragraph and considering the Naïve Bayes theorem, it
can be stated that the independent variables (or predictors in the specific case)
would be the availability probability and the average availability time whereas the
dependent variable will be the channel availability. According to this, the Naïve
Bayes prediction model works fine on predicting various classes and assumes the
independence between them.
In simple terms, a Naïve Bayes classifier assumes that the presence of one
characteristic in particular does not relate in any way with the presence of any
other characteristic. Even if one these characteristics depend from each other or
the existence or other characteristics, all these properties contribute independently.
One of the advantages is the utility to operate over large datasets and even surpass
highly sophisticated classification methods.
The Bayes theorem allows the calculation of the posterior probability P (c | x),
P(c), P(x) and P (x | c) in equation (1).
𝑃(𝑐|𝑥) =𝑃(𝑥|𝑐)𝑃(𝑐)
𝑃(𝑥) (1)
Here,
P (c|x) is the posterior probability of class c (c, target) given the predictor (x,
attributes)
P (c) is the previous probability of the class
598 Hans Marquez et al.
P (x|c) is the predictor’s probability given the class
P (x) is the predictor’s probability
According to equation (1) and considering our independent variables AP and
AAT as it was described in previous paragraphs as well as the dependent variable
or class (in the specific case, it is the channel availability that will be denoted as
occupied or available), we have equations (2) and (3).
𝑝𝑜𝑠𝑡𝑒𝑟𝑖𝑜𝑟(𝑜𝑐𝑐𝑢𝑝𝑖𝑒𝑑) =𝑃(𝑜𝑐𝑐𝑢𝑝𝑖𝑒𝑑)𝑝(𝑇𝐸𝐷|𝑜𝑐𝑢𝑝𝑝𝑖𝑒𝑑)𝑝(𝑃𝐷|𝑜𝑐𝑐𝑢𝑝𝑖𝑒𝑑)
𝑒𝑣𝑖𝑑𝑒𝑛𝑐𝑒
(2)
𝑝𝑜𝑠𝑡𝑒𝑟𝑖𝑜𝑟(𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒) =𝑃(𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒)𝑝(𝑇𝐸𝐷|𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒)𝑝(𝑃𝐷|𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒)
𝑒𝑣𝑖𝑑𝑒𝑛𝑐𝑒
(3)
The evidence would be given by equation (4).
𝑒𝑣𝑖𝑑𝑒𝑛𝑐𝑒 = 𝑃(𝑜𝑐𝑐𝑢𝑝𝑖𝑒𝑑)𝑝(𝑇𝐸𝐷|𝑜𝑐𝑢𝑝𝑝𝑖𝑒𝑑)𝑝(𝑃𝐷|𝑜𝑐𝑐𝑢𝑝𝑖𝑒𝑑)
+ 𝑃(𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒)𝑝(𝑇𝐸𝐷|𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒)𝑝(𝑃𝐷|𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒) (4)
3.3 Channel Allocation Prediction
In this block, the algorithm uses the probabilities generated through the Naïve
Bayes algorithm and the prediction function takes into account the probability
distributions shown in equation (1), and predicts the availability per channel
during simulation time. This process is iterative since 551 channels were
processed for availability prediction. Therefore, the algorithm has to predict the
availability for each one of them.
3.4 Design Model
The proposed model has for input a spectral occupation training matrix. Before it
is used in the predictor’s training process, the spectral information passes through
the spectral information-processing block, which converts the data into
dichotomous series where a “0” represents the occupation of the channel and a “1”
represents the availability of the channel to train the Naïve Bayes algorithm.
The variables AP and AAT that contain the necessary data to establish the
occupation/availability probabilities of the channel are now considered. During
the training process, two matrixes are built from the channel’s AP and AAT and
along with the predict function, each testing data row is classified into one of the
classes (occupied, available) according to the Naïve Bayes classifier and returns
the foreseen class level. These probabilities are assigned to a matrix called
Availability Prediction where the states of the channel are defined by “1”
(available) and “0” (occupied).
Channel availability prediction in cognitive radio networks 599
Once the prediction matrix is created, it is possible to compare the precision of the
channel allocation during the transmission time and assess whether the availability
predictions can lead to a beneficial use of the channels by establishing a more
effective allocation policy.
4. Results
To assess the performance of the proposed algorithm, a complete analysis was
carried out with two types of traffic: high and low traffic. For each type, eight
figures were designed that correspond to the metrics: anticipated handoffs, perfect
handoffs, failed handoffs, successful handoffs, average bandwidth, average delay,
interference handoff, average throughput.
Figures 2 through 9 show the obtained results for every metric previously
mentioned, for both high and low traffic.
4.1 Results Analysis
According to the results, the number of anticipated handoffs in high traffic is
slightly lower than in low traffic supposing a better performance in the spectrum’s
high occupation. This connects with the fact that the algorithm has a more
efficient use of the allocation time before it has to abandon the spectral
opportunity. This also validates the results for the number of perfect handoffs (see
Figures 2 and 3).
Figure 2. Number of anticipated handoffs
600 Hans Marquez et al.
Figure 3. Number of perfect handoffs
Figure 4. Number of failed handoffs
Although the number of accumulated handoffs during simulation was relatively
high for both scenarios, this cost is compensated by the fact that the number of
handoffs with interference was very low in comparison (see Figures 5 and 6).
Channel availability prediction in cognitive radio networks 601
Figure 5. Number of successful handoffs
Figura 6. Number of handoffs with interference
In both scenarios, the behavior of the bandwidth and throughput was relatively
constant (see Figures 7 and 9) thanks to taking advantage of the maximum
channel transmission time according to the prediction performed by the algorithm.
In addition, the delay in both cases was not very significant (see Figure 8) due to
profiting more efficiently from the spectral opportunities.
602 Hans Marquez et al.
Figure 7. Average bandwidth
Figure 8. Accumulated average delay
Figure 9. Average throughput
Channel availability prediction in cognitive radio networks 603
5. Conclusiones
It is concluded that the number of failed handoffs increased in high traffic
scenarios because the algorithm prioritizes the fact of maintaining a low number
of handoffs with interference to avoid the degradation of the transmission for the
network’s primary users. In spite of this increase, the algorithm tries to
compensate this cost with the number of perfect handoffs to guarantee a higher
transmission time for secondary users. The algorithm’s availability prediction is
validated with an increase in the number of perfect handoffs, a stable management
of the bandwidth allocation during transmission time for both traffic scenarios
and, consequently, a stable behavior in the throughput and the number of perfect
handoffs, which confirm that the allocation was effective as well as the
transmission time for each secondary user.
Acknowledgements. The authors would like to thank the Universidad Distrital
Francisco José de Caldas for the support given during the research process of this
project.
References
[1] J. Attridge, An overview of hardware security modules, SANS Institute,
InfoSec Reading Room, 1 (2002), no. 1, 1-10.
[2] H. A. Alkhzaímí and M. M. Lauridsen, Cryptanalysis of the SIMON family of
block ciphers, Technical University of Denmark, 1 (2013), no. 1, 1-26.
[3] P. Valla and J. Kaps, Lightweight cryptography for FPGAs, 2009
International Conference on Reconfigurable Computing and FPGAs, ReConFig
’09., (2009), 225- 230. https://doi.org/10.1109/reconfig.2009.54
[4] A. Bogdanov, L. Knudsen, G. Leander, C. Paarl, A. Poschmann, M. J. B.
Robshaw, Y. Seurin, C. Vikkelsoe, PRESENT: An Ultra-Lightweight Block
Cipher, Chapter in Cryptographic Hardware and Embedded Systems, Springer-
Verlag Berlin Heidelberg, 2007, 450-466.
https://doi.org/10.1007/978-3-540-74735-2_31
[5] R. Azuero, E. Jacinto and J. Castano, A low-memory implementation of 128
aes for 32 bits architectures, En Congreso Argentino de Sistemas Embebidos
CASE 2012, 2012, 67-73.
[6] M. Kumar and A. Singhal, Efficient implementation of advanced encryption
standard (AES) for ARM based platforms, 2012 1st International Conference on
Advances in Information Technology (RAIT), (2012), 23-27.
https://doi.org/10.1109/rait.2012.6194473
604 Hans Marquez et al.
[7] K. M. Abdellatif, R. Chotin-Avot and H. Mehrez, Lightweight and compact
solutions for secure reconfiguration of FPGAs, 2013 International Conference on
Reconfigurable Computing and FPGAs (ReConFig), (2013), 1-4.
https://doi.org/10.1109/reconfig.2013.6732304
[8] J. Pospiil and M. Novotny, Evaluating cryptanalytical strength of lightweight
cipher present on reconfigurable hardware, 2012 15th Euromicro Conference on
Digital System Design (DSD), (2012), 560-567.
https://doi.org/10.1109/dsd.2012.53
[9] E. Kavun and T. Yalcin, RAM-based ultra-lightweight FPGA implementation
of present, 2011 International Conference on Reconfigurable Computing and
FPGAs (ReConFig), (2011), 280-285. https://doi.org/10.1109/reconfig.2011.74
[10] A. Aysu, E. Gulcan and P. Schaumont, SIMON says: Break area records of
block ciphers on FPGAs, IEEE Embedded Systems Letters, 6 (2014), no. 2, 37-40.
https://doi.org/10.1109/les.2014.2314961
[11] S. Feizi, A. Ahmadi and A. Nemati, A hardware implementation of simon
cryptography algorithm, 4th International eConference on Computer and
Knowledge Engineering (ICCKE), 2014, (2014), 245-250.
https://doi.org/10.1109/iccke.2014.6993386
[12] Shuangqing Wei, Jian Wang, Ruming Yin, Jian Yuan, Trade-off between
security and performance in block ciphered systems with erroneous ciphertexts, 8
(2013), no. 4, 636-645. https://doi.org/10.1109/tifs.2013.2248724
[13] C.-P. Fan and J.-K. Hwang, Implementations of high throughput sequential
and fully pipelined AES processors on FPGA, International Symposium on
Intelligent Signal Processing and Communication Systems, 2007. ISPACS 2007,
(2007), 353-356. https://doi.org/10.1109/ispacs.2007.4445896
[14] F. Chih-Peng and H. Jun-Kui, Implementations of high throughput sequential
and fully pipelined AES processors on FPGA, International Symposium on
Intelligent Signal Processinq and Comunication Systems ISPACS 2007, (2007),
353-356. https://doi.org/10.1109/ispacs.2007.4445896
[15] J. J. Tay, M. M. Wong and I. Hijazin, Compact and low power AES block
cipher using lightweight key expansion mechanism and optimal number of s-
boxes, International Symposium on Intelligent Signal Processing and
Communication Systems (ISPACS), 2014, (2014), 108-114.
https://doi.org/10.1109/ispacs.2014.7024435
Channel availability prediction in cognitive radio networks 605
[16] R. Beaulieu, D. Shors, J. Smith, S. Treatman-Clark, B. Weeks and L.
Wingers, The simon and speck families of lightweight block ciphers, Cryptology
Print Archive, Report 2013/404, 4 2013, http://eprint.iacr.org
[17] V. Trujillo-Olaya, T. Sherwood, and Çetin Kaya Koç, Analysis of
performance versus security in hardware realizations of small elliptic curves for
lightweight applications, Journal of Cryptographic Engineering, 2 (2012), no. 3,
179-188. https://doi.org/10.1007/s13389-012-0039-x
[18] Urbano-Molano, V. Trujillo-Olaya and J. Velasco-Medina, Design of an
elliptic curve crypto-processor using optimal normal basis over GF(2233), 2013
IEEE Fourth Latin American Symposium on Circuits and Systems (LASCAS),
(2013), 1-4. https://doi.org/10.1109/lascas.2013.6519014
[19] R. Beaulieu, D. Shors and J. Smith, Stefan Treatman-Clark, Bryan Weeks,
Louis Wingers, The SIMON and speck block ciphers on avr 8-bit
microcontrollers, Chapter in Lightweight Cryptography for Security and Privacy,
Springer, Cham, 2014, 3-20. https://doi.org/10.1007/978-3-319-16363-5_1
[20] N. Hanley and M. O'Neill, Hardware comparison of the ISO/lEC 29192-2
Block ciphers, 2012 IEEE Computer Society Annual Symposium on VLSI
(ISVLSI), (2012), 57-62. https://doi.org/10.1109/isvlsi.2012.25
[21] D. Klinc, C. Hazay, A. Jagmohan, H. Krawczyk and T. Rabin, On
compression of data encrypted with block ciphers, IEEE Transactions on
Information Theory, 58 (2012), no. 11, 6989-7001.
https://doi.org/10.1109/tit.2012.2210752
[22] F. Qatan and I. Damaj, High-speed katan ciphers on-a-chip, 2012
International Conference on Computer Systems and Industrial Informatics
(ICCSII), (2012), 1-6. https://doi.org/10.1109/iccsii.2012.6454511
[23] S. Mane, M. Taha and P. Schaumont, Efficient and side-channel-secure block
cipher implementation with custom instructions on FPGA, 2012 22nd
International Conference on Field Proqrammable Logic and Applications (FPL),
(2012), 20-25. https://doi.org/10.1109/fpl.2012.6339236
Received: July 26, 2017; Published: August 21, 2017