RADIO Frequency Identification (RFID) devices are widely. A Multiple Hashing Approach to Complete Identification of Missing RFID Tags

Transcription

1 A Multiple Hashing Approach to Complete Identiication o Missing RFID ags Xiulong Liu, Keqiu Li*, Geyong Min, Yanming Shen, Alex X. Liu, Wenyu Qu Abstract Owing to its superior properties, such as ast identiication and relatively long interrogating range over barcode systems, Radio Frequency Identiication (RFID) technology has promising application prospects in inventory management. his paper studies the problem o complete identiication o missing RFID tag, which is important in practice. ime eiciency is the key perormance metric o missing tag identiication. However, the existing protocols are ineective in terms o execution time and can hardly satisy the requirements o real-time applications. In this paper, a Multi-hashing based Missing ag Identiication (MMI) protocol is proposed, which achieves better time eiciency by improving the utilization o the time rame used or identiication. Speciically, the reader recursively sends bitmaps that relect the current slot occupation state to guide the slot selection o the next hashing process, thereby changing more empty or collision slots to the expected singleton slots. We investigate the optimal parameter settings to maximize the perormance o the MMI protocol. Furthermore, we discuss the case o channel error and propose the countermeasures to make the MMI workable in the scenarios with imperect communication channels. Extensive simulation experiments are conducted to evaluate the perormance o MMI, and the results demonstrate that this new protocol signiicantly outperorms other related protocols reported in the current literature. Index erms RFID, missing tag identiication, multi-hashing. I. INRODUCION RADIO Frequency Identiication (RFID) devices are widely deployed in many application scenarios, such as supply chain management [] [2] and inventory control [3] [4] [5], where the missing tag problem is an important but challenging issue [6] [7] [8] [9]. his issue can be generally classiied into three categories: () missing-tag event detection protocols ocus on detecting whether any RFID tags are missing or not instead o exactly pinpointing which tags are missing [6] [7] [9]; (2) probabilistic missing-tag identiication protocols can pinpoint which RFID tags are missing (i.e., ind out the ID inormation o the missing RFID tags), but do not guarantee to report all missing ones (e.g., Protocol in [8]); (3) complete missing-tag identiication protocols ocus on pinpointing the ID inormation o all missing RFID tags and guarantee 00% reporting (e.g., IIP in [0], Protocols 2 and 3 in [8]). X. Liu, K. Li and Y. Shen are with the School o Computer Science and echnology, Dalian University o echnology, No 2, Linggong Road, Dalian 6023, China. Corresponding Author: K. Li. keqiu@dlut.edu.cn. G. Min is with the Department o Computing, University o Bradord, Bradord, BD7 DP, United Kingdom. G.Min@brad.ac.uk. A. X. Liu is with the Department o Computer Science and Engineering, Michigan State University, East Lansing, MI, USA. alexliu@cse.msu.edu. W. Qu is with the School o Inormation and echnology, Dalian Maritime University, Dalian 6026, China. eunice.qu@gmail.com. Manuscript received January 30, 203. his paper ocuses on the third sub-problem complete missing-tag identiication, which is o great importance and irreplaceable in some scenarios. For example, in a warehouse that suers rom burglary, it is essential to monitor the items. However, simply detecting the missing-tag event is not enough. In addition, we need to obtain the detailed inormation (e.g., category, price, etc.) o the missing items so as to assess the seriousness o the loss and take dierent countermeasures. In this situation, complete missing tag identiication protocol is preerred. Despite o its practical importance, the problem o complete identiication o missing tags is still underinvestigated and solicits new eicient solutions. o the best o our knowledge, the existing advanced protocols or addressing the problem o missing tag identiication include: () the Iterative ID-ree Protocol (IIP) proposed in [0]; and (2) a group o protocols reported in [8]. In what ollows, we will present and analyze these schemes, respectively. he IIP scheme is based on the classical Framed Slotted Aloha communication mechanism. By pseudo-randomly employing a hash unction (shared by both the reader and tags), the reader can predict the singleton slots, in which only one tag is expected to respond; the collision slots, in which two or more tags are expected to respond; and the empty slots, in which no tag is expected to respond. Based on the observation o the actual state o each slot, the reader identiies the missing tags. In IIP, or time-eiciency, the tag response is -bit. I an expected singleton slot turns out to be empty, the reader asserts that the tag corresponding to this slot is missing. Although many recent studies (e.g., [], [2], and [3]) have been reported to make use o the collision slots, IIP does not leverage the expected collision slots. he reasons are exempliied as ollows. For an expected collision slot, i all tags corresponding to this slot respond as expected, the reader senses a busy slot. On the other hand, i only some o them are missing, and thus at least one tag is present and responds, the reader still senses a busy slot. Clearly, in this case, the reader cannot identiy the missing tags. he reader can identiy the missing tags during an expected collision slot unless all the tags corresponding to this slot are missing and this slot turns out to be empty. But this probability is very small. hereore, without loss o generality, IIP only leverages the expected singleton slots. Whereas, the expected empty slots and the expected collision slots that account or nearly 48% are not used and become wasted, which leads to the deiciency o IIP. In [8], Zhang et al. proposed three protocols to identiy the missing tags in the multi-reader scenarios. Protocol 2 and 3

2 2 in [8] can identiy all the missing tags but Protocol cannot. Beneiting rom the cooperation among RFID readers, these protocols can reduce the time or identiying all the missing tags by up to 75% when compared to IIP. In the single reader scenarios (or in the scenarios where the number o readers is small), IIP still runs aster than the protocols in [8]. his paper investigates a multi-hashing approach to relieve the deiciency o IIP and proposes the Multi-hashing based Missing ag Identiication (MMI) protocol. In this protocol, multiple hashing processes are repeated to increase the proportion o the expected singleton slots improving the utilization o time rame. he challenge is how to guarantee that the achieved singleton slots will not be selected in the next hashing process. Accordingly, we investigate a bitmap to guide the next hashing process. Speciically, since the slot occupation states o the rame is predictable to the reader, it could construct a bitmap, in which s indicate the singleton slots that cannot be selected in the next hashing process; and 0s indicate the collision or empty slots that can be selected in the next hashing process. he reader then broadcasts this bitmap to guide the next hashing process. o maximize the perormance o the proposed MMI, we investigate the optimization o the involved parameters including rame size and the hashing count. Since the communication channel is error-prone in the real environment, this paper discusses the impact o channel error and investigates the corresponding countermeasures. Suicient analysis and experiments maniest that MMI reduces 32% and about 90% o the required execution time, when compared to IIP and Protocol 3 (the best protocol in [8]), respectively. he major contributions o this paper are summarized as ollows: ) A Multi-hashing based Missing ag Identiication (MMI) protocol is proposed to reduce the proportion o the expected empty slots and expected collision slots that are not leveraged and trigger the deiciency o the existing IIP scheme. 2) he optimal parameter settings o the proposed MMI protocol are thoroughly investigated in order to maximize its perormance. 3) he impact o channel errors on the proposed MMI protocol and the corresponding countermeasures are investigated. 4) he perormance o the proposed MMI protocol is thoroughly evaluated by virtue o extensive simulation experiments and is compared to the other related protocols reported in the current literature, maniesting its superior eiciency. he rest o this paper is organized as ollows. he related work is reviewed in Section II. Section III describes the problem to be addressed in this paper and presents the system model. We propose the MMI protocol and present the related proos in Section IV. In Section V, extensive simulation experiments are conducted to evaluate the perormance o the MMI protocol. Finally, this paper is concluded in Section VI. II. RELAED WORK In the current literature, many studies have been conducted to address various important problems in the ield o RFID. Most o the previous work concentrated on the problem o tag identiication, which is to identiy the IDs rom a large number o tags as quickly as possible. he existing tag identiication schemes can be generally classiied into Alohabased schemes [4] [5] [6] and tree-based schemes [7] [8] [9]. In recent years, a new technical problem o inormation collection has attracted much attention, which aims to collect the inormation (e.g., the environment temperature) generated by the sensor-augmented RFID tags [20], instead o just simple ID inormation. In [20], Chen et al. proposed a multi-hashing approach named Multi-hash Inormation Collection protocol (MIC) to increase the utilization o the time rame. In MIC, the reader assigns tags to slots by using k hashing unctions, which is equivalent to multi-hashing. Intuitively, a single hashing unction can generate about 37% singleton slots (when the number o tags is the same as the number o slots). wo hashing unctions can generate 37% + ( 37%) 37% singleton slots. he more hashing unctions are employed, the more singleton slots can be achieved. hen the reader sends a hash-selection vector to inorm the tags which hashing unctions they should adopt. However, in MIC, the tags have to store the whole hash-selection vector when searching their proper slots, which poses challenges on the very limited storage capacity o RFID tags, especially or the passive tags. he missing tag problem, which is o great practical importance, also attracted much attention. In [6], an et al. proposed the rust Reader Protocol (RP) to detect the missing-tag event with a given probability when the number o missing tags exceeds a threshold. In [7], Luo et al. investigated birthday paradox to detect the missing tag event and presented the corresponding energy-time tradeo. o urther accelerate the detection process, a multi-hashing method was proposed to increase the utilization o the time rame. he Multi-Seed Missing-tag Detection (MSMD) [9] uses multiple hashing seeds to increase the proportion o the expected singleton slots in the time rame. Speciically, the reader uses a hashing seed to map the tags to slots, which is logically equivalent to a hashing process (multiple seeds correspond to multiple hashing processes). A slot may be an expected singleton using a seed, but a collision using another seed. o maximize the proportion o the expected singleton slots, the reader selects the best hashing seed or each slot such that this slot can be singleton. Ater that, the reader constructs a seed-selection vector V, which contains selectors, one or each slot in the time rame. hen, the reader broadcasts the seed-selection vector V to tell the tags to choose the seed thereby guiding their slot selection (i.e., the multi-hashing processes). MSMD also suers rom the storage limitation. he protocols reported in [6] [7] [9] are able to detect the missing-tag event only, but cannot exactly pinpoint which RFID tags are missing. he Iterative ID-ree Protocol (IIP) presented in [0] can completely identiy the missing tags and guarantee 00% reporting. As aorementioned, the singleton slots in the IIP protocol are used to veriy the presence o the tags. he

3 3 expected empty slots and collision slots contribute nothing to missing tag identiication and are wasted. he ineiciency due to the collision slots has been noticed in [0], and the authors investigated a method to turn some o the collision slots into singleton slots. However, the empty slots are not discussed and still wasted directly. According to the theoretical analysis in [0], we ind that the expected empty slots and expected collision slots still account or nearly 48% even though they have tried to turn some collision slots into singleton slots. Clearly, IIP still has a large space or improvement. In reality, a single reader usually cannot cover the whole monitoring area. In [8], Zhang et al. proposed three protocols to identiy missing tags in the multi-reader scenarios, where all the readers perorm synchronized and parallel scans. Protocol 2 and 3 in [8] can identiy all the missing ones while Protocol cannot. Experimental results in [8] demonstrate that Protocol 3 reduces the time or identiying all the missing tags by up to 75% in comparison to IIP. he superiority over IIP beneits rom the cooperation o the readers, whereas, in the single reader scenarios, IIP in [0] still outperorms Protocol 3 in [8]. Actually there is another type o missing tag problem [2], in which the missing tags represent the tags that are let unread due to errors in the communication link towards the reader, e.g., caused by the obstacles in the radio path. In other words, the study in [2] investigated the problem o tag identiication (i.e., reading the tag IDs) in a scenario with non-perect communication channels. he authors studied how to minimize the probability o missing (miss-reading) a tag, which is dierent rom the missing tag identiication problem addressed in this study. A. Problem Description III. SYSEM AND PROBLEM high-rate network link { ID, ID2, K, IDN} Fig. : Problem description. Assume each item under monitoring in a warehouse is attached with an RFID tag. An RFID reader located in the center o this warehouse periodically scans the tags within its interrogation range. his scenario is illustrated in Fig. where there is one reader and N tags and all the tags are within the range o this reader. It is worth noting that the proposed MMI protocol can be easily extended to the multireader scenarios [22]. For the purpose o clarity, we consider the case o a single reader. Let S denote the tag set, where S = { ag,..., ag i,..., ag N }, i (, N). All tags are equipped with the same uniorm hash unction H( ), and each o them possesses a unique tag ID. he reader is able to access the tag IDs stored in a database [6] [7] [8] [9] [0]. Some tags may be missing due to thet or other reasons. he problem studied in this paper is to completely identiy all the missing RFID tags in a ast way. able I summarizes the notations used this paper. Symbols N S N S slot x ρ R H( ) m B. ime Slots ABLE I: Notations Descriptions he number o tags whose IDs are stored in the data base he set o tags that should be present, S = N he number o tags participating in this slotted rame he set o tags whose presence has not been veriied, S = N he rame size, i.e. the number o slots in this rame he x th slot in a time rame he load actor that is equal to N / he random number that is resh in each round he hash generator with a uniorm distribution he hashing count in a round o MMI he proposed MMI protocol is based on the slotted Aloha communication mechanism which will be brieed below. he communication between the reader and the tags is in a timeslotted way. he reader synchronizes the slots by broadcasting the end slot command. Each tag has a slot clock which is initialized with a random slot number. A tag down-counts its slot clock one each time when the reader indicates that the current slot has ended. A tag responds when its slot clock reaches zero. According to Philips I-Code [23], we have the ollowing two claims: () i each tag response is at least 0 bits, the reader can distinguish three types o slots: the empty slot in which no tag responds in the slot; the singleton slot in which exactly one tag responds; and the collision slot in which more than one tag responds. (2) i each tag response is less than 0 bits (e.g., bit only), the reader can distinguish two types o slots only: the idle slot in which no tag responds; and the busy slot in which at least one tag responds. he above two claims have also been adopted in the literature [5] [7] [9] [0]. Based on Philips I-Code [23], Li et al. [0] presented a method o classiying the time slots based on their length: tag slots, long slots and short slots. he length o a tag slot, denoted as t id, is set to 2.4ms and allows the transmission o a -bit tag ID. he length o a long slot, denoted as t l, is set to 0.8ms and aords transmitting a long response containing 0-bit data. he length o a short slot, denoted as t s, is set to 0.4ms and allows the transmission o a short response conveying only -bit data. his gives an approximate transmission rate o /( ) = 40Kb/s [8]. IV. MULI-HASHING BASED MISSING AG IDENIFICAION PROOCOL Recall the state-o-the-art IIP scheme, empty and collision slots accounting or about 48% are wasted and thus trigger

4 4 its deiciency. o overcome this problem, a method that can increase the proportion o singleton slots is desirable. Inspired by [9] [20], this paper leverages multi-hashing idea to relieve the ineicient o the IIP scheme. But it is worth noting that this paper and literature [9] [20] address dierent problems. Moreover, the detailed implementation o our multihashing idea is dierent rom that in [9] [20]. Compared with the methods in [9] [20], the proposed MMI protocol does not suer the limitation o storage. In this section, we irst present the intuitive advantage o multi-hashing that inspires the proposed MMI protocol. Ater that, we give the detailed protocol design and investigate the involved parameter settings including the rame size as well as the hashing count m to maximize its perormance. A. Intuitive Motivation o Multi-Hashing he MMI protocol is proposed to reduce the expected empty slots and the expected collision slots. Fig. 2 illustrates the intuitive motivation o multi-hashing. Beore exempliying the basic multi-hashing idea, or the purpose o clarity, we deine two types o tags: () singleton tag that picks an expected singleton slot in the multi-hashing process; (2) collision tag that picks an expected collision slot. he process that each tag pseudo-randomly picks a slot rom a given slot set is reerred to as a hashing. As illustrated in Fig. 2 (a), ater the irst hashing, 4 singleton slots (marked by circle) can be used to identiy the presence o the corresponding 4 singleton tags (marked by circle). Whereas, 6 slots (empty or collision) are not used and wasted. Clearly, i the hashing is implemented or just once, the slotted rame is o low utilization. A multi-hashing method can improve the eiciency. As illustrated in Fig. 2(b), keeping the singleton mapping (derived rom the irst hashing) unchanged, we urther implement the second hashing between the 6 collision tags and the 6 nonsingleton (empty or collision) slots to improve the utilization o the slotted rame. In other words, the singleton tags will not participate in the second round o hashing. Moreover, only the non-singleton slots can be picked in the second hashing. Ater that, we can obtain 2 more singleton slots (marked by hexagon). Note that, the hashing process is just to get a virtual mapping between the slots and the tags. he slotted time rame has not been executed yet. As exempliied in Fig. 2, ater two hashing processes, 6 tags are assigned to exclusive slots. hat is, 6 slots are supposed to be singleton in the ollowing actual time rame. I we iteratively implement more hashing processes, more slots will become singleton. So intuitively, the multi-hashing based protocol can better utilize the time slots thereby relieving the deiciency o the IIP scheme. B. Protocol Design he proposed MMI protocol consists o multiple rounds, each o them consists o three phases: Pre-identiication phase, Identiication phase and Acknowledgment phase. In the Pre-identiication phase, the hashing process is iteratively implemented or m times. In an arbitrary hashing process, each non-singleton tag pseudo-randomly chooses a slot rom is used to mark the singleton map derived rom the irst hashing is used to mark the singleton map derived rom the second hashing singleton slot slot 0 slot slot 2 slot 3 slot 4 slot 5 slot 6 slot 7 slot 8 slot 9 ag ag 2 ag 3 ag 4 ag 5 ag6 ag7 ag8 ag9 ag0 singleton slot derived rom the irst hashing (a) he irst hashing singleton slot derived rom the second hashing slot 0 slot slot 2 slot 3 slot 4 slot 5 slot 6 slot 7 empty slot or collision slot slot 8 slot 9 ag ag 2 ag 3 ag 4 ag 5 ag6 ag7 ag8 ag9 ag0 (b) he second hashing Fig. 2: Multi (two)-hashing processes. the non-singleton slot set. For clarity o description, we irst describe the MMI protocol with m = 2 (i.e., in each round, the hashing process is iteratively implemented twice in the Pre-identiication phase). Ater the irst phase, each RFID tag determines a time slot within the ollowing rame and responds in the picked slot. Since the tags have ability to send a -bit response to the reader [9] [0] [] [24], or the purpose o saving time, the proposed MMI lets each tag respond only bit inormation to announce its presence in the picked slot. According to the actual state o an expected singleton slot, the reader identiies the presence o the corresponding tag. Speciically, i an expected singleton slot turns out to be an empty slot, the tag corresponding to this slot is missing. In the Acknowledgment phase, the reader transmits a bitmapack to inorm the tags i they have declared their presence successully in this round. According to the bitmapack, each tag determines to participate in the next round or not. In an arbitrary round, the details o three phases are presented as ollows. ) Pre-identiication phase: In the irst phase, the proposed MMI protocol aims to generate (m ) bitmaps on the reader side when m hashings are made. Generally, between any two consecutive hashing processes, each bitmap represents the slot occupation states o the current round o hashing, and is used to guide the slot selection o the next hashing. Ater the bitmaps are generated, they are transmitted one by one. hen the tags go through the (m ) bitmaps, one ater another, until they ind their slots to respond and the announce their presence. he detailed procedures are presented as ollows. he reader irst broadcasts a query R,, where R is a random number (resh in each round) and is the rame size. Each tag, say ag i, receives the query R, and uses the hash generator H( ), R, and its ID i to pseudo-randomly

5 5 pick slot x, where x = H(ID i, R) mod, whose result is within [0, ]. his is the irst hashing process. By employing the same H( ), R and, the reader can predict the locations o the expected empty slots, the expected singleton slots and the expected collision slots ater the irst hashing process. he reader constructs an -bits bitmap, in which s represent the expected singleton slots that cannot be selected in the next hashing process; 0s represent the expected empty slots or expected collision slots that can be selected in the next hashing. he reader counts the number o non-singleton (empty or collision) slots denoted as z, and broadcasts it together with a new random number R. Each tag receives z and R and computes j = H(ID i, R ) mod z, whose result is within [0, z ], where j means that the j th non-singleton slot is the candidate slot that ag i may pick in the second hashing process. he reader broadcasts the bitmap constructed above. Each tag, say ag i, receives this bitmap sequentially. ag i checks the x th bit in this bitmap and inds out the index (denoted as y) o the j th 0 in the bitmap as well. I ag i inds that the x th bit in the bitmap is, it learns that the slot x derived rom the irst hashing is an expected singleton slot, then it picks the slot x. Otherwise, ag i picks the slot y, which is equivalent to the second hashing. Because the reader knows all the parameters, tags-slots mapping results o the above two-hashing processes are predictable to the reader. 2) Identiication phase: In this phase, the MMI protocol actually executes the slotted rame. Similar with IIP [0], the proposed MMI leverages the observations o the expected singleton slots to identiy the missing tags. Speciically, i the reader receives a response in an expected singleton slot, it can assert the presence o the corresponding tag. On the other hand, i an expected singleton slot turns out to be empty, the corresponding tag must be missing and is pinpointed by the reader. o check i an expected singleton slots is empty or non-empty, -bit tag response is adequate. At the end o this phase, the reader constructs the - bits bitmapack. I slot k is expected to be a singleton slot, the reader sets bitmapack[k]=. Otherwise, the reader set bitmapack[k]= 0. 3) Acknowledgment phase: When all the slots in the rame have been counted, the reader broadcasts the bitmapack constructed above. Each tag receives the bitmapack sequentially and checks whether it has been veriied successully. Speciically, i a tag picked and responded in the slot k, it checks the k th bit in the bitmapack. I the k th bit is, the tag learns that its presence is noticed by the reader, then this tag keeps silent and will not participate in the ollowing rounds. Otherwise, this tag continues to participate in the next round. In this round, a raction o the tags are veriied and the other tags will participate in the next round. he MMI protocol repeats the round described above until all the presence o all tags is veriied. In the Pre-identiication phase and the Acknowledgement phase, a tag sequentially receives the bitmaps which may be very long. As there is no need or a tag to store the whole bitmap, the long bitmap is divided into segments o -bits or transmission in long slots. A segment o the bitmap becomes useless and can be erased ater being checked by a tag. Hence, a tag needs to store only one segment (-bits) at the same time. So the storage requirement is not an obstacle to the proposed MMI protocol. A random number R is picked in the above protocol design to perorm the hashing. It is worth noting that, because the reader knows all tag IDs, instead o randomly picking R, it is able to select an ideal R that can achieve a better mapping between the tags and the slots. hen, the perormance o the proposed MMI can be urther improved. But or a air comparison with the other two main benchmark protocols proposed in [8], [0], where R is also picked as a random number, this paper still randomly picks R in each process o hashing. C. Choosing an Optimal Frame Size ABLE II: Notations used in the ollowing proving process Symbols s s n n s 2 n 2 Descriptions he expected number o singleton slots derived rom the irst hashing he expected number o non-singleton slots derived rom the irst hashing he expected number o singleton tags derived rom the irst hashing he expected number o collision tags derived rom the irst hashing he expected number o new singleton slots derived rom the second hashing he expected number o new singleton tags derived rom the second hashing he total number o expected singleton slots derived rom two hashing processes, i.e., = s + s 2 he whole execution time o this round he MMI protocol repeats multiple rounds to identiy the presence o all tags. In an arbitrary round, let N denote the number o the tags that should participate in this round. Clearly, N = N in the irst round. In what ollows, we will present the detailed analysis o how to choose the optimal rame size in order to achieve the best time-eiciency in each round. able II summaries the notations used below. heorem : For the special case o m = 2 (i.e., two hashing processes are conducted in each round), MMI achieves the best time-eiciency i we set = N in each round. Proo: First, let us consider how many tags are expected to be singleton in this round. In the irst hashing process, given the rame size, each tag has the probability to select a speciic slot during the irst hashing process. For N tags in total, the probability p that a slot becomes expected singleton slot is given as ollows: ( ) N p = ( N ) N N () e = ρ e ρ Since is normally large, in Eq. (), p can be simpliied to N e. For the clarity o presentation, we denote N as N

6 6 ρ, and ρ is reerred to as the load actor meaning the number o tags loaded in the current rame. Each o the slots in the current rame has the probability p to be a singleton slot. Let s be the expected number o the singleton slots derived rom the irst hashing. We have: s = p = ρ e ρ = N e ρ (2) Let s denote the expected number o non-singleton slots derived rom the irst hashing process. Clearly, s can be written as: s = s = N e ρ (3) Let n and n denote the expected number o the singleton tags and the expected number o the collision tags, respectively. n and n can be given by: n = s = N e ρ (4) n = N n = N N e ρ (5) In the second hashing process, the n collision tags re-select slots rom the set o s non-singleton slots. he probability p 2 that a certain non-singleton slot becomes a singleton slot can be given as ollows: ( ) n p 2 = = n s ( ) ( n ) s s ( (6) ) n s Each o the s non-singleton slots has the probability p 2 to be a singleton slot in the second hashing process. Let s 2 denote the expected number o new singleton slots derived rom the second hashing process. According to Eq. (6), s 2 is given by: s 2 = s p 2 = n ( s ) n n e n s (7) According to Eqs. (3) and (5), by replacing s and n in Eq. (7), we have: s 2 (N N e ρ ) e N N e ρ N e ρ (8) he expected number,, o singleton slots derived rom two hashing processes is given as ollows: = s + s 2 N e ρ + (N N e ρ ) e N N e ρ (9) N e ρ hen let us consider how long it takes in this round. he size o the bitmap transmitted in the Pre-identiication phase and the bitmapack transmitted in the Acknowledgment phase may be very long. he reader divides them into segments o -bits (equivalent to the length o the tag ID) and transmits each segment in a tag slot with length o t id, i.e. 2.4ms. he length o each short slot in the Identiication phase is t s, i.e., 0.4ms. he time or transmitting the parameters, R and z is negligible and can be ignored. Hence, the execution time o this round is: = t id + t s + t id (0) = 2 t id + t s According to Eqs. (9) and (0), in this round, the average time or identiying the presence o a tag is: = 2 tid + t s N e ρ + (N N e ρ ) e N N e ρ N e ρ 2 t id + t s N e ρ + (N N e ρ ) e N N e ρ N e ρ 0.45 ρ e ρ + ρ ( e ρ ) e ρ ρ e ρ ρ e ρ () Following [0], we ind the optimal rom the average values, without considering the variance. Clearly, the average time or identiying a tag,, is a unction o ρ, where ρ > 0. o obtain the value o ρ that minimizes the, we dierentiate and set the result to zero as ollows: ( ) = where A(ρ) = 0.45(ρ ) A(ρ) = 0, (ρ e ρ + ρ ( e ρ ) e ρ ρ e ρ ρ e ρ ) 2 e ρ + ( + ρ2 )e ρ + ( + ρ)ρe 2ρ ρ 2 e 3ρ e ρ ρe ρ ( ρe ρ ) 2 (2) ρe ρ (3) In the Appendix-A, we prove that A(ρ) in Eq. (2) is always larger than 0. Hence, we have: ( ) > 0 when ρ > ; ( ) = 0 when ρ = ; ( ) < 0 when ρ <. hen it is proven that is minimized when ρ = (i.e., N = ). In other words, MMI (m = 2, i.e., two hashing processes in each round) achieves the best time-eiciency when = N. D. Choosing an Optimal Hashing Count m In the irst phase, the proposed MMI protocol employs a bitmap to guide the next hashing process. We have described the MMI protocol with the special case o m = 2. Intuitively, we can implement the hashing process or more times in each round to urther improve the eiciency. o achieve the MMI protocol with m = τ, where τ > 2, we just need to iteratively execute the hashing processes in the Pre-identiication phase. Can we repeat the hashing process until each tag chooses an expected singleton slot in the rame? In this case, the utilization o the rame is 00%. However, the transmission o a bitmap or guiding the next hashing is not cost-ree and generates overhead, and excessive hashing may beget ineiciency instead. In what ollows, we will present the analysis to optimize the hashing count m in Pre-identiication phase o each round. In the previous subsection, we have proved that, under the condition o m = 2, the MMI protocol achieves the best eiciency when = N in each round. Hence, we still set = N in each round o the MMI protocol with m > 2. heorem 2: Under the condition o = N in each round. I the hashing process is iteratively implemented or 5 times in

7 7 the Pre-identiication phase o each round, the MMI protocol achieves the best eiciency. Proo: We irst consider how many tags are expected to be veriied in this round. According to Eq. (2), we have: s = N e ρ = N e (4) Ater the irst hashing, we expect to obtain N e singleton slots. hen there are N ( e ) non-singleton (empty or collision) slots remaining. According to Eq. (8), and due to ρ=, we have: s 2 (N N e ρ ) e N N e ρ N e ρ (5) N ( e ) e hat is, ater the second hashing process, we expect to obtain N ( e ) e more singleton slots. here are N ( e ) 2 non-singleton slots remaining. We iteratively implement the hashing process or more times to achieve more singleton slots. Ater each hashing, e o the remaining non-singleton slots are expected to become singleton slots. Let s i denote the number o the new singleton slots derived rom the i th hashing process. s i is given by: s i = N ( e ) i e (6) In other words, we expect to obtain N ( e ) i e more singleton slots rom the i th hashing process. Let denote the total number o all singleton slots we expect to obtain ater implementing hashing or m times in this round. As the singleton slots are used to identiy the presence o the corresponding tags, the presence o tags can be veriied in this round. We have: m m = s i = N ( e ) i e (7) i= i= hen, let us consider the execution time o this round. In the Pre-identiication phase, in order not to disturb the achieved singleton slots in the next hashing, the reader sends a bitmap, which relects the current slot occupation states in time rame ( slots), to guide the next hashing. Hence, the size o the bitmap relecting the states o the whole time rame should always be bits. For m times hashing, the reader needs to broadcast bitmap ( bits) or m times. Each bitmap is divided into -bits segments, and each segment is transmitted in a long slot with length o t id, i.e., 2.4ms. here are short slots in the Identiication phase and the length o a short slot is t s, i.e. 0.4ms. In the Acknowledgment phase, the reader transmits the bitmapack, which is also divided into -bits segments and the transmission o each segment occupies a long slot t id. hus, the total time o this round, m, is denoted as: m = (m ) = m t id + t s + t id (8) t id + t s As a result, the average time or identiying the presence o a tag in this round is: m m tid + t s = m i=0 N ( e ) i e = m m i=0 N ( e ) i e = m N [ ( e ) m ] (9) Since we set to N in this round, hence, we get m m (20) ( e ) m According to Eq. (20), the average time m is a discrete unction o m, where m =, 2, 3,. We need to ind the optimal hashing count m that minimizes m N. We irst assume m is a continuous unction o m, where m > 0. he derivative o m is denoted as ( m ) and given as ollows: ( m ) = 2.4 [ ( e )m ] + ( 2.4 m + 0.4) ( e )m ln( ) e [ ( e )m ] 2 (2) By solving ( m ) = 0 in Eq. (2), we get m = 5.6. Moreover, in Appendix-B, we prove that ( m ) is always ). ) = 0 larger than 0, where ( m ) denotes the derivative o ( m hereore, we have ( m ) > 0 when m > 5.6; ( m when m = 5.6; ( m ) < 0 when m < 5.6. hat is, m = 5.6 is the minimum point o m. Actually, m is a discrete variable and cannot be set to 5.6. Hence, the adjacent two points m = 5 and m = 6 are the only two optimal candidates. m is ms when m = 5, and ms when m = 6. Clearly, m = 5 is the optimal setting, i.e., the hashing count m is optimized to 5 in each round. In heorem 2, it has been proven that 5 times hashing processes are optimal in each round. Intuitively, ewer than 5 hashings decrease the utilization o the time rame (i.e., the proportion o the singleton slots is low). On the other hand, more than 5 hashings increase the slot utilization but cause too much overhead or transmitting the bitmaps, and the added overhead outweighs the gain o slot utilization. he simulation results shown in able III also validate the above analysis. E. Impact on Channel Errors he paper irst assumes that there is a perect communication channel between the reader and tags. However, in the real environment, the communication channel is error-prone [0] [2] [20] [25] [26]. he white noise may corrupt the data exchanged between the reader and tags, e.g., 0 becomes or becomes 0 [0]. he signals are even not detected at all due to path loss. In most literature [2] [20] [26] that consider the channel error, CRC (Cyclic Redundancy Code) is used to veriy the correctness o the exchanged data between the reader and tags. Since most previous literature assumes the reader has adequate power to interrogate all tags in its vicinity [20] [22], this paper assumes that the signals on the readerto-tags link will not be lost. In contrary, the signals on the tags-to-reader link may be lost because o the weak capability o tags. In what ollows, we discuss the various channel errors in detail. ) Channel Errors during ransmission o the Parameters R, and z: he correct transmission o the parameters R, and z is crucial to the proposed MMI protocol. o ensure the correct transmission o those parameters, the reader can

8 8 generate a 0-bit checksum [26] o these parameters and appends the checksum to the back o the transmitted parameters. he tags check the correctness o the checksum ater receiving the < parameters, checksum >. I the received parameters and the received checksum do not match, the tag will reply an NACK signal so as to orce the reader to retransmit the < parameters, checksum >. I the reader senses a strong pulse (may be a single NACK or a mix o multiple NACKs), it will retransmit this binary group until it is transmitted correctly. 2) Channel Errors during ransmission o the Bitmaps: As the bitmaps are used to guide the hashing processes, the correct transmission o the bitmaps is very crucial. o tackle the channel errors during transmission o the bitmaps, we can also use the method o appending a checksum. Speciically, we divide a bitmap into segments o 75-bit, each o which is given a segment number ( bits), e.g., the segment number o the 0 th segment is (equal to the decimal value 0 ). Later on, we will explain the unction o the segment number. he reader generates a 0-bit checksum o < segment number, segment > and appends the checksum to the back so as to generate a triple, i.e., < segment number, segment, checksum >. he reader transmits the triples to the tags one by one. Similarly, the tags check the correctness o the received triple by veriying the checksum. For a certain triple, some tags can correctly receive this triple, whereas, some other tags cannot correctly receive it and reply NACK signals to orce the reader to retransmit the same triple. Note that, the retransmitted triples do not conuse the tags that have already correctly received the same triple. I a tag successively receives two or more triples with the same segment number, they are treated as the same one. A triple is successully transmitted only when no NACK signal is replied. hen the next triple is transmitted. 3) Channel Errors when the ags Reply Responses: Clearly, based on the above countermeasures, we can guarantee that the inal virtual mapping between tags and slots is correct and is predictable by the reader. We do not care the channel errors in the expected empty slots and the errors in the expected collision slots, because the proposed MMI protocol only leverages the expected singleton slots to identiy the missing tags. Hence, we only discuss the impact o channel errors occurring in the expected singleton slots. I an expected singleton slot corresponds to a present tag, but due to the channel errors (e.g., path loss), its response is not sensed by the reader, then it is wrongly considered as a missing tag, namely the alse positive. On the other hand, i an expected singleton slot corresponds to an actually missing tag, due to the channel errors (e.g., white noise), this missing tag is wrongly veriied as a present one, namely the alse negative. An eicient method to tackle the alse positive (i.e., a present tag is considered as a missing one) is presented below. he reader requests the reported missing tags by their IDs one by one, then the ake missing tags can be iltered out. As or the alse negative (i.e., a missing tag is veriied as a present one), even i a missing tag is not detected in the current execution, it will be detected with a high probability in the next round o execution, because the missing tag identiication is usually periodically executed in reality. A. Simulation Setting V. PERFORMANCE EVALUAION In this section, we evaluate the perormance o the proposed MMI protocol. he execution time or identiying all the missing tags is used as the perormance criterion. Since IIP [0] and Protocol 3 [8] are the main benchmarks in this paper, we adopt the same parameter settings in [0] [8]. Speciically, transmission o a segment o bits takes a tag slot (i.e., t id = 2.4ms); transmission o bit response to the reader takes a short slot (i.e., t s = 0.4ms). We run each simulation or 000 times and collect the average results. B. Validating the Optimal Hashing Count m he irst simulation set in able III intends to investigate the impact o dierent hashing count m on the perormance o the proposed MMI protocol, where we simulate a single reader. We vary the hashing count m rom 2 to 6 in the irst simulation set. As shown in able III, the proposed MMI protocol achieves the best time-eiciency when the hashing count m = 5, which coincides with the proo presented in Section IV-D. ABLE III: he execution time o MMI with dierent hashing count m. N Execution time (s) m=2 m=3 m=4 m=5 m= C. Execution ime In this subsection, we mainly compare the MMI protocol with the state-o-the-art protocols, Iterative ID-ree Protocol (IIP) in [0] and the protocols in [8]. As Protocol 3 is the best one in [8], we only select Protocol 3 as one o the benchmarks in this paper. he solutions to tag identiication problem can also solve the problem o identiying the missing tags, by comparing the collected ID inormation with the original ID inormation stored in database. Hence, we also compare the MMI protocol with the most outstanding tag identiication protocols, the Enhanced Dynamic Framed Slotted ALOHA (EDFSA) [6] and the Binary ree Protocol (BP) [7]. In [0], the authors presented a lower bound or missing tag identiication protocol or the single reader case. Speciically, N tags respond -bit announcement to the reader one by one. I we assume no coordination inormation is transmitted, the execution time is N t s, which is treated as the lower bound or any missing tag identiication protocols. In this paper, we also compare the proposed protocol with this lower bound. As shown in able IV, extensive simulations are conducted to

9 9 ABLE IV: he Execution ime With Respect to ag Number N. N Execution time (s) BP EDFSA Protocol 3 in [8] IIP MMI (m=5) Lower Bound evaluate the eiciency o the MMI protocol, where a single reader is simulated. he tag number N varies rom 5,000 to 50,000. he simulation results in able IV demonstrate that the proposed MMI protocol perorms much better than the BP protocol and the EDFSA protocol. For example, when N is 50,000, the execution time o BP and EDFSA is seconds and seconds, respectively. And the MMI protocol is seconds, which outperorms the BP and EDFSA by 92.78% and 92.47%, respectively. Moreover, the MMI protocol cuts the execution time by about 32% when compared to the IIP protocol. For example, when N is 50,000, the execution time o the proposed MMI protocol is 32.3% less than that o IIP. Because Protocol 3 does not possess superiority in the single reader scenarios, it is not surprising that Protocol 3 perorms the worst in this simulation set. In multi-reader scenarios, Protocol 3 will show its superiority. But due to the space limitation, we do not present the simulation results o multi-reader case. Average time or veriying a tag in milliseconds heoretical value o IIP Simulation result o IIP heoretical value o MMI Simulation result o MMI he number N o tags in the RFID system Fig. 3: Comparison with the IIP protocol in a small scale RFID system. D. he Impact o the ag Number According to the simulation results shown in able IV, we can ind that the average time taken by the proposed MMI protocol to veriy a tag is approximately the same as the theoretical value obtained in Section IV-D, 0.58 ms per tag in a large scale system (i.e. with tens o thousands o tags). However, when the system is in a small scale, i.e., containing just hundreds o tags, the average time or veriying a tag in simulations is slightly higher than the theoretical value, as illustrated in Fig. 3. Why does this phenomenon occur? he reason is briely exempliied as ollows. Assume there are just 50 tags in the small scale MMI system. According to the above description o the MMI protocol, the bitmap will be 50 bits (shorter than bits) in the irst round, but still occupies a tag slot, 2.4 ms. his overhead becomes notable, when N (i.e. the number o tags in the system) is small. Hence, the average time taken or veriying a tag becomes slightly higher than the theoretical value. On the other hand, in a large scale system, this overhead still exists. But, due to the large number o tags, this small overhead is shared by a large number, N, o tags, then this overhead can be ignored. he simulation results approach the theoretical value as the system scales up. As illustrated in Fig. 3, the MMI protocol still outperorms the IIP protocol even in a small-scale RFID system. E. he Impact o Channel Error Since the communication channel is not always perect in reality, the channel errors (e.g., white noise, path loss, etc.) oten degrade the perormance o MMI or even give rise to the alse o the identiication results. Hence, it is necessary to evaluate the perormance o the proposed MMI protocol when the channel is not error-ree. Let P error denote the probability that each bit o the transmitted parameters or the bitmaps becomes wrong. I the channel error occurs, the retransmission is required. ) he Impact on the otal Execution ime: In this set o simulation experiments, we investigate the impact o the channel errors on the total execution time o the proposed MMI protocol under dierent channel conditions, where P error varies rom 0.% to %. he simulation results depicted in Fig. 4 reveal that the MMI protocol consumes more time in the scenarios with channel errors than in the perect scenarios. Given a perect communication channel, the MMI with CRC consumes more time than the pure MMI because the transmission o CRC generates extra overhead. 2) he Impact on the Identiication Accuracy: I the communication channel is error-ree, the proposed MMI can completely identiy all the missing tags without any alse. However, as aorementioned, the channel errors lead to alse negative (i.e., missing tags are veriied as present ones) and alse positive (i.e., present tags are reported as missing ones) o identiication. We use the ratio o alse negative and the ratio o alse positive to evaluate the identiication accuracy o the MMI protocol. he simulation results shown in Figs. 5 (a) and (b) demonstrate that both the ratio o alse negative and the ratio o alse positive luctuate around the bit error probability

10 0 he total execution time in seconds bit error rate % bit error rate 0.5% bit error rate 0.% perect channel, with CRC perect channel, without CRC 0 5,000 0,000 5,000 20,000 25,000 30,000 he number N o all tags Fig. 4: Evaluating the perormance o the MMI protocol under dierent channel conditions. P error, which is normally quite small. As mentioned in Section IV-E3, the relatively small number o alse positive tags can be reveriied by a sample polling method. Moreover, the small number o alse negative tags will be discovered with a high probability in the next execution [0]. he ratio o alse negative % % bit error rate % bit error rate 0.5%.25 bit error rate 0.% he number M o missing tags he ratio o alse positive bit error rate % bit error rate 0.5% bit error rate 0.% 0 5,000 0,000 5,000 20,000 25,000 he number N o all tags Fig. 5: Evaluating the identiication accuracy o the MMI protocol under dierent channel conditions. (a) he ratio o alse negative; (b) the ratio o alse positive. VI. CONCLUSIONS his paper has studied a practically important problem o missing tag identiication, which is very popular in many application scenarios such as warehouse management, supply chain management and prison monitoring. For relieving the deiciency o the state-o-the-art protocols, this paper has investigated the multi-hashing idea to improve the utilization o the time rame and thus proposed a Multi-hashing based Missing ag Identiication (MMI) protocol. We have urther investigated the optimal setting o the parameters and m to maximize the perormance o the proposed MMI protocol. he impact o channel errors on the proposed MMI protocol and the corresponding countermeasures are discussed. Extensive simulation experiments have been conducted to evaluate the eiciency o the proposed MMI protocol. he results maniest that this protocol considerably outperorms the stateo-the-art protocols in terms o time-eiciency. APPENDIX A. Proving A(ρ) > 0: Proo Objective: A(ρ) in Eq. (2) is always larger than 0, where A(ρ) is given as ollows: e ρ + ( + ρ2 )e ρ + ( + ρ)ρe 2ρ ρ 2 e 3ρ e ρ ρe ρ ( ρe ρ ) 2 ρe ρ (22) Proo: For clarity, we denote ( + ρ 2 )e ρ + ( + ρ)ρe 2ρ ρ 2 e 3ρ as B(ρ); and denote ρe ρ as C(ρ), then we have: A(ρ) = e ρ + B(ρ) ρ ρe ρ C(ρ) 2 e C(ρ) (23) Clearly, i we prove that B(ρ) > 0 and C(ρ) 0, then A(ρ) > 0 is proved. Step : Proving B(ρ) > 0: B(ρ) = ( + ρ 2 )e ρ + ( + ρ)ρe 2ρ ρ 2 e 3ρ > ( + ρ 2 )e ρ + ρ 2 e 2ρ ρ 2 e 3ρ = ( + ρ 2 )e ρ + ρ 2 (e 2ρ e 3ρ ) > ( + ρ 2 )e ρ }{{} denoted as D(ρ) (24) Let ( + ρ 2 )e ρ be denoted as D(ρ), and we get its derivative as ollows: D(ρ) = (ρ ) 2 e ρ 0 (25) As shown in Eq. (25), the derivative o D(ρ) is always larger than (or equal to) 0, thence, D(ρ) is a increasing unction with respect to ρ. hereore, we have D(ρ) > D(0), where D(0) = 0, i.e., D(ρ) > 0. According to Eq. (24), we have B(ρ) > D(ρ) > 0. Step 2: Proving C(ρ) 0: We get the derivative o C(ρ) = ρe ρ as ollows: C(ρ) = (ρ )e ρ, (26) where we have C(ρ) > 0 when ρ > ; C(ρ) = 0 when ρ = ; C(ρ) < 0 when ρ <. Hence, C(ρ) is minimized to C(), where C() = e. Clearly, C(ρ) 0. Since we have proved B(ρ) > 0 and C(ρ) 0, it is easy to know that A(ρ) in Eq. (23) is larger than 0. B. Proving ( m ) > 0: Proo Objective: he derivative ( m ) o ( m ) is always larger than 0, where ( m ) is given as ollows: ( m ) = 2.4 [ ( e )m ] + ( 2.4 m + 0.4) ( e )m ln( e ) [ ( e )m ] 2 (27) Proo: Also or clarity, we denote the 2.4 [ ( e )m ] + ( 2.4 m + 0.4)( e )m ln( e ) as F (ρ); and denote [ ( e )m ] 2 as G(ρ), then we have: ( m ) = F (ρ) G(ρ) (28)

11 he derivative o ( m ) is denoted as ( m ) and given as ollows: ( m ) = F (ρ) G(ρ) = F (ρ) G(ρ) F (ρ)g(ρ) G(ρ) 2 (29) Obviously, G(ρ) is not equal to 0, hence, i we prove that F (ρ) G(ρ) F (ρ)g(ρ) is larger than 0, the inal proving objective ( m ) > 0 is proved. In what ollows, we prove that F (ρ) G(ρ) F (ρ)g(ρ) > 0. F (ρ) G(ρ) F (ρ)g(ρ) =( 2.4 m + 0.4) ln2 ( e ) ( e )m {[ ( e )m ] 2 + 2( e )m [ ( e )m ]} ln( e ) ( e )m [ ( e )m ] 2 >( 2.4 m + 0.4) ln2 ( e ) ( e )m [ ( e )m ] ( e )m [ ( e )m ] 2 >( e )m [ ( e )m ] 2 [ 2.4 ln2 ( )m ] e >0 (30) According to Eq. 29 and Eq. 30, we prove that ( m ) is larger than 0. ACKNOWLEDGMEN he authors would also like to sincerely thank the editors and anonymous reviewers or their thoughtul suggestions and constructive comments, which have greatly helped and inspired us to improve the quality o this paper. his work was supported by NSFC (Grant No.s , 67360, 67362, , and 63249), New Century Excellent alents in University (NCE) o Ministry o Education o China, the National Science Foundation or Distinguished Young Scholars o China (Grant No ), and the Project unded by China Postdoctoral Science Foundation. REFERENCES [] L. M. Ni, Y. Liu, Y. C. Lau, and A. P. Patil, Landmarc: Indoor Location Sensing Using Active RFID, Wireless networks, vol. 0, no. 6, pp , [2] L. Xie, B. Sheng, C. C. an, H. Han, Q. Li, and D. Chen, Eicient ag Identiication in Mobile RFID Systems, Proc. o IEEE INFOCOM, 200. [3] M. Kodialam,. Nandagopal, and W. C. Lau, Anonymous racking using RFID tags, Proc. o IEEE INFOCOM, [4] B. Sheng, Q. Li, and W. Mao, Eicient Continuous Scanning in RFID systems, Proc. o IEEE INFOCOM, 200. [5] K. Bu, B. Xiao, Q. Xiao, and S. Chen, Eicient Pinpointing o Misplaced ags in large RFID Systems, Proc. o IEEE SECON, 20. [6] C. C. an, B. Sheng, and Q. Li, How to Monitor or Missing RFID tags, Proc. o IEEE ICDCS, [7] W. Luo, S. Chen,. Li, and S. Chen, Eicient Missing ag Detection in RFID Systems, Proc. o IEEE INFOCOM, 20. [8] R. Zhang, Y. Liu, Y. Zhang, and J. Sun, Fast Identiication o the Missing ags in a Large RFID System, Proc. o IEEE SECON, 20. [9] W. Luo, S. Chen,. Li, and Y. Qiao, Probabilistic Missing-tag Detection and Energy-ime radeo in Large-scale RFID Systems, Proc. o ACM MobiHoc, 202. [0]. Li, S. Chen, and Y. Ling, Identiying the Missing ags in a Large RFID System, Proc. o ACM MobiHoc, 200. [] J. Wang, H. Hassanieh, D. Katabi, and P. Indyk, Eicient and Reliable Low-Power Backscatter Networks, Proc. o ACM SIGCOMM, 202. [2] M. Zhang,. Li, S. Chen, and B. Li, Using Analog Network Coding to Improve the RFID Reading hroughput, Proc. o IEEE ICDCS, 200. [3] L. Kang, K. Wu, J. Zhang, and H. an, Decoding the Collisions in RFID Systems, Proc. o IEEE INFOCOM, 20. [4] J. I. Capetenakis, ree Algorithms or Packet Broadcast Channels, IEEE ransactions on Inormation heory, vol. I-25, no. 5, pp , 979. [5] L. G. Roberts, Aloha Packet System with and without Slots and capture, ACM SIGCOMM Computer Communication Review, vol. 5, no. 2, pp , 975. [6] S. Lee, S. Joo, and C. Lee, An Enhanced Dynamic Framed Slotted ALOHA Algorithm or RFID ag Identiication, Proc. o IEEE MobiQuitous, [7] J. Myung and W. Lee, Adaptive Splitting Protocols or RFID ag Collision Arbitration, Proc. o ACM MobiHoc, [8] N. Bhandari, A. Sahoo, and S. Iyer, Intelligent Query ree (IQ) Protocol to Improve RFID ag Read Eiciency, Proc. o IEEE ICI, [9] V. Namboodiri and L. Gao, Energy-Aware ag Anti-Collision Protocols or RFID Systems, Proc. o IEEE PerCom, [20] S. Chen, M. Zhang, and B. Xiao, Eicient Inormation Collection Protocols or Sensor-augmented RFID Networks, Proc. o IEEE IN- FOCOM, 20. [2] R. Jacobsen, K. F. Nielsen, P. Popovski, and. Larsen, Reliable Identiication o RFID ags Using Multiple Independent Reader Sessions, Proc. o IEEE RFID, [22] M. Shahzad and A. X. Liu, Every Bit Counts - Fast and Scalable RFID Estimation, Proc. o ACM MobiCom, 202. [23] P. Semiconductors, I-CODE Smart Label RFID ags, download/other/identiication/sl pd, Jan [24] Y. Zheng and M. Li, Fast ag Searching Protocol or Large-Scale RFID Systems, Proc. o IEEE ICNP, 20. [25] Y.-C. Lai and C.-C. Lin, wo blocking algorithms on adaptive binary splitting: single and pair resolutions or rid tag identiication, IEEE/ACM ransactions on Networking (ON), vol. 7, no. 3, pp , [26] Y. Qiao, S. Chen,. Li, and S. Chen, Energy-eicient Polling Protocols in RFID Systems, Proc. o ACM MobiHoc, 20. Xiulong Liu received the B.E. degree rom the School o Sotware echnology, Dalian University o echnology, China, in 200. Currently, he is a Ph.D. candidate in the School o Computer Science and echnology, Dalian University o echnology, China. His research interests include RFID systems and wireless sensor networks. Keqiu Li received the bachelor s and master s degrees rom the Department o Applied Mathematics at the Dalian University o echnology in 994 and 997, respectively. He received the Ph.D. degree rom the Graduate School o Inormation Science, Japan Advanced Institute o Science and echnology in He also has two-year postdoctoral experience in the University o okyo, Japan. He is currently a proessor in the School o Computer Science and echnology, Dalian University o echnology, China. He has published more than 00 technical papers, such as IEEE PDS, ACM OI, and ACM OMCCAP. He is an Associate Editor o IEEE PDS and IEEE C. He is a senior member o IEEE. His research interests include internet technology, data center networks, cloud computing and wireless networks.

12 2 Geyong Min is a Proessor o High Perormance Computing and Networking in the Department o Mathematics and Computer Science within the College o Engineering, Mathematics and Physical Sciences at the University o Exeter, United Kingdom. He received the PhD degree in Computing Science rom the University o Glasgow, United Kingdom, in 2003, and the B.Sc. degree in Computer Science rom Huazhong University o Science and echnology, China, in 995. His research interests include Next Generation Internet, Wireless Communications, Multimedia Systems, Inormation Security, Ubiquitous Computing, Modelling and Perormance Engineering. Alex X. Liu received his Ph.D. degree in computer science rom the University o exas at Austin in He is an Associate Proessor with the Department o Computer Science and Engineering, Michigan State University. He is an Associate Editor o IEEE/ACM ransactions on Networking and an Area Editor o Elsevier Journal o Computer Communications. He received the IEEE & IFIP William C. Carter Award in 2004 and an NSF CAREER award in He received the Withrow Distinguished Scholar Award in 20 at Michigan State University. He received Best Paper Awards rom ICNP-202, SRDS-202, and LISA-200. His research interests ocus on networking and security. Yanming Shen received the B.S degree in automation rom singhua University, Beijing, China, in 2000, and the Ph.D. degree rom the Department o Electrical and Computer Engineering at the Polytechnic Institute o New York University, Brooklyn, in He is a Proessor with the School o Computer Science and echnology, Dalian University o echnology, DaLian, China. His general research interests include packet switch design, data center networks, peer-to-peer video streaming, and algorithm design, analysis, and optimization. He is a recipient o the 20 Best Paper Awards or Multimedia Communications (awarded by IEEE Communications Society). Weyu Qu received the bachelor s and master s degrees rom Dalian University o echnology, China in 994 and 997, and the doctorate degree rom Japan Advanced Institute o Science and echnology in She is a proessor at the School o Inormation and echnology, Dalian Maritime University, China. She was a lecturer in Dalian University o echnology rom 997 to Her research interests include mobile agent-based technology, distributed computing, computer networks, and grid computing. She has published more than 50 technical papers in international journals and conerences. She is on the committee board or a couple o international conerences.