Machine Learning based Intelligent Cognitive Network using Fog Computing

Machine Learning based Intelligent Cognitive Network using Fog Co mputing Jingyang Lu* a , Lun Li a , Genshe Chen a , Dan Shen a , Khanh Pham b , Erik Blasch c a Intelligent Fusion Tec hnology, Inc., 20 271 Goldenrod Ln, Ger mantown, MD, 20876 ; b Air Force Research Lab, Kirtland AFB, NM, 87 117; c Air Force Research Lab, Rome, NY, 1344 1. ABSTRACT In this paper, a Cognitive Rad io Net work (CRN) based o n artificial intelligence is p roposed to distribute the li mited radio spectrum resources more efficiently. The CRN framework can analyze t he time-sensitive signal data close to t he signal source using fog computing with different types of machine learning techniq ues. Depending on the co mputationa l capabilities of the fog n odes, different f eatures and m achine lear ning techniques are chosen to optim iz e spectrum allocation. Also, the co mputing nodes se nd the periodic signal summary which is much s maller than the original sig nal to the cloud so that th e overall system spectrum source allocation strategies are dynamically updated. Applying fog computing, t he s ystem is m ore adaptive to the local environment and rob ust to spectrum changes . As m ost o f the signal data is processed at the fo g level , it further stre ngthens the s ystem security by reducing the communicatio n burde n of the communications net work. Keywords: Cognitive Network, Artificial Intell igence, Machine Lear ning, Cloud Co mputing 1. INTRODUCTION With large growth of wireless services and app lications, there exists an exponential demand for the available rad io spectrum. Some radio spectrum resources are so ld to the pri mary users, which so metimes are not full y used. In some countries, most of available spec trum resources have been fully utilized resulting in t he spec trum scarcity problem. On the other hand, a large amou nt of the licen sed spectrum exper iences lo w utilization based on the recent studies on the actual spectru m u tilization measurements, which indicates inefficiency i ssues existing in the radio spec trum r esources allocation. In the dynamic spectrum acces s (DSA) system, the opportunistic unlicensed secondary users sense the channel to fi nd the available radio spectrum over which they can transmit signal without interfering with the primary users. The secondar y users o f such capability utilize a n intel ligent method called Cognitiv e Radio (CR). There have been different cog nitive rad io applications develo ped such as sp ectrum sen sing [1] , au tonomous learning, channel estimatio n and data detection [2] , user cooperation, modeling, target trac king [3][4][5][6] and reasoni ng. The inef ficiency of spectrum use ha s triggered exciting activities in engineering, ec onomic, and regulation co mmunities to develop im proved spectrum mana gement and spectru m shari ng. T he concept of Cog nitive Rad io (CR ) w as first proposed by Mitola and Ger ald Q. Maguire [7] . The opp ortunistic spectrum acces s model [8] is a land mark e xample. Given the co nstraints of collision probabilit y and overlapping time, the scenario in which secondar y users can opp ortunistically access the unused s pectrum v acated b y idle pri mary users w as i nvestigated. T hree spectrum acce ss sc hemes using different sen sing, b ack-off, and trans mission mechani sms ar e sho wn to achieve indistinguishable secondar y per formance. Once detecting one or multiple spectrum hol es, t he secondary users reconfigure their tra nsmission par ameters such as the carrier frequency, modulation sche me, etc. to operate in the detected spectr um holes. Di fferent d ynamic spectrum access models suc h as d ynamic exc lusive use m odel, open s haring model, and hierarchical access model are methods to im prove spectrum efficiency [9] . The fundamental capacit y limits and associated tr ansmission tech niques for differe nt wireless net work design p aradigms are surveyed [10] . Based on the p rior rules r elated to spectrum usage, the interac tion between cognitive radios’ si gnals a nd the tra nsmissions of noncognitive nodes is of interest . Besides the opp ortunistic sp ectrum access model, the c oncurrent spectrum access model is investigated by many researchers, where second u sers coexist with pr imary user s within a licensed b and. Under the constraint s that the interference caused by second users’ transmitted si gnal is within the prim ary users’ reasonable threshold, three tasks i ncluding radio -scene analysis, chan nel- state estimation, and tran smit-po wer control and dynamic spe ctrum manageme nt are co mpared [11] . Regarding spectrum detectio n and allocation, many di fferent approaches have bee n utilized such as machine learning, artificial i ntelligence, model p rediction [12] ,fuzzy lo gic rea soning [13] , greedy search [14] , and dyn amic programming [15] . A s for machine learning, the system first learns the pattern of the object through the training samples, then the system can perform the op eration o n the data which is n ot e ncountered befor e. Machine learni ng has w ide application areas such as target d etection, image processing [16] , and pattern recognitio n [17] . Fo r exa mple, ne ural networks (NN) can learn the patterns fro m the structured train ing data with different features suc h as po wer spectr um energy, wave for m, modulatio n, and etc. through supervised or unsupervised methods, which ca n be seen in Figure 1 . Based on the weights ac hieved from training process, each nod e or the whole syste m can predict or conduct the cor responding spectrum decision. Figure 1 Different Machine Learning Approaches A contemporar y technique is to co mbine machine learning a nd game -theoretic methods. Since the over co mplexity of the selected model may lead to o verfitting, there is a need to adaptively lear n the sit uation when other users tr y to acces s the system. In the sce nario of o pportunistic spec trum access , it is ai med to find the spectrum holes to tran smit the signa l without interfere nce the pri mary users. Based on the non-coop erative independent d ecisions [18] , two dynamic spectr um access algorithms includi ng H ungarian and greed y sear ch ar e prop osed allowing t he seco nd users to ac cess the a vailable spectrum. T he greedy algorithm can red uce the sear ching co mplexity to linear level b y selecti ng the best candidate i n each iteration is investigated [19] , [20] . It is supposed that the channel availability statistics [21] and the number of secondary users are unkno wn, where th e pro blem is for mulated as a stochastic game , which can also be ap plied in the threat detection and sit uation awareness [22] [23] . The Nash eq uilibrium is first achieved based on perfect en vironment kno wledge. A stocha stic lear ning automata -based channel selection algor ithm is propo sed so that the second users can approach a Nash equilibrium point. A cognitive net work consists of multiple cognitive rad io nodes, where different nod es can conduct t he spectrum sen sing and choose the bes t spectru m to tran smit the signal. T wo methods include a centrali zed ap proach where all node s communicate with the clo ud or a decentralized ap proach where o nly a few fog nodes communicate b etween the clo ud and the other nodes. These two frameworks are illustrated in Figure 2 . I t is n ecessary for each cognitive node to communicate with each other , because without i nformation fro m the o ther nodes, t he node is pro ne to have a loca l optimal decision. It will increase the system’s spectrum e fficiency if all t he spectru m sensing i nformation is accessible to the ce ntral nodes or cloud. If the central node has t he acces s to all the information, it is capable of achieving t he global optimal spectru m allocation strategies. However there also exist some li mitations such as c ommunication b urden, system security, and system’s respon se to the change o f local environmen t when all fog nodes se nd the infor mation to the central node or clo ud which has more power computatio n capability. Since each node keeps r eceiv i ng signal fro m its nearb y neighbors, a large a mount of data needs to be transmitted to th e clo ud even in the case th e transmitting signal is compressed before transmitting to cloud [24][25][26] . In some cases, some sig nals m ay of no meaning to be sent back to the central no de in ter ms of t he sp ectrum detection or allocatio n. A lso , d uring the pr ocess that the signal is being se nt back, the signal may b e jammed or compromised b y j ammers or neutral user s eit her on pu rpose or not, which will brin g interference to the pr imary us er s’ signal transmission. Figure 2 Spectrum Sensing Framework Cognitive radio methods are promising techniques that can mitigate t he spectrum scarcity by enablin g the second u sers to access the limited spectrum resources licensed b y the primary u sers. Ho wever, spectrum sensing secu rity n eeds to be considered, as the false i nformation injected into the s ystem can severel y inter ference th e spec trum se nsing p rocess an d reduce the cha nnel availabil ity to the unlicensed legitim ate users [27] . The cognitive rad io attack surface is studied from the perspective of p rimary user emulation attack s , which demonstrates that the spectrum efficiency can drop largely by emulating the incumbent sig nal. The corr esponding d efending strategy called localization -based defense is proposed to defense this sort of attack stra tegies, which ca n detect the attack by esti mating the incu mbent tran s mitter’s location and its si gnal c haracteristics. There are also other types o f attac k strategie s which take adva ntage of the syste m con figuration to inject the false in formation t hat t he traditional C hi-Square detecto r cannot detect. I t is s hown [28] [29][30][31] that the adversary can take the fully advanta ge of the syste m configuration p arameters to attack t he system incurring l arge system mean square error. Considering security, co mmunication co st, a nd d ynamic u sers; it is not fea sible for the system to provide real-tim e response by sending all the received s ignals to the clo ud. In most case s, it is infeasible to find optimal spectrum allocation strategies ; however subo ptimal solution s are fair enough to pro vide comparatively efficient s ystem spectrum usage perfor mance. As for the distrib uted cognitive radio network, each node has to coord inate to the nodes in the neighborhood in ord er to avoid the interfere nce to the p rimary users. This will incur the co mmunication b urden and system sec urity iss ue. Ma ny cases need agile re sponse i n a real -time fas hion to suppor t the system functionalit y. For example [32] , an agile and low cost atmospheric measurement system for energy harvest and real time m ission support is developed such that a candidate sensor set can be dynamicall y deployed to wind filed estimation. In ord er b alance the advantages and limitatio ns b etween ce ntralized and distr ibuted frame, in this paper, we prop osed fog co mputing framework i n which fog nodes can analyze s the most time-sensitive data, close to th e data sources and send the s elected data o r data summary for give n time to t he cloud for historical analysis a nd global opti mization of the dynamic spectrum access. 2. SYSTEM MODEL In this section, the system mo del is introd uced to sho w how the DS A fog-computin g frame w ork can optimally d etect and allocate the limited spectrum resources. Under this framework, each edge node can sense spectrum individually. Based on the rules provided by the cloud, each node has the ab ility to choo se the best spectrum selection based on the reaso ning process. Each node sen ds the data summar y to cloud periodically for wh ich th e centralized cloud server can analyze the entire system a nd determine the updated rules for each edge node. A lso, whenever there is abnormal data or a situation that the edge node cannot p rocess, the data related to this certain situat ion is also sent to th e cloud. The DAS fog s ystem structur e can be seen in Figure 3 , where each node can c o nduct its o wn spectrum sensing. As a sequence of si gnal is achieved, the fog node co nducts feat ure extraction, ano maly detectio n, and decision making . The feature extraction includ es waveform-based sen sing, energy sen sing, radio i dentification, c yclostationarit y based se nsing, and match filtering. Figure 3 Fog Computing Framework 2.1 Machine Learning Appro aches All extracted features are used to deter mine t he e xistence of spectrum a nd allocate t he spectrum for the trans mitting signals. Based on the co mputation cap ability of each fog node, different types o f machine learning methods ar e incorporated in the framework such as least square regression, support vector machine (SVM) , and manifold learning. As for lea st square logic regression , it is supp osed that vect or  denotes t he training data, of which each element d enotes the o utput fro m feature extraction pro cess. It is supposed that   󰇟              󰇠  are the N training samples.  denotes whether certain spectru m is being used or not at the current time. T he objective is to find the opti mum   that can minimize the estimation erro r          ,. The objective functio n can be characterized as follo ws,                   (1) where  is the sparsit y constraint. The cor responding Lagrangian formula ca n be rewritten as                 (2) where  is penalt y coefficient,  is independent of the  in the p rimal for mula. As    ,    󰇛    󰇜     . In this case, the node which cond ucts the least sq uare regression save s the para meter  . As for the Supp ort Vector Machine (SVM) ,                     󰇛    󰇛   󰇜   󰇜                 (3) The corr esponding dual form is                                (4) where  is a col umn vector of all ones,  is the upper bound,  is an    semidefinite p ositive matrix, of w hich          󰇛     󰇜 , the kernel          󰇛   󰇜   󰇛  󰇜 , kernel f unctions can b e linear, polynom ial, sigmoid, and Gaussia n RBF, which can also b e applied to the least square regres sion. A third appro ach is manifold learning . When a non-linear relation exist s between the features and d esired output, each data   is represented in a Euclidean space      , where  deno tes the number of signal features.   is assumed to be an i mage o f so me   in a topological space. L et sp ace  resemble   with    .  is kno wn as a manifold with reduction di mensionality  . Figure 4 illustrates t he examples of 2 -D manifolds embedded in a 3-D E uclidean space. The objective of manifold lear ning is to directly u ncover t he o ne - to -one map from   to   . T he mapping function should preserve th e geo metric structure of space   . Geo metrically, this can be interpreted as uncurl ing a curved surface into a super-plane. P opular ap proaches for manifold lear ning i nclu de locally linear embedding [33] , I soMap [34] , Laplacia n eigenmap [35] , and Maximum variance unfolding [36] . Figure 4 2-D Manifolds Embedded in A 3- D Sp ace (Linear Subspace, S-curve, and Swiss roll) Figure 5 sho ws the dimensio nality reductio n of the 3 -D S wiss roll based on manifold learning. Figure 5 d isplays the results o f several manifold functions, e.g., IsoMap, LLE, Hessia n LL E, and provides the comparison results fro m li near approaches, su ch as M ultidimensio nal scaling (MD S) and Principle component a nalysis ( PCA). Since the 2 D embed ded manifold in Swill Roll is nonlinear, the linear approac hes always fail to “uncurl” it. Also, in Figure 5 , w e can see that fo r manifold learning, IsoMa p and Hessian LLE can s uccessfully “uncurl” the or iginal data, but LLE cannot. Hence, based on the co mputational capabili ty of each ed ge node, manifold learning can be applied to feature extraction for dynamic spectrum access. Figure 5 Dimensionality Reduction of 3-D Swiss Roll Based on Nonlinear Manifold Learning 3. SPECTRUM SENSING FOR FOG-BASED COGNITIVE RADIO Spectrum se nsing is a collect ive name of multiple tec hniques that are applied to cognitive radio to detect the unused frequency band and assign it to one or multiple secondar y users. In this sect ion, we briefly i ntroduce and su mmarize some widely used signal processing tec hniques that have bee n rep orted in the literature and are suitable for o ur proposed fog framework, T he methods deter mine whether the chann el is occupied by the p rimary user and whether an y unused spectrum is availab le for a seco ndary user. In t he pro posed framework, each fog nod e serves as a machine lear ning engine, and uses variou s outp uts of a given spectrum sensing method based on the computational cap ability of ea ch node. 3.1 Energy Detect or Energy detector -based sensin g is the most widely used spectr um sensing algorit hm due to its low co mputational complexity. Another advanta ge of energy as sessment i s that the energy detector is developed based on the assumptio n that r eceivers need no pr ior information o f the p rimary user signals, which makes this sensing tech nique much more practical as compared to other existing app roaches. Ho wever, in some cases, t he second users ha ve some infor mation about the target to be d etected. A po pular Ba yesian detector can be also utilized for detection [37] . The received signal is as sumed as  󰇛  󰇜   󰇛  󰇜   󰇛 󰇜 (5) where  󰇛󰇜 is the received signal,  is the channel coefficie nt,  󰇛  󰇜 is the primary user ’s signal to be detected ,  󰇛 󰇜 is the add itive white Gaussian noise (AWGN), and  is the sample index. Note that when  󰇛  󰇜   , there is no transmission from the pri mary user over this frequency band. The total energy observed over  samples can be written as      󰇛  󰇜    (6) Using (6 ), the spectrum occupanc y d ecision ca n be obtained by co mparing a threshold  , which i s equi valent to the following hypothesis.    󰇛  󰇜   󰇛 󰇜 (7)    󰇛  󰇜   󰇛  󰇜   󰇛󰇜 Equation (7 ) deter mines whether the primary user ’s signal occupies the spectrum. Note t hat  is stored in “rule” block for fog proce ssing, and  can also be modified by the clo ud in a real time fashion ins tead of a fixed value of a threshold used in the traditional energ y detector sensing method. 3.2 Wavefor m-Based Sensing In the ca se that the signal patterns are known to the receive rs, the w aveform - based sensing technique can be applied to the cognitive r adio system. W e use the sa me transmission m odel (5) , from whic h the wavef orm-based sensing metric is written as       󰇛  󰇜   󰇛󰇜     (8) where  denotes the co njugate oper ation. When primary user signal p resents on the spectrum, (8) can be further written as       󰇛  󰇜      󰇛  󰇜     󰇛󰇜   (9) When primary user signal i s absent, the sensi ng metric is written as        󰇛  󰇜     󰇛󰇜  ( 10 ) Similarly, a threshold val ue needs to be set i n t he s ystem to make spectrum occupa ncy dec ision, and this value is modified by the cloud. 3.3 Cyclostatio narity-Based Sensing Cyclostationarity-ba sed sensing is a techniq ue t hat e xploit the signal c yclostationarity feat ures to d etect the primar y user transmissions. In t his method, the cyclic correlation functio n is used to detect the signals in the frequency ba nd. The cyclic spectral densit y function of a received signal is written as  󰇛    󰇜      󰇛  󰇜     ( 11 ) where  is the c yclic frequency, and    is the cyclic autoco rrelation function, which ca n be written as    󰇛  󰇜   󰇟 󰇛    󰇜   󰇛    󰇜   󰇠 ( 12 ) When the cyclic freque ncy  equals t he frequency of transmitt ed signal  󰇛󰇜 , the output of ( 11 ) reaches the peak value. 3.4 Feature Det ection Sensing Since primar y user si gnal usually has its unique transmi ssion pattern suc h as modulatio n sc heme, ca rrier frequenc y, bandwidth, etc.; feature detectio n and machi ne learni ng alg orithms can also be applied for spectrum sensing. This category of se nsing technique s u sually uses data-driven model rather than p hysical mo del to d ifferentiate the pr imary user signal from the secondary user signal and noise. I n t his work, all types of features extracted b y feature d etection sensing are sent to the machine learning engine t o make the decision on spectr um access. 4. CONCLUSION AND FUTURE WORK In this paper, we propose a manifold lear ning dynamic s pectrum allocatio n framework co mbining fog co mputing a nd cloud co mputing so that t he r eceived signal is proce ssed clo se to where it is generated. Machine learnin g app roaches are also incorporated to determine the available spectru m. Based on the features of t he signal received, different machine learning met hods are as applied such as least square lo gic re gression, support vector machine, and manifold leaning. T he feature extracted ca n also be projected to a higher feature space so that categorizatio n p erfor mance can b e furt her improved. Based on the rules defin ed, each fog nod e has th e ability to reason and choose the best sp ectrum ca ndi date to transmit the signal without interfering with the licensed legitimate primary users. T he signal could be processed by th e local nodes immediatel y, while the local node can also send the signal su mmary to the clo ud so that t he decision r ules at the e dge nod es r ely on can be updated d yn amicall y. Conside ring t he co mputational capability of t he clo ud, the ab normal data f rom t he ed ge nodes is also sent to the cloud for further analysis which enhances system security. T he future work will include the hard ware imple mentation regarding fog nodes of different computatio nal capabilities and machine learning appro ach selection considering t he feature of transm itted signal to optimize the d ynamic spectrum allocation . REFERENC ES [1] Tian, X., T ian Z. , B lasch, E. , Pham , K. , Shen, D. and Chen, G. , “ Slidi ng windo w energ y detection for spectrum sensing under lo w SNR conditions ” , Wirel. Co mmun. Mob. Comput, Vol. 16, Issue 12, pp. 1654 – 1663, Aug. 2016 . [2] Li L. , Ding Y. , Zha ng, J. K. an d Zhang, R. , "Blind Detection with U nique Identific ation in T wo-Wa y Relay Channel", in IEE E Transactions on Wireless Co mmunications, vol. 11, no. 7, pp. 2640 -2648, July 2012. [3] C hen, Y., Blasch, E., Chen, N. , Deng, A. , Ling, H. , Chen, G. , “Real - ti me W AMI streaming target tracki ng in fog,” in Proc. SP IE, Vol. 9838, 2016. [4] Wu, R ., Liu, B ., Chen, Y ., B lasch, E ., Ling, H ., Che n, G ., “A Container -based Elastic Cloud Architecture for Pseudo Real-time Exploitation of Wide Ar ea Motion Imagery (WAMI ) Stream ” , T he Journal of Signal P rocessing Systems, pp 1-13, Nov. 2016. DOI 10.1007/s11 265 - 016 -1206-6 [5] L iu, B. , Blasch, E. , C hen, Y. , Av ed, A. J., Hadiks, A. , Shen, D., Chen, G. , “Information Fusion in a Clo ud Computing E ra: A S ystems-Level Per spective ” , IEEE Aerospace and Electronic Systems Ma gazine, Vol. 29, No. 10, pp. 16-24 , Oct. 2014. [6] L iu, B. , Chen, Y. , Blasch, E. , Pham, K. , Shen, D. , Chen, G. , “A Holistic Cloud -Enabled Robotics Syste m for Real- Time Video T racking Application ” , Fut ure In formation T echnology, Lecture No tes in Electrical Engineering, Vol. 276, pp 455 -468, 2014. [7] Mitola, J., Maguire, G. Q., "Cognitive rad io: making software radio s more personal," in IEEE Personal Communication s , vol. 6, no. 4, pp . 13-18, Aug ( 1999 ). [8] Niy ato, D. , Ho ssain, E., "Competitive spectrum sharing in cognitive r adio networks: a dynamic game appr oach," in IEEE Transaction s on Wireless Communication s , vol. 7, no. 7, pp . 2651-2660, July ( 2008 ). [9] Zh ao , Q. , and Sadler, B. M., "A Survey of Dynamic Spectrum Access," in IEEE Signal Processing Magazine, vol. 24, no. 3, pp. 79 -89, May 2007. [10] Goldsmith, A., J afar, S. A., Maric, I., Srinivasa, S., “ Breaking Spectrum Gr idlock Wi th Cogniti ve Radios: An Information Theoretic P erspective," in Proceedings o f the IEEE , vol. 97, no. 5, pp. 894 -914, May ( 2009 ). [11] Haykin, S., “ Cognitive radio: brain-e mpowered wireless communications," in IEEE Journal on Selec ted Area s in Communication s , vol. 23, no. 2, p p. 201-220 , Feb. (2005 ). [12] Xiao, S., Metry, M., Tr abelsi, M. , Balog , R. S. and Abu-Rub, H. , "A Model P redictive Co ntrol techn ique for u tility- scale grid connected b attery systems usin g packed U cells multilevel inverter," I ECON 2016 - 42nd Annual Conference of the IEEE Indus trial Electronics Society, Flo rence, 2016, pp . 5953- 5958. [13] Zadeh, L. A. , “ Fuzzy logic, neural networks, and so ft computing ” . Co mmun. ACM 37, 3 (March 19 94), 77 - 84. [14] Lu, J ., Niu, R. " Sparse attack ing strate gies in multi -sensor dynamic s ystems ma ximizing state esti mation er rors," IEEE International Conference on Acoustics, Speech and Signal Processing (ICASS P), pp. 3151-3155 , (2016 ). [15] Lu, J., Niu, R ., Han, P. "Op timal space -time attacks o n syste m state esti mation under a sparsity co nstraint", in Pro c. SPIE Sensors and Syste ms for Space Applications I X, 2016 [16] Park, S. , Lee, S. J ., Weiss , E. , and Mo tai, Y. " Intra- and Inter -Fractional Variatio n Pred iction of Lung T umors Usin g Fuzzy Deep Learning," in IEEE Journal of T ranslational Engineering in He alth and Medicine, vol. 4, pp. 1-12, 2016. [17] Zuo, H ., Fan, H ., Blasch, E ., and Ling, H ., “Combining Convolutional and Recurrent Neural Networks for Human Skin Detection, “I EEE Si gnal Proce ssing Letters, Jan 2017. [18] Jung , E. and L iu, X. , " Opportunistic Spectrum Access in Heterogeneous User Environments," 2008 3rd IE EE Symposium on Ne w Frontiers in Dynamic Spectr um Access Net works, Chicago, IL, 2008 , pp. 1 - 11. [19] Luo, C. , H uang, Y. F. , G upta, V. , "Stochastic Dynamic Pricing for EV Charging Stations with Renewables Integration and Energ y Storage," in IEEE Transactions o n Smart Grid , vol.PP , no.99, pp.1-1 [20] Luo, C. , Huang , Y. F. and Gupta, V. , " A Consumer Behavior Based Appro ach to Multi-Stage E V Char ging Station Placement," 20 15 IEEE 81st Vehicular Technolog y Conference (VT C Spring), Glasgow, 2015 , pp. 1 - 6. [21] Xu, Y. , W ang, J. , Wu, Q. , Anpalagan , A. , and Yao , Y. D., "Opportunistic Spectrum Access in Unknown Dynamic Environment: A Game-Theo retic Stochastic Learning Solution," in IEEE Transactions o n Wireles s Communications, vol. 1 1, no. 4, pp. 1380 -1391, April 2 012. [22] Chen, G. , Shen, D., K wan, C., Cru z, J., et al., “Gam e T heoretic Approach to T hreat Pred iction and Situ ation Awareness,” Jo urnal of Advances in Infor mation Fusion, Vol. 2 , No. 1, 1 -14, J une (2007). [23] Lu, J., Niu, R. " A state esti mation and malicious attack game in multi -sensor dynamic systems," Interna tional Conference on Information Fusion (Fusion), pp. 9 32-936, ( 2015) [24] Nicolae, B., “High T hroughput Data - Co mpression for C loud Storage”, Glob e 10 in Pro ceedings of the 3rd International Conference on Data Management in Gr id and P2P Systems, Sep 2010, B ilbao, Spain. 6265, pp.1 -12, 2010, LNCS [25] Li, L. , Wan g, G. , Chen, G. , Chen, H. M. , Blasch, E. and Ph am, K. , "Robust airbo rne i mage transmission u sing j oint source-channel coding with UEP," 2016 IEEE Aerospace Conference, Bi g Sky, MT, 2016, pp . 1 - 7. [26] Welch, T. A., “ A Technique for High-Perfor mance D ata Compressio n. Computer”, June 198 4, 8- 19. [27] Chen, R. , Park, J. M. and Reed, J. H., "Defense ag ainst Primary User Emulation Attacks in Cogniti ve Radio Networks," in IEEE J ournal on Selected Areas in Communications, vol. 26 , no. 1, pp. 25 -37, Jan. 2008. [28] Kim, J. , T ong, L. and Tho mas, R. J., " Data Fra ming Attack on State Estimation," in I EEE Jo urnal on Selected Areas in Communications, vol. 32 , no. 7, pp. 1460 -1470, July 2014. [29] Lu , J. , Niu, R. , "False information injection a ttack on d ynamic state esti mation in m ulti -sensor systems," 17th International Confere nce on Information Fusion (FU SION), Salamanca, 201 4, pp. 1 - 8. [30] Lee, K. , Emami, T. , Ibdah, Y. , Bi , Y. and Li, L. , "Cooperative Distributed MISO Wireless Rela y Networks Under Jamming Environments w ith Po wer Constraints," 20 11 T hird Internation al Confer ence on Co mputational Intelligence, Co mmunication Systems and Net works, Bali, 201 1, pp. 84 - 88. [31] Li, L., Wang, G., Tian, X. Shen, D., Pham, K., Blasch, E., Chen, G., “SINR estimatio n for SATCOM in the environment with ja mming signals”, in Proc. SPIE 9838, Sensors and S ystems for Space Applications I X [32] Xiang, X. , Wang, Z. , Mo , Z. , Chen, G. , Pha m, K. , and Blasch , E. , "W ind field estimation through autono mous quadcopter avionics," 2016 IEEE/A IAA 35 th Digital Avionics S ystems C onference (DASC), Sacra mento, CA, 2016, pp. 1 - 6. [33] Roweis , S. T. and Saul, L. K., " Nonlinear d imensionality reduction b y lo cally linear embedd ing," Science, vol. 290, pp. 2323 -2326, 2000. [34] Tenenbaum, J. B., De Silva, V. , and Langford, J. C. "A global geometric fra mework fo r n onlinear dimensionalit y reduction," science, vol. 2 90, pp. 2319 -2323, 20 00. [35] Donoho , D. L., and Gri mes, C. , " Hessian eigen maps: Locall y linear e mbedding tech niques for high -dime nsional data," Proceed ings of the National Acade my of Sciences, vol. 100 , pp. 5591 -5596, 2 003. [36] Weinberger, K. Q. and Saul , L. K. , “ An introduction to nonli near dimensionalit y reduction b y maximum variance unfolding ” . In proceed ings of the 21st natio nal conference on Artificial intelligence , 1683 - 1686. [37] Niu , R. , Lu, J. , "False infor mation detection with mi nimum mean squared er rors for Bayesian est imation," 2015 49th Annual Conference on I nformation Scie nces and Systems (C ISS), Baltimore, MD, 2 015, pp. 1 - 6.