CatchIt: Detect Malicious Nodes in Collaborative Spectrum Sensing

Transcription

1 CatchIt: Detect Malicious Nodes in Collaborative Spectrum Sensing Wenkai Wang, Husheng Li, Yan (Lindsay) Sun, and Zhu Han Department of Electrical, Computer and Biomedical Engineering University of Rhode Island, Kingston, RI 0288 Department of Electrical Engineering and Computer Science University of Tennessee, Knoxville, TN Department of Electrical and Computer Engineering University of Houston, Houston, TX Abstract Collaborative spectrum sensing in cognitive radio networks has been proposed as an efficient way to improve the performance of primary users detection. In collaborative spectrum sensing schemes, secondary users are often assumed to be trustworthy. In practice, however, cognitive radio nodes can be compromised. Compromised secondary users can report false detection results and significantly degrade the performance of spectrum sensing. In this paper, we investigate the case that there are multiple malicious users in cognitive radio networks and the exact number of malicious users is unknown. An onion-peeling approach is proposed to defense against multiple untrustworthy secondary nodes. We calculate suspicious level of all nodes according to their reports. When the suspicious level of a node is beyond certain threshold, it will be considered as malicious and its report will be excluded in decision-making. We continue to calculate the suspicious level of remaining nodes until no malicious node can be found. Simulation results show that malicious nodes greatly degrade the performance of collaborative sensing, and the proposed scheme can efficiently detect malicious nodes. Compared with existing defense methods, the proposed scheme significantly improves the performance of primary user detection, measured by ROC curves, and captures the dynamic change in the behaviors of malicious users. Keywords: Collaborative Spectrum Sensing, Security. I. INTRODUCTION The increasing spectrum demand for new wireless applications leads to the idea of cognitive radio []. It allows secondary users to communicate in licensed spectrum provided that they do not cause harmful interference to primary users. A major challenge in cognitive radio networks is to detect available spectrum resources in given time and location without changing the existing system framework of primary users. There are several detection schemes, including matched filter, energy detection, cyclostationary detection and wavelet detection [], [2]. Among them, energy detection is a common method for detection of unknown signals because it does not depend on the structure of primary signal [3], [4]. The performance of spectrum sensing degrades significantly when wireless channel experiences fading or shadowing. To deal with this problem, collaborative spectrum sensing which combines sensing results of multiple nodes is proposed to This work is supported by NSF Award # and # achieve better performance. Roughly speaking, there are two types of collaborative sensing techniques. In the first type, multiple secondary users send sensing reports directly to a common decision center [3] [8]. In the second type, some secondary users relay sensing reports for other secondary users and the decision can be made in a distributed manner [9], [0]. Most current collaborative sensing schemes assume that all secondary users are honest. However, the attacker can compromise or control some cognitive radio nodes, and the compromised secondary users can report dishonest sensing information and degrade sensing performance. There are very limited works that address the problem of malicious secondary users. A straightforward defense is proposed in [] for harddecision case, in which secondary users report yes (i.e. primary user presents) or no (i.e. primary user absents). When there are k malicious nodes, the decision center decides yes only if there are at least k + nodes reporting yes. This simple defense scheme has several disadvantages. First, it is hard to accurately estimate the number of malicious nodes (i.e. k value). Second, the smart attackers may not report yes all the time. This scheme, which is very conservative, can greatly scarify the performance of collaborative sensing. Third, this scheme cannot be directly applied to the scenario when the nodes report sensed energy instead of hard decisions. An attack called Spectrum Sensing Data Falsification (SSDF) was discussed in [2]. The scheme deals with dishonest users in distributed spectrum sensing. It requires the secondary users to collect more reports from neighbors when the current reports are not sufficient for making confirmative decision. In [3], the credibility of secondary users is taken into consideration in primary user detection. The credibility scores are based on each secondary user s detection rate and false alarm rate, but do not consider potential malicious cheating behaviors. A scheme proposed in [4] addresses the compromised cognitive nodes sending false information in centralized spectrum sensing. The proposed scheme can differentiate honest and malicious nodes when there is only one malicious node. However, the scheme cannot be applied to multiple malicious users scenario. In this paper, we propose an onion-peeling approach to defend against multiple malicious nodes. We calculate suspicious level of all secondary users according to their reports.

2 2 Fig.. Secondary user Secondary user Primary user Decision Center Task : Malicious secondary user? Task 2: Primary user existing? Collaborative Spectrum Sensing Secondary user When the suspicious level of a node is beyond certain threshold, it will be regarded as malicious and its report will be excluded in decision-making. Then we continue to calculate the suspicious level of remaining nodes until no malicious nodes can be found. We compare the performance of the collaborative sensing scheme without security protection, the scheme using a straightforward existing defense method [], and the proposed scheme under various scenarios. Simulation results demonstrate that multiple attackers can significantly degrade the performance of collaborative sensing when there is no defense. The straightforward defense scheme alleviates the problem but still has high false alarm rate. The proposed scheme can efficiently detect malicious users and improve the sensing performance. For example, when there are 0 secondary users, with the primary user detection rate being, two malicious users can make the false alarm rate (P f ) increase to 85%, the straightforward defense method can reduce P f to 23%, while proposed scheme can decrease P f to 8%. Furthermore, when a good user suddenly turns bad, the proposed scheme can quickly increase the suspicious level of this user. If this user only behaves badly for a few times, its suspicious level can recover after a large number of good behaviors. As a summary, the proposed scheme has two major advantages. First, it is self-adaptive and does not need to know the number of attackers beforehand. Second, it has better performance compared to the existing defense scheme that even has the knowledge of maximum number of attackers. The performance is measured by the ROC (Receiver Operating Characteristic) curve of primary user detection. This paper is organized as follows. System model is given in Section II. In Section III, the attack models and the proposed scheme are explained. Simulation results are shown in Section IV and conclusion is drawn in Section V. II. SYSTEM MODEL In this section, we briefly describe the scenario of collaborative spectrum sensing and then propose two attack models. A. Collaborative Spectrum Sensing The objective of spectrum sensing is to detect whether the primary user is transmitting or not, based on the reports from the secondary users. Let H 0 stand for the absence of primary user and H for the presence of primary user. The goal is to decide the hypotheses test: { ni, H y i 0, () h i s + n i, H, where y i is the received signal at the i-th secondary user, n i is the thermal noise, h i is the channel gain from the primary transmitter to the i-th secondary user, and s is the transmitted symbol from the primary user. Suppose Y i is the sensed energy for the i-th cognitive user in T time slots, the distribution of Y i is χ 2 distribution as: { χ 2 Y i 2T W, H 0, χ 2 2T W (2γ (2) i), H, where γ i is the received signal to noise ratio, and T W is time-bandwidth product. From (2) it can be seen that under H 0, given T W and y i, the probability P (Y i y i H 0 ) is known when the channel statistics are known. Similarly, under H, given T W and y i, the probability P (Y i y i H ) is also known. Both the conditional probability can be calculated based on path loss model and the location information. However, results in [3], [4] show that the performance of spectrum sensing degrades significantly when channel experiences fading or shadowing. In this case, collaborative sensing which combines detection results from several secondary users is proposed to achieve better performance. This scenario, called collaborative spectrum sensing, is shown in Figure. Here, N collaborative secondary users send reports to the decision center and then decision center makes the final decision. B. Attack Model When a secondary user is malicious, it can cause false alarm by reporting yes or high energy level when the primary user absents. This is the major security concern. In addition, the attacker can increase the miss detection rate if it reports no or low energy level when the primary user presents. To demonstrate the performance of the proposed defense scheme, in this paper, we adopt the two attack models, FA and FAMD, presented in [4]. The attacks are modeled by three parameters: the attack threshold (η), attack strength ( ), and attack probability (P a ) For a continuous spectrum sensing process, let X n (t) denote the observation of node n about the existence of the primary user at time slot t. The two attack models are, False Alarm (FA) Attack: In each time slot, if the sensed energy X n (t) is higher than η, the attacker reports X n (t); otherwise, it will attack with probability P a. This means it does not necessarily attack all the time. It can randomly choose to attack or not in one round, with probability P a. If it chooses to attack, it will report X n (t) +. False Alarm & Miss Detection (FAMD) Attack: In each time slot, attacker chooses to attack or not with attack probability P a. If it doesn t attack, it will just report what it has. Otherwise, it will compare the sensed energy X n (t) with η. If it is higher than η, the attacker reports X n (t) ; otherwise, it reports X n (t) +.

3 3 With the FA attack, the malicious secondary user aims to cause high false alarm. While in the FAMD attack, the malicious user can cause both false alarm and miss detection. For the defense scheme proposed in this paper, we assume that the attack type, either FA or FAMD, is known. III. SECURE COLLABORATIVE SENSING To defend collaborative spectrum sensing against dishonest secondary users, we first consider an ideal malicious node detection scheme. The scheme enumerates all possible malicious nodes set and tries to identify the set with the largest possibility of being malicious. However, this method faces the curse of dimensionality. Then, a heuristic scheme called Onion-peeling approach is proposed. A. Ideal Malicious Node Detection We use S(t) to denote channel state which is either B (Busy) or I (Idle) for each sensing time slot. Channel states in different time slots are assumed to be independent of each other. Let q B (t) and q I (t) denote the a priori probabilities of states B and I for time slot t, respectively. Recall that X j (t) is the observation of node j at time slot t. Let p B and p I denote the observation probabilities of X j (t) under busy and idle states, respectively, i.e. p I (X j (t)) P (X j (t) S(t) I), (3) p B (X j (t)) P (X j (t) S(t) B). (4) Let T n be the type of node n, which could be H (honest) or M (Malicious), and F t be all observations from time slot to time slot t. For any Ω {,..., N} (note that Ω could be an empty set, i.e. there is no attacker), we define π Ω (t) P (T n M, n Ω, T m H, m / Ω F t ), (5) as the belief that all nodes in Ω are malicious nodes while all other nodes are honest. Given any particular set of malicious nodes Θ, by applying Bayesian criterion, we have π Ω (t) P (F t Ω)P (Ω) Θ P (F t Θ)P (Θ). (6) Suppose that P (T n M) ρ for all nodes. Then, we have P (Ω) ρ Ω ( ρ) N Ω, (7) where Ω is the cardinality of Ω. Next, we can calculate where P (F t Ω) τ P (X j (τ) F, F τ ) P (X j (τ) T j H) j / Ω j Ω ρ n (τ), (8) τ ρ n (t) j / Ω P (X j (τ) T j H) j Ω P (X j (τ) F, F τ ). (9) For each possible choice of Ω, using equation (6)-(9), we can calculate the probability that this Ω contains all malicious users but no honest users. Choose the Ω(t) with largest π Ω (t) value. Then we can define a threshold, when π Ω (t) of the chosen Ω(t) is beyond this threshold, the nodes in Ω are considered to be malicious. However, for a cognitive radio network with N secondary users, there are 2 N different choices of set Ω. Thus, the complexity grows exponentially with N. So this ideal detection of attackers faces the curse of dimensionality. When N is large, we have to use approximate and heuristic approach. B. Onion-Peeling Approach In this paper, we propose a heuristic onion-peeling approach that detects the malicious user set in a batch-by-batch way. Let suspicious level describe the possibility that a node is malicious. We calculate suspicious level of all users according to their reports. When the suspicious level of a node is beyond certain threshold, it will be considered as malicious and its report will be excluded in primary user detection in this round. Then we continue to calculate the suspicious level of remaining nodes until no malicious node can be found. First we initialize the set of malicious nodes, Ω, as an empty set. In the first stage, compute the a posteriori probability of attacker for any node n, which is given by π n (t) P (T n M F t ) P (F t T n M)P (T n M) P (F t T n M)P (T n M) + P (F t T n H)P (T n H) (0) where we assume that all other nodes are honest when computing P (F t T n M) and P (F t T n H). Note here we only calculate the suspicious level for each node rather than that of a malicious nodes set, the computation complexity is reduced from O(2 N ) to O(N). Let X(t) denote the collection of X n (t), i.e. reports from all secondary users at time slot t. It is easy to verify where P (F t T n M) P (X(τ) T n M, F τ ) τ τ j,j n P (X j (τ) T j H) P (X n (τ) F τ ) ρ n (τ), () τ ρ n (t) P (X n (t) F τ ) j,j n P (X j (t) T j H), (2) Here, P (F t T n M) means the probability of reports at time slot t conditioned that node n is malicious. Note that the first equation is obtained by repeatedly applying the following equation P (F t T n M) P (X(t) T n M, F t )P (F t T n M). (3)

4 4 Similarly, where φ n (t) P (F t T n H) P (X(τ) T n H, F τ ) τ P (X j (τ) T j H) τ j φ n (τ), (4) τ P (X j (t) T j H). (5) j As mentioned before, q B (t) and q I (t) are the priori probabilities of whether the primary user exists or not, p B (X j (t)) and p I (X j (t)) are the observation probabilities of X j (t) under busy and idle states. An honest user s report probability can be calculated by P (X j (t) T j H) p B (X j (t))q B (t) + p I (X j (t))q I (t). (6) Then for each reporting round, we can update each node s suspicious level based on above equations. We set a threshold ξ and consider n as a malicious node when n is the first node such that P (T n M F t ) ξ. (7) Then, add n into Ω. Through equation (0) - (7), we have shown how to detect the first malicious node. In the k-th stage, we compute the a posteriori probability of attacker in the same manner of (0). The only difference is that when computing P (F t T n M) and P (F t T n H), we assume that all nodes in Ω are malicious. Equation (2) and (5) now become (8) and (9) respectively, and they can be seen as the special cases of (8) and (9) when Ω is empty. ρ n (t) P (X n (t) F τ ) ( N j,j n,j Ω φ n (t) ( N j,j n,j / Ω P (X j (t) T j H). P (X j (t) T j M) ), (8) j,j / Ω j,j Ω P (X j (t) T j H). P (X j (t) T j M) ), (9) Add n k to Ω when n k is the first node (not in Ω) such that P (T nk M F t ) ξ. (20) Repeat the procedure until no new malicious node can be found. Based on the above discussion, the primary user detection process is shown in Procedure. The basic idea is to eliminate the reports from users who have suspicious level higher than threshold before making final decision. In this procedure, ξ can be chosen dynamically. This procedure can be used together with many existing primary user detection algorithms, for example, hard decision combing or soft decision combing. The study in [] have shown that hard decision performs almost the same as soft decision in terms of achieving performance gain when the cooperative users (0-20) face independent fading. For simplicity, in this paper, we will use the hard decision combining algorithm in [3], [4] to demonstrate the performance of the proposed scheme and other defense schemes. Procedure Primary user detection : initialize the set of malicious nodes. 2: collect reports from N secondary users. 3: calculate suspicious level for all users. 4: for each user n do 5: if π n(t) > ξ then 6: move node n to malicious nodes set, the report from user n is removed 7: exit loop 8: end if 9: end for 0: perform primary user detection algorithm based nodes that are currently assumed to be honest. : go to step 2 and repeat the procedure IV. SIMULATION RESULTS We simulate a cognitive radio network with N( 0) secondary users. Distance from the N cognitive radios to primary transmitter is randomly distributed from 000m to 2000m. The time-bandwidth product [3], [4] is m 5. Primary transmission power and noise level is set to 200mw and -0dBm respectively. We use the two-state Markov chain channel model described in [5] as shown by Figure 2. The transition probabilities, P BI (from busy to idle) and P IB (from idle to busy) are both set to 0.. The path loss exponent is 3 and Rayleigh fading is assumed. Channel gains are updated based on node s location for each sensing report. The attack threshold is η 5, the attack strength is 5, and the attack probability P a is 00%. The suspicious level threshold ξ is set to. Three schemes of primary user detection are compared. OR Rule: When one or more secondary users reported value is greater than certain threshold, the primary user is Fig. 2. Two-state Markov chain channel model

5 5 No Attacker,N0,OR No Attacker,N0,OR No Attacker,N8,OR No Attacker,N8,OR Two Attackers,N0,OR Two Attackers,N0,K3 Two Attackers,N0,OR Two Attackers,N0,K3 Proposed Scheme,t500 Proposed Scheme,t500 Proposed Scheme,t Proposed Scheme,t Fig. 3. ROC curves (False Alarm Attack, Two Attackers) Fig. 5. ROC curves (False Alarm & Miss Detection Attack, Two Attackers) No Attacker,N0,OR No Attacker,N7,OR Three Attackers,N0,OR Three Attackers,N0,K4 No Attacker,N0,OR No Attacker,N7,OR Three Attackers,N0,OR Three Attackers,N0,K4 Proposed Scheme,t500 Proposed Scheme,t500 Proposed Scheme,t Proposed Scheme,t Fig. 4. ROC curves (False Alarm Attack, Three Attackers) Fig. 6. ROC curves (False Alarm & Miss Detection Attack, Three Attackers) considered to be present. This is the most common hard fusion method. Ki Rule: When i or more secondary users reported values are greater than certain threshold, the primary user is considered to be present. This is the straightforward defense scheme proposed in []. Proposed Scheme: Use OR rule after removing reports of malicious nodes. As shown in Figure 3-6, the ROC curves for six cases are generated. Case is N honest users, no malicious node, and OR rule. Case 2 is N 2 (or N 3) honest users, no attacker, and OR rule. Case 3-6 is N 2 (or N 3) honest users and 2 (or 3) malicious users. OR rule is used in case 3 and Ki rule is used in case 4. Case 5 and case 6 are with the proposed scheme with different detection rounds. Case 5 is the performance evaluated at round t 500 and case 6 is at round t 000. When the attack strategy is the FA Attack, Figure 3 and Figure 4 show the ROC curves when the attacker number is 2 and 3, respectively. The following observations are made. By comparing the ROC curves for case and case 3, we see that the performance of primary user detection degrades significantly when there are multiple malicious users. When detection rate is 99%, corresponding false alarm rate is higher than 80%, which implies inefficient usage of available spectrum resource. The proposed scheme demonstrates significant perfor- mance gain over the scheme without defense (i.e. OR rule) and the straightforward defense scheme (i.e. Ki rule). Table shows the false alarm rate (P f ) when detection rate is P d 99%. Table False Alarm Rate (when detection rate ) OR Ki Proposed Proposed Rule Rule (t 500) (t 000) FA,2 Attackers FA,3 Attackers FAMD,2 Attackers FAMD,3 Attackers When there are three attackers, false alarm rate for all these schemes are larger, but the performance advantage of proposed scheme over other schemes is still large. In addition, as t increases, the performance of the proposed scheme becomes close to the performance of case 2, which is the performance upper bound. Figure 5 and Figure 6 shows the ROC performance when the schemes face the FAMD attack. We observe that the FAMD attack is stronger than FA. Compared to the cases with FA attack, performance of the OR rule and Ki rule is worse when facing the FAMD attack. However, the performance of the proposed scheme is almost the same under both attacks. That is, the proposed scheme is highly effective under both attacks, and much better than the traditional OR rule and the simple defense Ki rule. The examples of false alarm rate are listed in Table.

6 Malicious Node Malicious Node 2 Honest Node Suspicious Level Suspicious Level Malicious Node Malicious Node 2 Honest Node Detection Round Detection Round Fig. 7. Dynamic suspicious level in proposed scheme (two malicious nodes perform FA attack during time [20, 00].) Fig. 8. Dynamic suspicious level in proposed scheme (two malicious nodes perform FA attack during time [5, 5].) Finally, dynamic suspicious level of honest users and malicious users are shown in Figure 7 and 8. Please note that we only demonstrate suspicious level curve for one honest node. The malicious user adopts the FA attack and dynamically chooses which round to start attack and which round to stop attack. In Figure 7, the malicious users start to attack at t 20 and stop to attack at time t 00. We can see that the suspicious level of malicious nodes increases steadily when nodes turn from good to bad. In Figure 8, one user behaves badly in only 0 rounds starting at t 5. In this case, its suspicious level increases rapidly, stay high for a number of rounds and then drop to normal. This means that the proposed scheme allows slow recovery of suspicious level after occasional bad behaviors, which may due to channel variation and unintentional errors. Figure 7 and 8 also demonstrate the convergence speed of the proposed algorithm. In this set of experiments, it takes about 30 rounds to converge. Convergence means that the suspicious levels of malicious nodes and the suspicious levels of honest nodes are largely separated. V. CONCLUSIONS Multiple malicious nodes will significantly degrade performance of collaborative sensing. In this paper, we propose an onion-peeling defense scheme. It first computes suspicious level of nodes according to reports, and then removes malicious nodes one by one based on their suspicious level until all remaining nodes are honest. Comprehensive simulations are conducted to study the ROC curves and suspicious level dynamics for different attack models, attacker numbers and different collaborative sensing schemes. The proposed schemes demonstrate significant performance advantage. For example, when there are 0 secondary users, with the primary user detection rate equals to, two malicious users can make the false alarm rate (P f ) increases to 85%. While the simple defense scheme can reduce P f to 23%, the proposed scheme reduces P f to 8%. Furthermore, when a good user suddenly turns bad, the proposed scheme can quickly increase the suspicious level of this user. If this user only behaves badly for a few times, its suspicious level can recover after a large number of good behaviors. The onion-peeling approach used in this paper is novel and demonstrates effectively differentiation of malicious nodes and honest nodes. REFERENCES [] S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE J. Select. Areas in Commun., vol. 23, no. 2, Page(s): , Feb [2] E. Hossain, D. Niyato, and Z. Han, Dynamic Spectrum Access in Cognitive Radio Networks, in print, Cambridge University Press, UK, [3] A. Ghasemi and E. S. Sousa, Collaborative spectrum sensing for opportunistic access in fading environments, in Proc. of First IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, DySPAN, Nov [4] A. Ghasemi and E. S. Sousa, Opportunistic spectrum access in fading channels through collaborative sensing, Journal of Communications (JCM), vol. 2, no. 2, pp. 7-82, Mar [5] H. Sarvanko, M. Mustonen, A. Hekkala, A. Mammela, M. Matinmikko, M. Katz, Cooperative and noncooperative spectrum sensing techniques using Welchs periodogram in cognitive radios, in Proc. of IEEE First International Workshop on Cognitive Radio and Advanced Spectrum Management, CogART 08, Feb [6] K. B. Letaief and W. Zhang, Cooperative spectrum sensing, Cognitive Wireless Communication Networks, Springer, [7] C. Sun, W. Zhang, and K. B. Letaief, Cluster-based cooperative spectrum sensing in cognitive radio systems, in Proc. of IEEE International Conference on Communications, Glasgow, Scottland, Jun [8] C. H. Lee and W. Wolf, Energy efficient techniques for cooperative spectrum sensing in cognitive radios, in Proc. of IEEE Consumer Communications and Networking Conference, Jan [9] G. Ghurumuruhan and Y. (G.) Li, Cooperative spectrum sensing in cognitive radio: Part I: two user networks, IEEE Trans. on Wireless Commun., vol.6, no.6, p.p , Jun [0] G. Ghurumuruhan and Y. (G.) Li, Cooperative spectrum sensing in cognitive radio: Part II: multiuser networks, IEEE Transactions on Wireless Communications, vol.6, no.6, p.p , Jun [] S. M. Mishra, A. Sahai, and R. W. Broderson, Cooperative sensing among Cognitive Radios, in Proc. of IEEE International Conference on Communications(ICC), Jun [2] R. Chen, J. M. Park, and K. Bian, Robust distributed spectrum sensing in cognitive radio networks, in Proc. of IEEE Infocom 2008 miniconference, Apr [3] W. Yang, Y. Cai, Y. Xu, A fuzzy collaborative spectrum sensing scheme in cognitive radio, in Proc. of International Symposium on Intelligent Signal Processing and Communication Systems, ISPACS 07, Nov [4] W. Wang, H. Li, Y. Sun, and Z. Han, Attack-proof collaborative spectrum sensing in cognitive radio networks, 43rd Annual Conference on Information Sciences and Systems, Mar [5] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized Cognitive MAC for Opportunistic Spectrum Access in Ad Hoc Networks: A POMDP Framework, IEEE J. Select. Areas Commun. (JSAC): Special Issue on Adaptive, Spectrum Agile and Cognitive Wireles Networks, vol. 25, no. 3, pp , Apr