Internet Multimedia Traffic Classification from Qos Perspective Using Semi-Supervised Dictionary Learning Models

NETWORKS & SECURITY

Internet Multimedia Traffic Classification from QoS Perspective Using Semi-supervised Dictionary Learning Models

Zaijian Wang1,*, Yuning Dong2, Shiwen Mao3, Xinheng Wang

1 College of Physics and Electronic Information, Anhui Normal University, Wuhu, 241000 China 2 College of Communications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing, 210003 China 3 Department of Electrical and Computer Engineering, Auburn University, Auburn, AL 36849-5201 USA 4 School of Computing, University of the West of Scotland, Paisley, PA1 2BE, UK * The corresponding author, email: [email protected]

Aabstract: To address the issue of fine- sifier. Our experimental results demonstrate grained classification of Internet multimedia the feasibility of the proposed classification traffic from a Quality of Service (QoS) per- method. spective with a suitable granularity, this paper Keywords: dictionary learning; traffic clas- defines a new set of QoS classes and presents sication; multimedia traffic; K-singular value a modified K-Singular Value Decomposition decomposition; quality of service (K-SVD) method for multimedia identification. After analyzing several instances of I. INTRODUCTION typical Internet multimedia traffic captured in a campus network, this paper defines a new With the development of various Internet mul- set of QoS classes according to the difference timedia applications, cross-layer optimization in downstream/upstream rates and proposes a of network resource allocation for enhanced modified K-SVD method that can automatical- user experiences has attracted considerable ly search for underlying structural patterns in interests in the research community [1-8]. the QoS characteristic space. We define bag- On the other hand, Internet Service Pro- QoS-words as the set of specific QoS local viders (ISPs) need to consider different Qual- patterns, which can be expressed by core QoS ity-of-Service (QoS) requirements for data, characteristics. After the dictionary is con- voice and video applications because users structed with an excess quantity of bag-QoS- need better, more personalized services. How- words, Locality Constrained Feature Coding ever, due to the heterogeneity of the definition (LCFC) features of QoS classes are extracted. of QoS/service classes with different levels By associating a set of characteristics with a of granularity for various types of multimedia percentage of error, an objective function is traffic, ISPs still have difficulties guaranteeing formulated. In accordance with the modified the end-to-end QoS of multimedia services by K-SVD, Internet multimedia traffic can be allocating proper network resources, especially Received: Feb. 24,2017 classified into a corresponding QoS class with when they target high-bandwidth applications Revised: Jun. 8, 2017 a linear Support Vector Machines (SVM) clas- such as video streaming [9-12]. Granularity, Editor: Youping Zhao

202 China Communications • September 2017 in particular, should not be used when classi- cepts in various major benchmarks [17], clas- fying, existing types of the multimedia traffic sification performance should be improved by The authors intend to with new types of multimedia traffic. using dictionary learning models in the QoS address the problem For example, video traffic can be further domain. of effective classifying classified as Peer-to-Peer (P2P)-based or In this paper, we define seven QoS catego- multimedia traffic with a suitable gran- non-P2P-based, each with different dynamic ries with newly discovered features and pro- ularity, and presented characteristics. To efficiently utilize limited pose a characteristic-based method which will a modified K-SVD network resources, it is important to classify classify Internet multimedia traffic by utilizing Internet multimedia Internet multimedia traffic using a finer granu- a modified K-Singular Value Decomposition traffic identification larity, so that Internet multimedia services can (K-SVD) method. The K-SVD will learn the framework based on be improved with better end-to-end QoS guar- dictionary so that it can classify various Inter- the concept of QFAg in this paper. antee by ISPs. In order to effectively provide net multimedia traffic with appropriate differ- end-to-end QoS in heterogeneous networks, entiating granularity of multimedia services by authors in [12] emplyed an Application Ser- considering QoS requirements. By analyzing vice Map (ASM) to classify traffic, while traf- multimedia flows in a large-scale, real world fic was divided into eight Unified Communi- network, we find that the new QoS-related cation (UC) classes in [13]. In addition, online features correlate with the bidirectional rates allocation of communication and computation of different streams of Internet multimedia resources for different real-time multimedia traffic. After investigating the sparse nature of services was considered in [14] to provide an the QoS of multimedia streams, we extract the optimal service to users. training dictionary from training samples, uti- In order to offer uninterrupted services lize a modified K-SVD to learn the dictionary, when using a mobile Internet, a Flow Aggre- and propose a characteristic-based method to gation (FAg) concept was proposed in our pre- classify multimedia traffic. Specifically, our vious work [15], which involved aggregating main contributions in this paper are summa- traffic flows with similar QoS requirements rized as follows: into a FAg. We show that the FAg may act as a 1) by extensive statistical analysis of traffic bridge to help different network/QoS domains flow data collected from a real world net- along the transmission path understand/trans- work, we find that downstream/upstream late QoS requirements with a unified view. rates are suitable for differentiating Internet However, how to utilize the FAg in an effec- multimedia traffic (e.g., video and game tive way is still an unresolved issue. In this traffic) from a QoS perspective; paper we try to tackle this problem by using 2) based on the sparse property of the QoS an appropriate, differentiating granularity of characteristics, we show that the current multimedia services from a QoS perspective types of Internet multimedia traffic (e.g., to achieve the FAg objective. video and game traffic) can be classified To achieve this goal, it is necessary to clas- into seven QoS based FAg (QFAg) catego- sify Internet multimedia traffic into proper ries; granularity categories based on QoS related 3) we propose a modified, K-SVD-based characteristics. Most existing traffic classifica- Internet multimedia traffic classification tion methods focus on a specific application/ framework based on the concept of QFAg, protocol but not on QoS related characteristics which improves the classification perfor- [16]. However, traffic belonging to a single mance significantly by employing a unified QoS class may be from different application objective function incorporated with QoS types and generated by different protocols. features, for which the K-SVD algorithm Considering that the Bag-of-Words (BoW) can be utilized to efficiently solve the clas- method has been proven to be the most effi- sification problem. cient classification scheme for individual con- The remainder of this paper is organized as

China Communications • September 2017 203 follows. Section II presents related work in the Inter-Arrival Time (IAT), delay, jitter, and literature. Section III introduces the setup of packet loss [18]. These characteristics are not our data collection from a real world network suitable for effectively creating the QoS class- and presents the dataset. Section IV analyzes es because they cannot reflect the differences the related typical QoS characteristics, selects within the QoS classes. For example, packet

1 http://www.cntv.cn/. new QoS characteristics for differentiating size, frequently described in the literature, is 2 http://www.bittorrent. Internet multimedia traffic streams, and pres- an important characteristic, but cannot be used com/. ents the definition of the QoS class for Internet to make a distinction between Closed Circuit 3 http://skype.gmw.cn/. multimedia traffic. Section V develops a QoS Television (CCTV1) web video and BitTorrent2 4 http://iptv.cntv.cn/. class recognition framework based on modi- video. Similarly, IAT also needs to be further fied K-SVD for Internet multimedia traffic and processed before it can be applied to differen- Section VI analyzes its complexity. Experi- tiate traffic types [19]. mental results are given in Section VII, and To contrast Skype3 with IPTV4, authors Section VIII concludes the paper. in [20]. studied the corresponding statistical characteristics of IAT and packet size by II. RELATED WORK using an interactive mode and packet prop- erties. Researchers in [21] showed that IAT The International Telecommunication Union and packet size may affect jitter, packet loss, (ITU), the Internet Engineering Task Force and bandwidth. IAT, packet loss and jitter for (IETF), and some other standard organi- Skype in UMTS were analyzed in [22]. The zations, as well as network providers have results indicated that QoS characteristics have stipulated a large number of typical QoS char- different distributions in different flow direc- acteristics of Internet multimedia traffic [11]. tions. Researchers in [23] indicated that, in For example, ITU-T Recommendation G.1010 online gaming, there were obvious differenc- primarily adopts error tolerance, delay and jit- es between downstream and upstream video ter, and packet loss as QoS criteria. IETF uses traffic. However, related works only provided bandwidth, delay and jitter, and packet loss a simple comparison as opposed to QoS clas- to classify services into IntServ and DiffServ sification. Researchers in [24] proposed a flow models. The Third Generation Partnership aggregation method, which can effectively re- Project (3GPP) defines four QoS classes based duce the number of flows. Further research in on delay. IEEE 802.11e selects a maximum [25] indicated that statistical characteristics of contention window, a minimum contention flow aggregation can be used to identify relat- window, and an arbitration inter-frame space ed traffic streams. Researchers in [26] tried to to classify the traffic classes. IEEE 802.16m aggregate flows at the host/port level, but no adopts maximum sustained rate, maximum implementation details were given. Reference latency tolerance, jitter tolerance, etc. The [27] showed that flow could be aggregated classifications using these models are not ideal based on some common characteristics to re- for existing networks. In particular, the online duce the network operating costs. game traffic with various genres currently Different from our works, the authors in occupies the majority of network resources, [36] considered different classes of video especially in Asian countries. However, little flows, but focused on scheduling execution research has been done using current QoS time in Wireless Multimedia Sensor Net- models for online games. work(WMSN). The authors in [37] focused on Typical classification methods utilize flow control to balance delivery fairness and different features to construct classifiers for efficiency over Heterogeneous Wireless Net- differentiating the specific types of traffic or works (HWNs). The authors in [38] presented protocols. These features include protocol, an online unsupervised learning classification port, payload, packet/flow size, flow duration, of pedestrians and vehicles for video surveil-

204 China Communications • September 2017 lance, but not Internet multimedia. They ex- multimedia traffic flows with a natural granu- 5 http://www.icq.com; tracted the moving objects with their features larity. http://www.QQ.com; from the original video, but not QoS charac- The datasets are collected from 25 kinds of http://cn.msn.com/; http://dota.uuu9.com/; teristics. The authors in [39] detailed a novel multimedia traffic flows, each being 30 min- http://www.dota2.com.cn; approach to solve the assignment problem of utes long, on a machine with an AMD 2.10 http://lol.qq.com/; http:// finding optimum thresholds for axle-based ve- GHz AthlonTM X2 DualCore QL-64 proces- xyq.163.com/; http:// hicle classifiers. The authors in [40] proposed sor during day/night time in Spring 2014. We nz.qq.com/; a bilevel feature extraction-based text mining ran these traffic streams at the hosts several http://donkey4u.com/; that integrates features extracted at both syntax times and collected the traffic flows. When http://www.emule.org. cn/; http://www.uusee. and semantic levels with the aim to improve one of these is running, other applications are com/; PPStream is a net- the fault classification performance. stopped. Each application flow is identified work television software; Traffic identification is an automatic clas- by a five-tuple: (source/destination IP address, http://dl.xunlei.com/; sification process, and many methods have source/destination port, and protocol). If their http://www.sopcast.cn; achieved good results as for example, K-SVD, duration is longer than 0.1 second, they are http://tvants.en.softonic. Hidden Markov Model (HMM) [28], Support marked as sub-flows (Usually one traffic flow com/; PPLive is a live p2p TV application; http:// Vector Machine (SVM) [29], and Bayesian has many sub-flows). Control packets were ppmate.com/. network based classification methods [30]. dropped from the raw data, leaving only the K-SVD, especially, has advantages over other raw data to be analyzed. The types of multi- algorithms in convergence rate and adaptabil- media traffic flows are summarized in Table ity to complex data structure [31]. Based on 1.5 the existing research work on K-SVD [31], Table I Components of the dataset this paper proposes a modified K-SVD algo- Category Typical Multimedia Traffic Data volume (GB) rithm for Internet multimedia traffic QoS class ICQ 16.128 recognition. The proposed algorithm utilizes Real-time interactive QQ 16.128 sparse coding to learn the dictionary, reduces video MSN 16.128 the dimension of characteristics, and improve Dota 0.692 the accuracy of recognition. Dota2 3.203 Online Lol 1.592 ATASET REPARATION gaming III. D P Dreams Western adventure 1.285 Since there are no suitable datasets available in Against War 1.285 which the multimedia traffic is classified into eDonkey 104.48 different categories according to QoS related Emule 104.36 characteristics, we capture the typical Internet UUSee 31.2 multimedia traffic from a campus network PPStream 50.49 to build a basic dataset with Wireshark [32]. Tudou video (standard-definition) 8.1327 After processing the raw data, the following Youku video (standard-definition) 8.5114 statistical characteristics were obtained: Tudou video (high definition) 13.3597 1) average upstream rate (Mbps): the amount Real-time Youku video (high definition) 13.5831 of data per second sent to a server by cli- streaming BBC Web video 14.2819 media ents; and CCTV Web video 14.3264 2) average downstream rate (Mbps): the Xunlei 104.76 amount of data per second sent to clients BitTorrent 104.81 from the server. SOPCast 31.2 We analyze the statistical characteristics of TVAnt 31.2 different types of Internet multimedia traffic. Skype 3.1914 With the appropriate combination of QoS PPlive 52.02 characteristics, we aim to differentiate Internet PPMate 50.49

China Communications • September 2017 205 In order to construct a clear display, we IV. SELECTION OF TYPICAL QoS select data of 14 types of video and gaming RELATED CHARACTERISTICS AND traffic, and compute the corresponding nor- CLASSIFICATION OF QoS CATEGORIES malized byte volume, the normalized average IAT, and the normalized average packet size. Downstream/upstream rates embody differ- The results are shown in Fig. 1. By analyzing ences of QoS resource requirements of traffi c the QoS characteristics of the dataset, we fi nd in time, space and direction, for the following that: reasons. 1) different types of traffi c streams are notably 1) Validity: rate represents the number of different in data volume, e.g., Youku video bytes per second, and indicates the rela- Standard-Defi nition video (SD) has a small- tionship between packet/fl ow size and time. er byte volume than Youku High-Defi nition Considering that the majority of traffic (HD) video; fl ows have direct requests to the bandwidth 2) online games have a smaller byte volume resources, it seems to be more effective to and average packet size; classify traffic with downstream/upstream 3) online games generally have larger average rates. IAT since there are a lot of human computer 2) Generality: downstream/upstream rates em- interactions and local operations in online body the resource requirements of traffi c in games. different directions. Considering that new It is necessary to further study QoS charac- QoS classes should effectively cover and teristics of multimedia traffic for obtaining a differentiate the existing QoS classes, we more suitable QoS class. choose downstream/upstream rates to be In addition, a lot of QoS related characteris- consistent with most international standards tics show locality to corresponding transform to some extent. domain based on the above analysis results 3) Availability: since a typical QoS model pro- and [41]. According to [42],[43], locality can vides operations of two directions for band- bring sparsity, which inspires that we can uti- width allocation, downstream/upstream lizes locality to defi ne new QoS categories and rates can be obtained easily. For example, sparsity to classify QoS categories for typical broadcast operates the downstream and the multimedia traffi cs. Base Station (BS) controls the upstream by

1 Dreams Western adv. 0.9 dota2 0.8 QQ dota 0.7 BBC Web video 0.6 CCTV Web video SOPCast 0.5 Youku video (SD) 0.4 Youku video (HD) Skype 0.3 PPStream 0.2 lol 0.1 PPlive Xunlei 0 normalized byte volume normalized average IAT normalized average packet size

Fig. 1 The columnar distribution of 14 types of traffi c related characteristics

206 China Communications • September 2017 allocating resources to the traffic streams specifi c normalization equation is as follows:

according to their QoS requirements in Wi- Riu, riu, =,1 ≤≤ iN , (1) MAX. Rmax 4) Robustness: the dynamic network status R =id, ≤≤ has an obvious effect on QoS character- rid, ,1 iN , (2) Rmax istics. Considering that the time intervals where R denotes the upstream rate of traffi c between the upstream and downstream data iu, flow i, r is the corresponding normalized transmissions are very short, downstream/ iu, rate; R denotes the downstream rate of traf- upstream rates seems to be affected by id, fi c fl ow i, r is the corresponding normalized the same set of network factors. The ratio id, rate; N represents the number of traffi c fl ows; between the downstream/upstream rates and R denotes the maximum value between seems to be insusceptible to the interference max R and R , which is defi ned as caused by changes in network status. mu, md, Rmax = max{}m{ RR m , u , m , d } , m= 1,2, … , N . (3) Therefore, downstream/upstream rates are In order to understand all the captured traf- selected to classify multimedia traffic in this fi c streams, we plot the normalized logarithmic paper since they can be an effective classifi ca- value of downstream/upstream rates in Fig. 2. tion feature. The x-axis is the normalized downstream rate Since the network status is influenced by and the y-axis is the normalized upstream rate. many known (such as routing and user’s active For easy reading, Fig. 2 plots one out of every degree) and unknown factors that affect the ten points. From Fig. 2, we can compare the related QoS characteristic, traffic acquisition distributions of normalized logarithmic value should be carried out in a longer time span. of downstream and upstream rates for all cap- The inherent relationship is quite complex and tured traffi c. The entire distribution plane can will be left to our future work. be clearly divided into seven different zones, To obtain good statistical characteristics, as shown in Fig. 2. our experiments was continued for a period From the distribution of downstream/up- of four months to collect the traffi c data. The stream rates shown in Fig. 2, the traffi c fl ows maximum value of downstream/upstream rates can be broadly classified into the following is chosen to normalize all traffic rates. After seven distinct categories from the QoS level normalization, the relative value is used to dis- perspective: broadcast standard-defi nition vid- tinguish between different types of traffi c. The

−2

h −3

−4

−5

Zone 1 −6 Zone 2 Zone 3 −7 Zone 4

Normalized Upstream Bandwidt Zone 5 −8 Zone 6 Zone 7 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 Normalized Downstream Bandwidth

Fig. 2 The distributions of normalized logarithmic value of downstream/upstream rates (Mbps) for seven QoS categories

China Communications • September 2017 207 eo (BSDV), broadcast high-definition video mode. (BHDV), web video (WV), trade style video WV is designated for web video traffic (TSV), barter style video (BSV), interactive using the HTTP protocol, in which the HTTP video (IV), and game class (GC). The cor- server delivers the web objects embedded in responding relationships are shown in Table the web page. For this class, the downstream 2. To clearly describe the relationships, we rate is larger than the upstream rate since the divided the normalized value of downstream/ video traffic embedded in HTTP has very few upstream rates into five grades (extra small, instructions to be transmitted in upstream small, medium, large, extra large), which are direction. WV has similar downstream and up- represented by numbers 1, 2, 3, 4, and 5, re- stream rates for different traffic flows since it spectively. A larger number represents a higher maintains a stable rate and period of transmis- grade with more stringent QoS requirements. sion. BSDV represents live standard-definition TSV is ideal for traffic with the hybrid P2P video multimedia traffic such as multimedia model, which often perform near-optimally teleconferencing and video phone traffic. This in terms of uplink bandwidth utilization, and class can achieve grade 1 for the upstream rate download time except under certain extreme and grade 2 in downstream rate. The distribu- conditions. The traffic often generates a sig- tion of traffic characteristics is concentrated, nificant amount of upload traffic and presents and the amount of data generated is not siz- high packet length values. The Hybrid P2P able. model is currently one of the most common BHDV represents live high-definition video models for streaming video on the Internet, multimedia traffic of high fidelity with tradi- by which traffic is usually started by enabling tional live traffic characteristics such as high multiple concurrent TCP connections, and the quality video conferencing. Different from data is transmitted bidirectionally from multi- BSDV, this class has a larger downstream rate ple P2P nodes at the same time. Multiple tasks than BSDV does since high definition video performed simultaneously generally have a has a larger volume data than standard defi- longer concurrent transmission time. Peers nition video, and can be classified as the 3rd not only download content from the server but grade for the downstream rate. BHDV and also serve it to other peers. Peers often can BSDV have little difference in the upstream download more than they upload to the net- rate since they adopt the same transmission work when high bandwidth peers are present.

Table II Proposed new QoS classes Zone QoS Class Multimedia Traffic Downstream rate Upstream rate Description VoD, video phone, online Condensed distribution of downstream/upstream rates; 1 BSDV Small Extra small video large ratio of downstream/upstream rates 2 BHDV High quality Video, VoD Middle Extra small Similar with BSDV in condensed distribution and ratio BBC Web video, CCTV Large ratio downstream/upstream rates; largest span in 3 WV Extra large Small Web video the distribution of downstream rate Xunlei, BitTorrent, Emule, Large span in the distribution of downstream rate; mid- 4 TSV Large Middle Fileguri dle ratio downstream/upstream rates Sopcast, TVAnt, PPlive, Large span in the distribution of upstream rate; small 5 BSV Skype, PPMate, PPStream, Large Extra large ratio of downstream to upstream rate; more dispersed SinaLive, Coolstreaming distribution of downstream/upstream rates Condensed distribution of downstream/upstream rates; 6 IV ICQ, QQ, MSN Middle Middle the downstream rate is similar with upstream rate Multiplayer Condensed distribution of downstream/upstream rates; 7 GC Interactive Extra small Extra small middle ratio of downstream to upstream rate Gaming

208 China Communications • September 2017 This class is not sensitive to delay and band- client. width, and the upstream rate can be classified In typical QoS models, video traffic flows as grade 2 and the downstream rate as grade 4. are usually placed into one category without BSV is designated to traffic with typical considering that video traffic flows with differ- P2P live traffic characteristics. In the traffic ent transmission modes (P2P, HTTP, C/S, and streams, there are more nodes available to ob- Broadcast) have obvious difference in QoS tain data from other nodes and provide data to requirements. Moreover, video traffic with the other nodes. Since this class can preserve user same transmission mode and content also have satisfaction by using an adaptive multi-rate to different QoS requirements due to different guarantee QoS requirements, this class has a display standards. Utilizing a distinct QoS larger span than TSV in the distribution of the classification scheme, this paper evaluates the upstream rate, making the downstream rate above traffic characteristics in a new QoS clas- far smaller than the upstream rate. The distri- sification framework. We are not simply creat- bution of downstream and upstream rates is ing QoS classes according to traffic types, but more dispersed since longer traffic delay and are refining existing QoS classification based jitter can be tolerated. BSV can be classified on the actual distribution of characteristics in as having a grade 5 in upstream rate and grade downstream and upstream rates of multimedia 4 downstream rate. traffic. The newly proposed QoS class is more IV is intended for live multimedia traffic convenient for network resource management. that has a strict interactive response time with For example, an Internet Service Provider the highest class priority such as ICQ, QQ and (ISP) can allocate corresponding bandwidth MSN. IV has a very condensed distribution resources for one kind of data flow according of downstream and upstream rates, and the to their common QoS requirements. downstream rate is similar to the upstream rate. The traffic belonging to this class has V. QoS CLASS CLASSIFICATION little data fluctuation and bidirectional sym- FRAMEWORK BASED ON MODIFIED metrical data transmission. GC is designated K-SVD for game data transmission including anima- tion and real-time voice or video delivery to The proposed framework in this paper is the players with very high fidelity. The traffic illustrated in Fig. 3. In Fig. 3, QoS charac- belonging to this class is different from other teristics are extracted from traffic data. Since live traffic, and may adopt the Client/Server QoS characteristics have various parameter (C/S) or hybrid P2P mode. They generally types, they need be preprocessed. After pre- have much smaller packet size and shorter processing, the core QoS characteristics of inter-arrival time. This class needs more strin- Internet multimedia traffic are obtained and gent QoS requirements regarding interactive mapped into QoS-Word, which denotes the action. Because they often have stored a lot of specific QoS local pattern. According to the data when traffic software is installed, most of Bag-of-Words (BoW) model [17], bag-QoS- the data transmitted between server and client words, which employ a vector consisting of is instruction data with smaller packet size the weighted counts are constructed. Internet during playing, which results in the smallest multimedia traffic can be described by a dic- values of data rate. tionary (codebook) constructed with bag-QoS- GC has the smallest values of downstream words. and upstream rates among the seven catego- Since K-SVD does not consider charac- ries, since only a small amount of game data teristic differences of data structures, the is sent to the client by the server that performs over-complete dictionary, trained by K-SVD, heavy-weight computations, and only the cannot provide detailed characteristics of data player’s actions are sent to the server by the structures. By introducing a characteristics-set

China Communications • September 2017 209 from a training sample, a modifi ed K-SVD is highly effi cient with an effective sparse coding employed to efficiently learn the over-com- [34]. For reducing the objective values of the pleted dictionary constructed with bag-QoS- Dictionary Learning (DL) problem, K-SVD words. Through combining given probe traffi c mainly repeats two steps: (i) sparse representa- flows, QoS class classification models are tions and (ii) updating the dictionary. Regard- obtained with a characteristic-based classifi ca- ing the step of sparse representations, based on tion scheme. Multimedia traffic flows can be the well-defined dictionary, the sparse repre- classifi ed into a corresponding QoS category/ sentations are computed by means of a pursuit class with a linear SVM classifier. The pro- method [34]. When updating the dictionary, posed method can address the problems that both atoms and their coeffi cients are updated challenged many traditional methods of im- simultaneously. The algorithm continuously proving QoS requirements. optimizes the dictionary and sparse coeffi cient until the termination condition is met. 5.1 QoS model parameters of K-SVD It is an inherent issue of DL with the algorithm K-SVD algorithm to solve the following opti- Based on [33], we assume that each of the mization problem. ‖ ‖2 ‖‖ QoS parameters adopts a real number or min{,}DX { Y− DXF},..,, s t xi 0≤∀ T 0 i (5) N can be parameterized by a real number. Let where Yy= {}ii=1 denotes a sample set, MN× N X ∈  denote a set of N column traffic Xx= {}ii=1 is a sparse coefficient matrix cor- M vectors xn ∈  , which is the only descrip- responding to its sample set vector Y, which tion of the nth traffic in the M-dimensional contains the coefficients of the dictionary M real number Euclidian space  , and con- with xi being its ith column vector, D is an

sists of M corresponding QoS parameters. over-completed dictionary matrix with dk MK× Let Dd=[ 1,, … dK ] , dk ∈  , for all k, and being its kth column vector, which leads to KM> , be an over-completed dictionary with sparse representations and can be designed by K characteristics (atoms) for a sparse represen- adapting its content to fi t a given set of signal

tation of X. examples. T0 stands for the sparsity of vector

The goal of sparse modeling is to learn a xi , which is the nonzero element bounded in ‖‖ reconstructive dictionary D by solving the fol- vector xi . • F represents the Frobenius norm ‖‖ lowing objective function: and • 0 corresponds to the  0 norm. ‖‖2 Based on [44], the K-SVD algorithm alter- (,AD )= arg min{ X− DA F }, {,}AD (4) nates between sparse representations of the st.,.‖‖α ≤∀ T, n n 00 examples based on the current dictionary and ‖‖ where • F is a Frobenius norm and an update process for the dictionary atoms so KN× A =…∈[,,αα1 N ] is called a sparse coding as to better fi t the data. The update of the dic- of X for a fi xed D, which can be obtained by tionary columns is done jointly with an update employing a K-SVD algorithm. the sparse representation coeffi cients related to The K-SVD algorithm directly generalizes it, resulting in accelerated convergence [44]. the K-means algorithm, which is simple and By utilizing K-SVD algorithm to fi nd the best

Extracting QoS QoS characteristics Constructing Traffic data characteristics preprocessing bag-QoS-words

Traffic Obtaining QoS Template-based Learning classification class model classification scheme dictionary

Fig. 3 The ﬂ owchart of the proposed QoS Class classifi cation framework

210 China Communications • September 2017 dictionary, we can represent the data samples ing vector when sample {}yi as sparse compositions. According to [44], the is represented by atom dk , and k K-SVD procedure is given in detail as follows. ωkT={i|1 ≤≤ i Kx , ( i ) ≠ 0} as the in- Step 0. This step is the initialization phase. dex value of non-zero entries in the k Based on analysis results from cap- vector xiT (). Assuming that N× ||ωk

tured dataset, we choose the classical matrix Ωk has ones at the (ωk (ii ), ) QoS parameters and their combina- th entries and zeros elsewhere, kk R tions (including the downstream/up- xxR= Tk Ω and EEk= kk Ω . Eq. (6) can stream rates) with obvious difference be transformed as follows: ‖ k‖‖ Rk ‖ degree as elements. After preprocess- EkΩ− k dx kT Ω k = E k − dx kR. (7) R ing the captured dataset (such as nor- 2) SVD decomposes matrix Ek as RT malization), we generate a dictionary Ek = UV ∆ , in order to select both D by well-defined and holds it as the maximum singular value and its constant. Meanwhile, we set J = 1 . corresponding singular vector for up- k Step 1. This is the sparse representation step. dating atom dk and sparse vector xR , K-SVD initializes dictionary D by respectively. U is an N unitary matrix a random or well-defined dictionary and its 1 st column is used to update

matrix and holds it as a constant. Ac- dk . V is an ||ωk unitary matrix and k cording to predefined dictionary D, ∆ is a diagonal matrix. xR is updated K-SVD algorithm calculates sparse with the 1 st column of V multiplied

representations xi for each sample by ∆(1,1) .

vector yi with objective function (5) Step 3. Repeat Steps 1 and 2. The iterations by means of the pursuit method in will continue until the termination [34]. condition is met by a certain thresh- Step 2. In this step, we start to update the old or a maximum number of itera- dictionary. Based on sparse coeffi- tions.

cient xi obtained from the previous step, the K-SVD algorithm corrects 5.2 Dictionary learning in modified the dictionary matrix column by col- K-SVD algorithm

umn. Assuming to correct atom dk By formulating a new objective function by , the K-SVD algorithm isolates it by combining the original objective function and the following penalty term from the construction error, the problem of sparse rep- objective function (5). resentation of QoS classes can be addressed

2 K by solving the following new optimization 2 Y−=− DX Y d x j F ∑ jT problem for dictionary construction. = j 1 F *‖ ‖‖22 ‖ 2 (AD , )= arg min X − DA22 +−η T DC K {,ADT ,} (8) jk (6) =−−Y∑ dxjT dx kT j=1, jk ≠ relax F ≈−arg min‖‖X* DA *2 2 2 (9) =E − dxk , {AD* ,} k kTF ‖‖* j s.t. & αn 0 ≤∀Tn, (10) where xT stands for the jth row in vector Y, * and Ek denotes the error caused by the kth X= (, XTη ) (11) atom removed. A* = ( AC ,η ), (12) In order to meet the sparsity constraint by ensuring most of the elements are zeros in the where η denotes the relative weight indi- k cating the contribution between training new vector xT , the K-SVD algorithm rectifies sample reconstruction and characteristic-set matrix Ek as follows: ML× k reconstruction, T ∈  is the character- 1) Define xiT () as the correspond-

China Communications • September 2017 211 istic-set for traffic, C ∈ KL× is the sparse After obtaining LCFC features by utilizing coding of T. Note that the optimization of the dictionary learned by the modified K-SVD ‖ ‖‖22 ‖ min{,ADT ,} X− DA2 +−η T DC 2 in (8) is re- dictionary learning method (as described in laxed to a simpler form (9). the last section), the detailed description of

Applying K-SVD to update dk and the kth characteristic-based classification scheme is k row in A is represented by aT at a time. As- given in the following. j j  suming that Ek=() X − ∑ dajT , aT and Ek rep- Step 1. Select n traffic flows per class ran- ≠ resent the result of discardingjk the zero entries domly from traffic data for generating j k in aT and Ek , respectively, dk and aT can be a set of characteristic-sets. obtained by solving hte following problem: Step 2. Extract QoS characteristics from traf- kk‖‖ 2 fic data, and preprocess them, includ- (,)dk a T = argmin Ek− da kTF. k (13) {,}dakT ing quantization and normalization of  After performing SVD for Ek , i.e., parameter values; T  k UVΣ=SVD( Ek ) , we obtain dk and aT as: Step 3. Calculate the QoS distance between

dUk = (:,1) (14) the probe traffic and each of the traf-

k fic streams in the characteristic-set aVT = Σ(1,1) (:,1). (15) with the core function as follows: Lastly, the non-zero values in ak are replaced T ‖‖− k ri rpi with aT . JPi= Kr( , p ) = exp − , 2σ 2 (18) Here, we formulate the problem of classi- ri=class i, = 1,2, … , fying the Internet multimedia traffic with QoS i where J ri is the QoS distance of the features into a unified objective function con- p probe traffic p and characteristic-set structed with the original objective function traffic r ( r is the ith traffic of char- and reconstruction error. Obviously, it is a typ- i i acteristic-subset r), and Kr(, p ) is ical optimization problem that can be resolved i Radial Basis Function (RBF). efficiently with the K-SVD algorithm, while Step 4. According to (18), we now average the classification performance could be signifi- row by row for obtaining the final cantly improved. representation of the network traffic 5.3 Locality constrained coding of for classification to reduce the dimen- QoS characteristics sionality of the QoS distance matrix. K All QoS characteristics in this paper are cod- r ri Zp= mean∑ Jrpi ,= class i, i = 1,2, … . (19) ed with Locality Constrained Feature Coding i=1 (LCFC) based on [31]. Step 5. Based on (19), the represented feature The LCFC method needs to satisfy the fol- is normalized and SVM is selected to lowing constraint: classify the traffic.

2 Z p 2 dist(YB , ) = ‖‖ Z p . (20) min Y− BH 2 +⋅γ exp∗H (16) ‖‖ {}H σ Z p 2 2

T s.t. 1hii = 1, ∀ , (17) VI. COMPLEXITY ANALYSIS where Y is a vector to be encoded, B is a dictionary, H is the coding obtained through To further analyze the efficiency of the pro- LCFC, ∗ denotes the element-wise multi- posed modified-K-SVD method, we study the plication, dist(⋅⋅ , ) is the Euclidean distance, time complexity of the proposed method. As dist(YB , ) and exp is the weight vector for described in Section V, the proposed method σ different codes, which is used to measure the is constructed using the K-SVD approach. The relationship between each element and its cor- dictionary learning problem in the proposed responding columns in dictionary B. modified-K-SVD is casted as a regularized

212 China Communications • September 2017 least-squares problem, where sparsity and SVM, Naive Bayes and K-NN in terms of regularization terms are used. Following [45]- accuracy, precision, Recall and F1-measure [48], in the training phase, since updating the as showed in Tables 5 and 6, and Fig. 4 in the coding coefficient for each sample is a tradi- following section. tional sparse coding problem, the corresponding time complexity for each traffic is approx- VII. EXPERIMENTAL RESULTS AND 2  imately ()MK , where  ≥ 1.2 is a constant, DISCUSSIONS M represents the feature dimensionality of the sample and K denotes the number of dictio- In this section, we focus our experiments on nary atoms. The time complexity of all N sam- classifying multimedia traffic based on QoS ples is ()NM2 K  . The time complexity of characteristics from a QoS class perspective. C We generate the datasets by capturing popu- updating the dictionary atoms is ()∑MN K , = where C is the number of traffic classesc 1 and lar Internet multimedia traffic from a campus network. The datasets include 26 natural cate- Nc is the number of training samples of class c. The overall time complexity of the proposed gories of multimedia traffic, which are divided C 2  into seven QoS classes as shown in Table 4. method is ()INM K+ ∑ IMN K , where I is c=1 Different from that of the existing works the number of the iterations. [28]-[30], our novel research issue is to effec- According to [49], the time complexity of tively utilize QoS characteristics to classify the HMM method is (3QT ) , where Q is the traffic into the appropriate QoS class. number of states in the model and T denotes In simulations, we process the raw data the number of symbols in the observation. captured by the Wireshark software tool, while Referring to [50], the time complexity of the the following characteristics are obtained, SVM method is ()N 3 , where N is the num- including protocol, port, payload, packet/ ber of training samples. Referring to [51], the flow size, flow duration, arrival time, IP ad- time complexity of the Naive Bayes method dress, and etc. After analyzing their statistical is ()Nn , where N is the total number of characteristics including average upstream samples and n stands for the number of fea- rate, average downstream rate, byte volume, tures selected for the model. Referring to [52], the time complexity of the K-NN method is Table III Average computation times ()mTL , where m is the number of feature HMM SVM Naive Bayes K-NN Modified-K-SVD selected from the original feature set, T is the 43.1s 85.9s 17.5s 48.4s 671.3s number of samples in the test set, and L denotes the number of samples in the train. Table IV Components of the experimental dataset In this paper, the experiments are imple- Number of Data Amount QoS Class Traffic type mented with Matlab running in ThinkStation. Traffic Flows (GB) The training time of the proposed method is 1 BSDV Tudou, Youku video(SD) 100 1.05 about 8 hours. We compare our method against 2 BHDV Tudou, Youku video(HD) 100 2.12 HMM, SVM, Naive Bayes and K-NN in terms BBC Web video, CCTV Web 3 WV 100 2.06 of the average computation time for classify- video ing one testing traffic. The average compu- Xunlei, BitTorrent, eDonkey, 4 TSV 100 8.39 tational time of HMM, SVM, Naive Bayes, Emule K-NN, and the proposed modified-K-SVD Sopcast, TVAnt, UUSee, methods are given in Table 3. The time com- 5 BSV Skype, PPlive, PPStream, 100 4.87 plexity of the proposed modified-K-SVD PPMate method is the highest among HMM, SVM, 6 IV ICQ, QQ, MSN 100 4.78 Naive Bayes, and K-NN. However, the pro- dota, dota2, lol, Dreams 7 GC Western adventure, Against 100 1.21 posed method is more competitive than HMM, War

China Communications • September 2017 213 Inter-arrival Time (IAT), delay, jitter, the formula as follows: different combination of QoS characteristics TP P = , (22) (such as sub-flows), the different transforms TP+ FP corresponding to characteristics (such as Dis- where TP denotes the total number of sam- crete Fourier Transform ) and etc. We extract ples which are distinguished correctly by the QoS characteristics with obvious difference classifier and are the correct samples indeed, from above characteristics (including down/ and FP denotes the total number of samples up-stream rate) and preprocess them including which are distinguished correctly by the clas- quantization and normalization of parameter sifier and are the incorrect samples indeed. value to obtain corresponding parameter val- 3) Recall (R). It is expressed by the formula as ues. After the dictionary is constructed with an follows: excess quantity of bag-QoS-words, Locality TP R = , (23) Constrained Feature Coding (LCFC) features TP+ FN of QoS classes are extracted. Referring to [31], where FN denotes the total number of sam- we set the dictionary size to be 100 by using ples which are distinguished incorrectly by the an SVM classifier. We randomly partition the classifier and are the correct samples indeed. dataset into 50 training traffic flows per class 4) F1-measure ( F1 ) [35]. It is represented by and the rest into testing traffic flows. Then, we the formula as follows: evaluate our method according to new QoS 2× PR F = . (24) classes introduced in Table 2. 1 PR+ In order to evaluate the effectiveness of our The experimental results are presented in method, in combination with selected charac- Table 5. We can see that our method achieves teristics, we generate a characteristics vector the best performance among all the methods to label the original Internet multimedia traffic and has higher classification accuracy than the flow. On a Matlab platform, we compare our next best result by nearly 3.43%. Since Naive method against HMM [28], SVM [29], Naive Bayes, HMM, and SVM are affected by the Bayes and K-Nearest Neighbor (K-NN) [30] training dataset in the learning stage, they are in terms of accuracy, precision, Recall, and dependent on the dataset’s specific character- F1-measure. The results are obtained from 50 istics, and have different recognition effects runs. for different traffic streams belonging to the The following performance metrics are same QoS class. For example, BitTorrent has used in the experimental evaluation. a higher accuracy than eDonkey for Naive 1) Accuracy (Ac). For the given test datasets, Bayes although both of them belong to TSV. accuracy is determined by finding the ratio SVM has a very poor recognition performance of the number of samples correctly classi- for PPlive, but has an extremely good perfor- fied by the classifier to the total number of mance for BitTorrent and eDonkey. PPStream samples. has higher classification accuracy than Sop- TP Ac = Σ , (21) cast in HMM. Since K-NN only calculates the SUM “nearest” neighbor samples, the recognition

where TPΣ denotes the number of samples performance is also influenced by training set correctly classified by the classifier, and SUM samples in the classification process. represents the total number of samples in the In QoS class classification, different QoS dataset. classes may utilize the same protocol, such as 2) Precision (P) [35]. It is represented by the BSDV, BHDV, TSV and BSV classes all being able to use P2P mode, and the same QoS class Table V Comparison of traffic classification methods in terms of accuracy may access different protocols, such as GC Classification Method Proposed method Naive Bayes HMM SVM K-NN using both P2P and C/S modes. Since Naive Accuracy (%) 98.29 88 89.71 94.86 92.85

214 China Communications • September 2017 Bayes, HMM, SVM and K-NN are used to Table VI Comparison of traffi c classifi cation methods in terms of precision classify specific protocols that are not of the Classifi cation BSDV BHDV WV TSV BSV IV GC QoS class, they have poor performance in QoS Method class classification due to the existing iden- Proposed method 0.9796 0.9608 1 0.9608 0.98 1 1 tification error rate, in which some traffic is Naive Bayes 0.8125 0.7885 0.9574 0.8333 0.8654 0.92 1 classifi ed with an inappropriate QoS class with HMM 0.8511 0.8113 0.9583 0.8654 0.8846 0.9216 1 the same protocol, as in the exampling traffi c SVM 0.9167 0.8846 0.9796 0.9231 0.96 0.98 1 belonging to the BSDV class being classifi ed K-NN 0.8980 0.8824 0.96 0.92 0.9038 0.94 1 as traffic of the BHDV class, and traffic belonging to the BHDV class being classifi ed as traffi c of the BSDV class. This is because the difference between BSDV and BHDV classes is not very obvious as SD and HD standards 1 constantly change in network. In Table 6, we can see that our method 0.8 clearly has higher precision performance than e 0.6 the other four methods. This is because our

method considers the dispersion of characteris- F1−measur 0.4 tics caused by network dynamics, helps reduce Proposed Naive Bayes their influence, and improves the precision 0.2 HMM performance. In the same way, our method SVM K−NN improves F1-measure performance (shown in 0 BSDV BHDV WV TSV BSV IV GC Fig. 4). Our method appears to have a higher Traffic Type F1-measure value than other methods. In particular, the fi ve classifi cation methods have the Fig. 4 Comparison of traffi c classifi cation methods in term of F1-measure same precision regarding traffi c belonging to the GC class, but our method has a higher value in Recall than that of other methods since framework based on the concept of QFAg. some traffi c with dispersive QoS characteris- Based on statistical analysis of multimedia tics are divided into other QoS classes when flows collected from a large-scale, real net- using Naive Bayes, HMM, SVM and K-NN. work, we defined seven QoS categories with Unlike existing work [28]-[30], our method features of downstream/upstream rates. By in- takes into account feature selection from the vestigating the sparsity property of the multi- point of view of QoS class, which can effec- media streaming QoS characteristic, this paper tively improve the classification accuracy. In utilized a modifi ed K-SVD to train dictionary the proposed method, downstream/upstream extracted from training samples. By learning rates reflect the essential QoS characteristics a characteristic-set to obtain sparse represen- of multimedia traffi c and help improve recog- tation for multimedia traffic, we proposed a nition accuracy. As a result, with our method, characteristic-based method to classify multi- it is possible to obtain potential QoS patterns media traffi c. Experimental results reveal that to improve the classifi cation performance. the proposed method can improve the performance significantly compared to other state- VIII. CONCLUSIONS of-the-art methods.

This paper addressed the problem of effective ACKNOWLEDGEMENTS classifying multimedia traffic with a suitable granularity, and presented a modifi ed K-SVD This work was supported in part by the Na- Internet multimedia traffic identification tional Natural Science Foundation of China

China Communications • September 2017 215 (NO. 61401004, 61271233, 60972038), Plan [10] S. Mao, S. S. Panwar, and Y. T. Hou, “On optimal of introduction and cultivation of university traffic partitioning for multipath transport,’’ in Proc. IEEE INFOCOM 2005, Miami, FL, Mar. 2005, leading talents in Anhui (No.gxfxZD2016013), pp. 2325-2336. the Natural Science Foundation of the High- [11] Q.M. Qadir, A.A. Kist, and Z. Zhang, “A novel er Education Institutions of Anhui Province, traffic rate measurement algorithm for Quality China (No. KJ2010B357), Startup Project of of Experience-Aware video admission control,’’ IEEE Transactions on Multimedia, vol. 17, no. 5, Anhui Normal University Doctor Scientific pp. 711-722, May 2015. Research (No.2016XJJ129), the US Nation- [12] R. Misun, Y. Kim Youngmin, and H. Park, “Sys- al Science Foundation under grants CNS- tematic QoS class mapping framework over 1702957 and ACI-1642133, and the Wireless multiple heterogeneous networks,’’ in Proc. 8th International Conference on Next Generation Engineering Research and Education Center at Teletraffic and Wired/Wireless Advanced Net- Auburn University. working, St. Petersburg, Russia, Sept. 2008, pp. 212-221. References [13] A. Elnaka and Q.H. Mahmoud, “QoS traffic map- [1] Z. He, S. Mao, and T. Jiang, “A survey of QoE ping for a multi-participant session in unified Proc. 26th IEEE driven video streaming over cognitive radio communications networks,’’ in Canadian Conference on Electrical and Comput- networks,’’ IEEE Network, vol. 29, no. 6, pp. 20- er Engineering 25, Nov./Dec. 2015. , Regina, Canada, May 2013, pp. [2] Y. Xu and S. Mao, “A survey of mobile cloud 1-5. computing for rich media applications,’’ IEEE [14] J. Liao, P.A. Chou, C. Yuan, Y. Hu, and W. Zhu, Wireless Communications, Special Issue on Mo- “Online allocation of communication and com- bile Cloud Computing, vol. 20, no. 3, pp. 46-53, putation resources for real-time multimedia IEEE Transactions on Multimedia June 2013. services,’’ , vol. [3] Z. He, S. Mao, and S. Kompella, “Quality of Ex- 15, no. 3, pp. 670-683, Apr. 2013. perience driven multi-user video streaming in [15] Z.-J Wang, Y.-N. Dong, and X. Wang, “A dynamic cellular cognitive radio networks with single service class mapping scheme for different QoS IEEE Systems channel access,’’ IEEE Transactions on Multime- domains using flow aggregation,’’ Journal dia, vol. 18, no. 7, pp. 1401-1413, July 2016. , vol. 9, no. 4, pp. 1299-1310, Dec. 2014. [4] Z. He and S. Mao “Adaptive multiple description [16] J. Zhang, C. Chen, X. Yang, W. Zhou, and X. coding and transmission of uncompressed vid- Yong, “Internet traffic classification by aggre- IEEE eo over 60GHz networks,’’ ACM Mobile Com- gating correlated naive Bayes predictions,’’ Transactions on Information Forensics and Secu- puting and Communications Review (MC2R), vol. rity 18, no. 1, pp. 14-24, Jan. 2014. , vol. 8, no. 17, pp. 5-15, Jan. 2013. [5] D. Hu and S. Mao, “Streaming scalable videos [17] S. Samanta and B. Chanda, “Space-time facet IEEE over multi-hop cognitive radio networks,’’ IEEE model for human activity classification,’’ Transactions on Multimedia Transactions on Wireless Communications, vol. , vol. 16, no. 6, pp. 9, no. 11, pp. 3501-3511, Nov. 2010. 1525-1535, Oct. 2014. [6] S. Lin, Y. Wang, S. Mao, S.S. Panwar, “Video [18] S.-H. Yoon, J.-S. Park, M.-S. Kim, C. Lim, and J. transport over ad hoc networks using multiple Cho, “Behavior signature for big data traffic Proc. 2014 International Con- paths,’’ in Proc. IEEE ISCAS 2002, Scottsdale, AZ, identification,’’ in ference on Big Data and Smart Computing May 2002, pp. 57-60. (BIG- [7] S. Mao, X. Cheng, Y. T. Hou, H. D. Sherali, and J. COMP), Bangkok, Thailand, Jan. 2014, pp. 261- H. Reed, “On joint routing and server selection 266. for multiple description video in wireless ad [19] M. Jaber, R.G. Cascella, and C. Barakat, “Can we hoc networks,’’ IEEE Transactions on Wireless trust the inter-packet time for traffic classifica- Proc. IEEE ICC 2011 Communications, vol. 6, no. 1, pp. 338-347, Jan. tion?’’ in , Kyoto, Japan, June 2007. 2011, pp. 1-5. [8] Y. Huang and S. Mao, “Downlink power control [20] N.M. Markovich and U.R. Krieger, “statistical for multi-user VBR video streaming in cellular analysis and modeling of peer-to-peer multi- Next Generation Internet networks,’’ IEEE Transactions on Multimedia, vol. media traffic,’’ , LNCS 15, no. 8, pp. 2137-2148, Dec. 2013. 5233, pp. 70-97, Springer-Verlag: Berlin Heidel- [9] S. Mao, S. S. Panwar, and Y. T. Hou, “On mini- berg, 2011. mizing end-to-end delay with optimal traffic [21] N.M. Markovich and U.R. Krieger, “Statistical partitioning,’’ IEEE Transactions on Vehicular characterization of QoS aspects arising from Proc. First Technology, vol. 55, no. 2, pp. 681-690, Mar. the transport of Skype VoIP flows,’’ in International Conference on Evolving Internet 2006. , Cannes/La Bocca, France, Aug. 2009, pp. 9-14.

216 China Communications • September 2017 [22] H. Tobias and A. Binzenhofer, “Analysis of Skype worked Embedded Systems for Enterprise Appli- VoIP traffic in UMTS: End-to-end QoS and QoE cations (NESEA), Perth, Australia, Dec. 2011, pp. measurements,’’ Elsevier Computer Networks, 1-6. vol. 52, no. 3, pp. 650-666, Feb. 2008. [36] L. Zhou, Z. Yang, H. Wang, and M. Guizani, “Im- [23] M. Claypool, D. Finkel, A. Grant, and M. Solano, pact of execution time on adaptive wireless vid- “Thin to win? Network performance analysis of eo scheduling,’’ IEEE Journal on Selected Areas the OnLive thin client game system,’’ in Proc. in Communications, vol. 32, no. 4, pp. 760-772, 2012 Annual Workshop on Network and Systems Apr. 2014. Support for Games (NetGames), Venice, Italy, [37] C. Xu, Z. Li, J. Li, H. Zhang, and G.-M. Munte- Nov. 2012, pp. 1-6. an, “Cross-layer fairness-driven concurrent [24] M. Sanli, E.G. Schmidt, and H.C. Guran, “A flow multipath video delivery over heterogeneous aggregation method for the scalable and effi- wireless networks,’’ IEEE Transactions on Circuits cient quality of service support in Next Gener- and Systems for Video Technology, vol. 25, no. 7, ation Networks,’’ in Proc. IEEE GLOBECOM’13, pp. 1175-1189, July 2015. Atlanta, GA, Dec. 2013, pp. 2243-2248. [38] Y. He, N. Sang, C. Gao, and J. Han, “Online un- [25] G. Thatte, U. Mitra, and J. Heidemann, “Paramet- supervised learning classification of pedestrian ric methods for anomaly detection in aggregate and vehicle for video surveillance,’’ Chinese traffic,’’ IEEE/ACM Transactions on Networking, Journal of Electronics, no. 26, no. 1, pp. 145-151, vol. 19, no. 2, pp. 512-525, Apr. 2011. Jan. 2017. [26] S. Valenti and D. Rossi, “Identifying key features [39] N. Bitar, and H.H. Refai, “A probabilistic ap- for P2P traffic classification,’’ in Proc. 2011 IEEE proach to improve the accuracy of axle-based International Conference on Communications automatic vehicle classifiers,’’ IEEE Transactions (ICC), Kyoto, Japan, June 2011, pp. 1-6. on Intelligent Transportation Systems, vol. 18, [27] L. Noirie, E. Dotaro, G. Carofiglio, A. Dupas, P. no. 3, pp. 537-544, Mar. 2017. Pecci, D. Popa, and G. Post, “Semantic network- [40] F. Wang, T. Xu, T. Tang, M. Zhou, and H. Wang, ing: Flow-based, traffic-aware, and self-man- “Bilevel feature extraction-based text mining for aged networking,’’ Bell Labs Technical Journal, fault diagnosis of railway systems,’’ IEEE Trans- vol. 14, no. 2, pp. 23-38, Apr. 2009. actions on Intelligent Transportation Systems, [28] S. Maheshwari, S. Mahapatra, C.S. Kumar, and K. vol. 18, no. 1, pp. 49-58, Jan. 2017. Vasu, “A joint parametric prediction model for [41] Y.-T. Han and H.-S. Park, “Game traffic classi- wireless internet traffic using Hidden Markov fication using statistical characteristics at the Model,’’ Springer Wireless Networks, vol. 19, no. transport layer,’’ ETRI Journal, vol. 32, no. 3, pp. 6, pp. 171-1185, Aug. 2013. 22-32, Feb. 2010. [29] J. He, Y. Yang, Y. Qiao, and C. Tang, “Accurate [42] S. Bu and Y. Zhang, “Single-frame and multi- classification of P2P traffic by clustering flows,’’ frame image super-resolution based on local- IEEE/CIC China Communications, vol. 10, no. 11, ity-constrained linear coding,’’ Journal of Jilin pp. 42-51, Nov. 2013. University ( Engineering and Technology Edi- [30] H. Dibeklioglu, A.A. Salah, T. Gevers, “Recogni- tion), vol. 43, no. 1, pp. 365-370, Mar. 2013. tion of genuine smiles,’’ IEEE Transactions on [43] J. Wang, J. Yang, K. Yu, F. Lv, T. Huang, and Y. Multimedia, vol. 17, no. 3, pp. 279-294, Mar. Gong, “Locality-constrained linear coding for 2015. image classification,’’ IEEE Conference on Com- [31] Q. Li Qun, H. Zhang, J. Guo, B. Bhanu, and A. puter Vision and Pattern Recognition (CVPR) Le, “Reference-based scheme combined with 2010, San Francisco, CA, June 2010, pp. 3360- K-SVD for scene image categorization,’’ IEEE 3367. Signal Processing Letters, vol. 20, no. 1, pp. 67- [44] M. Aharon, M. Elad, and A. Bruckstein, “K-SVD: 70, Jan. 2013. An algorithm for designing overcomplete dic- [32] S. Gold, “Hacking on the hoof,’’ IET Engineering tionaries for sparse representation,’’ IEEE Trans- and Technology, vol. 7, no. 3, pp. 80-83, Apr. actions on Signal Processing, vol. 54, no. 11, pp. 2012. 4311-4322, Nov. 2006. [33] K. Nahrstedt and J. Smith, “The QoS broker,’’ [45] G. Somasundaram, R. Sivalingam, V. Morellas, IEEE Multimedia, vol. 2, no. 1, pp. 53-67, Mar. and N, Papanikolopoulos, “Classification and 1995. counting of composite objects in traffic scenes [34] Z. Zhu, F. Guo, H. Yu, and C. Chen, “Fast single using global and local image analysis,’’ IEEE image super-resolution via self-example learn- Transactions on Intelligent Transportation Sys- ing and sparse representation,’’ IEEE Transac- tems, vol. 14, no. 1, pp. 69-81, Mar. 2013. tions on Multimedia, vol. 16, no. 8, pp. 2178- [46] R. Rubinstein, M. Zibulevsky, and M. Elad, “Ef- 2190, Dec. 2014. ficient implementation of the K-SVD algorithm [35] Z. Yu, G.B. Kil, Y.-D. Choi, and S.-H. Kim, “Traffic using batch orthogonal matching pursuit,’’ classification based on visualization,’’ in Proc. Tech. Rep., Dept. Comput. Sci., Technion, Haifa, 2011 IEEE 2nd International Conference on Net- Israel, Apr. 2008.

China Communications • September 2017 217 [47] R. Ptucha and A.E. Savakis, “LGE-KSVD: Robust interests include wireless networking, multimedia sparse representation classification,’’ IEEE Trans- communications and network traffic identification. actions on Image Processing, vol. 23, no .4, pp. 1737-1750, Apr. 2014. Shiwen Mao, (S’99-M’04- [48] X. Shu, J. Tang, G.-J. Qi, Z. Li, Y.-G. Jiang, and SM’09) received his Ph.D. in S. Yan, “Image classification with tailored fine- electrical and computer engi- grained dictionaries,’’ IEEE Transactions on Cir- neering from Polytechnic Uni- cuits and Systems for Video Technology, vol. PP, versity, Brooklyn, NY in 2004. no. 99, pp. 1-1, Sept. 2016. Currently, he is the Samuel [49] J.R. Deller and R.K. Snider, “Reducing redundant Ginn Distinguished Professor computation in HMM evaluation,’’ IEEE Trans- and Director of Wireless Engi- actions on Speech and Audio Processing, vol. 1, neering Research and Education Center at Auburn no. 4, pp. 465-471, Oct. 1993. University, Auburn, AL, USA. His research interests [50] H. Chen, P. Tino, and X. Yao, “Probabilistic clas- include wireless networks, multimedia communica- sification vector machines,’’ IEEE Transactions tions, and smart grid. He is a Distinguished Lecturer on Neural Networks, vol. 20, no. 6, pp. 901-914, of IEEE Vehicular Technology Society. He is on the June 2009. Editorial Board of IEEE Transactions on Multimedia, [51] D. Ruta, “Robust method of sparse feature se- IEEE Internet of Things Journal, IEEE Multimedia, lection for multi-label classification with Naive IEEE/CIC China Communications, among others, and Bayes,’’ in Proc. 2014 Federated Conference on a Steering Committee Member of IEEE Transactions Computer Science and Information Systems, on Multimedia and IEEE Transactions on Network Sci- Warsaw, Poland, Sept. 2014, pp. 375-380. ence and Engineering. He serves as TPC Chair of IEEE [52] Z.-H. Wang, Z.-S. Hou, Y. Gao, and Q. Liu, “Study INFOCOM 2018, Area TPC Chair of IEEE INFOCOM on scale development of boolean medicine 2017 and 2016, Technical Program Vice Chair for data based on the GA and improved k-NN Information Systems (EDAS) of IEEE INFOCOM 2015, algorithm,’’ in Proc. 2008 International Confer- symposium co-chairs for many conferences, includ- ence on BioMedical Engineering and Informatics, ing IEEE ICC, IEEE GLOBECOM, ICCCN, among others, Sanya, China, May 2008, pp. 367-371. and Steering Committee Voting Member of IEEE ICME and AdhocNets. He is the Chair of IEEE ComSoc Biographies Multimedia Communications Technical Committee. Zaijian Wang, received his He received the 2015 IEEE ComSoc TC-CSR Distin- BE degree (2002) from Anhui guished Service Award, the 2013 IEEE ComSoc MMTC Polytechnic University, MSc Outstanding Leadership Award, and the NSF CAREER degree (2005) from University Award in 2010. He is a co-recipient of the Best Demo of Science and Technology of Award of IEEE SECON 2017, the Best Paper Awards of China, and PhD degree (2015) IEEE GLOBECOM 2016, IEEE GLOBECOM 2015, IEEE from Nanjing University of WCNC 2015, and IEEE ICC 2013, and the 2004 IEEE Posts and Telecommunications. Communications Society Leonard G. Abraham Prize His current research interests focus on multimedia in the Field of Communications Systems. big data, end-to-end QoS provisioning and wired/ Xinheng Wang, wireless multimedia streaming. He is currently an as- (SM’14) re- sociate professor at College of Physics and Electronic ceived the B.Eng. and M.Sc. de- Information, An’hui Normal University, Wuhu, China. grees in electrical engineering from Xi’an Jiaotong University, Yuning Dong (M’07), received Xi’an, China, in 1991 and 1994, his B.E and M.E degrees from respectively, and the Ph.D. Nanjing University of Posts & degree in computing and elec- Telecommunications (NUPT), tronics from Brunel University, his Ph.D degree from Southeast Uxbridge, U.K., in 2001. He is currently a Professor of University, all in electrical engi- networks with the School of Computing, University neering, and his M.Phil degree of the West of Scotland, Paisley, U.K. His current re- in Computer Science from The search interests include wireless networks, Internet Queen’s University of Belfast (QUB). He is currently a of Things, converged indoor positioning, cloud com- professor with the College of Communications and puting, and applications of wireless and computing Information Engineering at NUPT. He was a Brit- technologies for health care. He has close engage- ish Council postdoctoral fellow at Imperial College ment with the industry. London, 1992-93; a visiting scientist at University of Texas, 1993-95; and a research fellow at QUB and the University of Birmingham, 1995-98. His research

218 China Communications • September 2017