Real-time Framework for Trust Monitoring in aNetwork of Unmanned Aerial Vehicles
Mahsa Keshavarz, Alireza Shamsoshoara, Fatemeh Afghah, Jonathan Ashdown
AA Real-time Framework for Trust Monitoring in aNetwork of Unmanned Aerial Vehicles
Mahsa Keshavarz ∗ , Alireza Shamsoshoara ∗ , Fatemeh Afghah ∗ , Jonathan Ashdown †∗ School of Informatics, Computing and Cyber Systems, Northern Arizona University, Flagstaff, AZ, USAEmail: { mk959,alireza shamsoshoara,fatemeh.afghah } @nau.edu † Computer Information Systems Department, SUNY Polytechnic Institute, Utica, NY, USA,E-mail: [email protected]
Abstract —Unmanned aerial vehicles (UAVs) have been increas-ingly utilized in various civilian and military applications suchas remote sensing, border patrolling, disaster monitoring, andcommunication coverage extension. However, there are still proneto several cyber attacks such as GPS spoofing attacks, distributeddenial-of-service (DDoS) attacks, and man-in-the-middle attacksto obtain their collected information or to enforce the UAVs toperform their requested actions which may damage the UAVsor their surrounding environment or even endanger the safetyof human in the operation field. In this paper, we propose atrust monitoring mechanism in which a centralized unit (e.g. theground station) regularly observe the behavior of the UAVs interms of their motion path, their consumed energy, as well asthe number of their completed tasks and measure a relativetrust score for the UAVs to detect any abnormal behaviorsin a real-time manner. Our simulation results show that thetrust model can detect malicious UAVs, which can be undervarious cyber-security attacks such as flooding attacks, man-in-the-middle attacks, GPS spoofing attack in real-time.
Index Terms —Trust monitoring, UAV networks, cyber attacks,selfish behavior.
I. I
NTRODUCTION
Unmanned aerial vehicles (UAVs) recently play a major rolein several civilian and military operations including remotesensing, surveillance, package delivery, and medical services[1]–[6]. Their unique features including high-mobility, ease ofdeployment, and their ability to hover enable them to provideservices in hard-to-reach regions or time-critical missionsduring man-made and natural disasters in order to provideurgent Internet and communication services when necessaryor for imaging purposes [7], [8].Despite this wide range of often critical UAV missions, theUAV networks are still vulnerable to several attacks includingcyber-attacks such as false data injection [9], physical attackssuch as targeting the UAVs using firearms [10], and cyber-physical attacks such as GPS spoofing [11]. Therefore, animportant step toward safe applications of UAVs networksis developing robust trust monitoring mechanisms to identifypotential attacks on these systems during the flight.There are several existing trust management mechanismsas further discussed in Section II to detect the maliciousUAVs. However, the majority of these methods either focuson a specific type of aircraft, or an specific type of cyber-security attacks. Another challenge with several current pro-posed methods is that they are not able detect the attacks inreal-time, since they often look at the history of the UAVs’trust to evaluate if they have deviated from normal behaviorlong enough to impact their reputation [12]. It means that if the UAVs with a good reputation are targeted by an attacker,the attack may not be detected immediately. The reputation-based methods also involve storing the reputations of all agentsover the course of time at the central unit, which require alarge memory space. Another trend in trust monitoring is theclass of cooperative trust monitoring methods, where all theagents coordinate with one another to monitor the behavior oftheir teammates. While such methods are robust against the”single point of failure” problem, where the centralized trustmonitoring unit may be attacked, they involve large energyconsumption and computation load at the UAVs to observeother agents’ behavior and also impose a heavy signaling loadto the system. Some other trust monitoring approaches [13]were proposed based on observing the statistical measures todetect malicious UAV in real-time, however, these methodscannot distinguish between the malicious behavior or potentialchange of behavior due to harsh environmental conditions.In this paper, we propose a trust monitoring framework todetect various attacks such as DDoS attack, GPS spoofingattack, man-in-the-middle attack, and the agent’s potentialselfish act in a network of UAVs in a real-time manner. Inthis method, the relative trust scores of all UAVs are regularlymeasured at a central unit by observing several factors includ-ing (i) the observation of the central unit from the environmentto estimate how successful a UAV has been in completingits assigned tasks, (ii) the deviation of the UAVs from theiroriginal path, and (iii) the energy consumption of the UAVs.The UAVs under the DDoS and GPS spoofing attacks consumemore energy than normal UAV, and are more likely to deviatefrom the predicted paths. Also, the UAVs under the man-in-the-middle attack often complete a less number of assignedtasks related to a normal condition. Therefore, the selectedfactors to monitor the trust can capture various common at-tacks. The trust score of each UAV will be evaluated comparedto the trust scores of its teammates to identify if a UAV hasbeen attacked. This method can be applied to any types ofaircraft, can detect the attacks in real-time instead of relyingon the history of the agents’ behavior and does not requirelarge memory or computation capabilities. Furthermore, theproposed method can differentiate the malicious behavior fromthe unstable behavior due to harsh environmental conditions.II. R
ELATED W ORKS
There are several methods to monitor the safety of differ-ent aircraft such as the fault detection methods which candetect any abnormalities in the aircraft. Qi et al. [14], [15] a r X i v : . [ ee ss . SP ] J u l rovides a comprehensive review of recent methods for faultdiagnosis. Birnbaum et al. [16] propose a method in which theRecursive Least Squares (RLS) algorithm is applied to performestimation and tracking of the controller parameters. In thismethod, data in batches of size 500 samples are processedand compared to detect discrepancies indicating anomalies.Several anomaly detection methods are tailored to specifictypes of attacks. [16] overviews of the most popular six typesof threats for UAVs, such as hardware failure, malicious hard-ware, and control computer attacks. In order to detect sometypes of hijacking or hardware failures, a Hardware HealthMonitor is utilized, which monitors all of the UAV sensor data.This data consists of the positioning data and flight surfacesmonitored by using the RLS algorithm. Another commonattack in UAV networks is the GPS spoofing attack. In orderto detect this type of attack, [17] discussed a technique, inwhich the receiver observes the received signal strength andcompares it to the predicted signal strength over time. [18]proposed a method that uses two GPS receivers and check theircross-correlations to detect GPS spoofing attacks. However,this technique cannot detect the spoofing attack when thesignals are weak. In order to detect the GPS spoofing attacksand determine the real locations of UAVs, [10] proposes aframework, where the UAVs are allowed to travel on theshortest path between any given two locations by using theoptimal UAVs controller. This model can capture the possibleGPS spoofer on the UAVs traveling path. This method enablesthe UAVs to determine their real location under GPS spoofingattack by using the neighbor UAVs real location and theirrelevant distances.[13] proposed a model-based trust monitoring approachwhere several basic statistical measures are used to track thecharacteristics of a flight. These statistical measures includingthe mean, variance, and covariance across multiple flights atdifferent stages of the flight are stored as a fingerprint of theflight, where the variations of the flight pattern can alarm thatthe UAV was hijacked. A threshold is set to determine at whatpoint the flight is no longer normal. In this method, the normalflight statistics are added to the baseline and any abnormalones are removed from the baseline. This method is tested byusing 20 potential hijacking simulations and 50 safe flights.All of the obvious hijacking scenarios were detected correctly,however some unstable flights that were close to the baselinewindy flights remained undetectable.III. S YSTEM M ODEL
Here, we propose a centralized trust monitoring mechanismto monitor the behavior of multiple UAVs during the flightand detect any potential abnormal behaviors in real-time. Letus consider a homogeneous network of N UAVs which flyover a M × M area, where M is a positive arbitrary integer. M is chosen randomly from [1 . km, . km ] range. EachUAV is assigned with a particular task which determines theUAV’s flight pattern during its operation. Some examples ofthese tasks include package delivery, aerial photography, andgeophysical survey. It is assumed that the UAVs are randomlydistributed in the area of the network prior to the mission. It isalso assumed that among the UAVs which operate in a closeproximity of each other, only one UAV is under an attack in a given time [10]. In proposed trust monitoring model, wecalculate the trust score of each UAV and compare it withthe trust scores of its neighbor UAVs. Since all the neighborUAVs experience similar environmental conditions, if a UAVpresents an out-of-range trust score it can be identified asan under-attack UAV. This trust monitoring method can bescaled up to a large network by dividing the UAVs to severalclusters, where the UAVs withing each cluster experiencesimilar environmental conditions.After clustering the neighbor UAVs, the central unit (e.g.,an audit unit or the ground station) regularly measure thetrust score of all the UAVs within a cluster to keep track oftheir behavior in order to detect any abnormalities in a real-time manner. After each mission, the trust score of the UAVsis reset to zero. To calculate this trust score , we considerthree factors: the task success rate, the energy consumptionrate, and the deviation of the UAV’s flight trajectory from itsexpected path. The reason behind choosing these three factorsis that most common attacks impact one or more of thesebehavioral factors. For instance, when a UAV is under theman-in-the-middle attack, it usually skips some assigned tasks(i.e. present a low task success rate). When a UAV is underthe DDoS attack or the hijacking attack, it will present anunusual energy consumption pattern, where it may consumemore energy (because the attacker forces the UAV to performunexpected tasks) or less energy (because the attacker preventsthe UAV from performing its task), or when the UAV is underthe GPS spoofing which the attacker wants to mislead the UAVinto another location (deviation from the predicted path). Incontinue, we define the three factors of trust. A. Task Success Rate
One of the main factors to evaluate the trustworthiness ofan agent is the task success rate. During a flight, each UAVis assigned with a couple of tasks and they are expectedto complete all of these tasks. It means that in a normalcondition, the number of successfully completed tasks shouldbe higher than the number of failed tasks. Sometimes due toenvironmental and technical issues, the UAVs are not able tocomplete all of their tasks, but still, the number of failed tasksis lower than the number of successful tasks. However, whenthe agents are under certain attacks such as the man-in-the-middle attack, or the forwarding attack, they may deliver aless number of successful tasks.To evaluate the behavior of UAVs in terms of performingtheir assigned tasks, we consider the history of completedtasks for the UAVs during each mission. The trust factor isexpected to differentiate whether the underlying reason for notcompleting the assigned tasks is related to an attack or becausethe UAV is experiencing any environmental or technical issues.Therefore, the trust factor considers the relative performanceof the UAVs rather than an individual performance. In thiscase, if the UAVs are operating under bad weather conditions,they all will underperform and deliver a low task success rate.However, if one of the UAVs is under an attack, only thisUAV will experience a low task completion rate. Moreover,since the task success rate is measured upon the completionof a mission, it gives the UAVs the time to recover from thepotential technical issues they may face. For example, if theAV has a low battery level, it has the opportunity to puta hold on performing the tasks to land on a recharging spotand then complete the rest of the assigned tasks. Hence, whenlooking at the record of this UAV during the entire mission,the number of successful tasks is higher than the failed tasks.The history of completed tasks can be considered eitherbased on the direct reports of the UAVs which is prone toreceiving false reports from the under-attack UAVs or it canbe estimated based on the direct or indirect observation of thecentral unit. For instance, if the UAV’s task was to survey aspecific area, the central unit can determine if the task has beensuccessfully completed based on whether obtaining the imageof the entire region or not rather than relying on the UAV’sself-report. Indeed, the central unit cannot always have sucha direct observation for all the assigned tasks. For instance,if the UAV’s task was to drop a fire ball to ignite fire in aparticular area (a mechanism used to initiate controlled firesto prevent wildfires [19] ), then the central unit may not beable to accurately determine if the ball was dropped. However,it can estimate if the task was performed by the UAV byobserving its impact on the field (i.e. fire in the target area)with some level of uncertainty. To account for such potentialuncertainty in the assessment of the central unit in terms ofthe number of successful or failed tasks for each UAV, wepropose a trust factor based on subjective logic framework(SLF) [20]. Let us assume that s is the number of successfultasks, f is the number of unsuccessful tasks, and x denotes thenumber of tasks that the central unit can not certainly declareas successful or unsuccessful. Based on this theory, the trustis composed of a vector T = { b, d, u } . The parameters b, d, u respectively represent the probability of trust, the probabilityof distrust, and the chance of uncertainty, where they satisfy b + d + u = 1 and b, d, u ∈ [0 , . The aforementionedparameters are calculated as follows: b = ss + f + x , d = fs + f + x , u = xs + f + x (1)The calculation of the task success rate for binary state-ments, which in our case, it is reliable and unreliable UAVsis as below [20]: T task = 2 b + u (2) B. Energy Consumption
Another important factor to asses the level of trust foran agent is to calculate its energy consumption noting theexpected amount of energy for its assigned task. In a normalcondition, the energy consumption rate for each task shouldbe within a certain range. However, when the UAVs are underan attack (e.g., the flooding and GPS spoofing), the energyconsumption rate would be higher or lower than the normalrange. Further, in bad weather conditions (e.g., strong winds),the UAVs need to consume more energy to complete theirassigned tasks. In order to distinguish between the maliciousconditions and environmental conditions, we compare theenergy consumption of the UAVs in one cluster. For example,if the energy consumption of one UAV is higher than the other neighbor UAVs, it means that the UAV is under a flooding or aGPS spoofing attack. Also, a low or high energy consumptionof a single UAV can be an indicator that the control of the UAVhas been taken over by an adversary. However, under unstableenvironmental condition like heavy wind, all the neighborUAVs in a cluster present an unusual energy consumption asthey all experience similar a weather condition. Noting theassumption that the attackers can target one UAV at a time inone cluster [10], this proposed approach can reduce the rateof false alarms in reporting the UAVs as malicious in extremeenvironments.As we mentioned before, each UAV in the same clusterhas the same initial energy, and since all UAVs are assignedwith similar types of tasks, they are expected to have a similarenergy consumption. In order to evaluate the behavior of theUAVs within a cluster in terms of their energy consumption,we need to compare the energy consumption of each UAV(let us say UAV i ) with the average consumption of all otherUAVs in that cluster based on (3). Let K denote the number ofUAVs in one cluster, and E i denote the consumption energyof the UAV i . The energy trust factor of UAV i , T ene i canbe calculated based on the absolute value of the differencebetween E i and E ave i as shown in (4). E avg − i = 1 K − K (cid:88) j =1 ,j (cid:54) = i E j , (3) T ene i = | E i − E avg − i | E avg − i (4) C. Path Deviation
Each UAV in a cluster is assigned with several similar tasksduring their mission, and a particular path is expected for eachtask before the mission is started. We assume that the UAVsshould follow the expected path; however, due to the differentreasons, the actual trajectory of the UAVs may be differentfrom their expected paths. One reason can be an extremeweather condition (e.g., strong wind) to cause deviation fromthe predicted path. Also, an obstacle in the UAV’s expectedpath can cause a deviation from the original path for a shorttime. Another reason for such deviation from the expected pathcan be different cyber-physical attacks such as GPS spoofing.In order to find out the main reason behind the UAV’spath deviation, the audit unit observes the current location ofthe UAVs based on the GPS information, and also it has anestimation of the existing obstacles in the predefined region.In order to calculate the deviation trust, the audit unit looksat the history of the actual locations for α consecutive timeslots. At each time slot, the audit unit calculates the differencebetween the actual and expected location; then, it calculates theaverage of the differences for the α time slots. We can define α based on the size of predefined obstacles. (5) expresses thetrust deviation at each time slot: dev = (5) t (cid:80) i = t − α (cid:112) ( x ex i − x ac i ) + ( y ex i − y ac i ) + ( z ex i − z ac i ) α , where x ex , y ex , z ex is the expected location of the UAV and x ac , y ac , z ac is the actual location at time i .If the deviation from the expected path is related to facingan object or temporal harsh weather conditions, the UAV isexpected to return to the original path withing a short portionof time; however, if the UAV has been attacked this deviationwill be observed for a long period of time. In this case, thevalue of deviation trust is increased, which impacts the totaltrust score. Since, the trust scores are compared with oneanother, if the UAVs face extreme weather condition, they willall have a considerable long-term deviation from their originalpaths and this condition can be differentiated from potentialattacks.Environmental issues may consume more energy than nor-mal condition and cause deviation from the expected path. Aswe discussed before, we compare the energy consumption ofthe UAV with the neighbor UAVs, which they are in the samecluster. In this case, by considering the amount of both energyconsumption and the deviation, we can recognize which oneis attack deviation and which one is a weather deviation. D. Calculation of Trust score
The overall trust score of a UAV is defined by integratingthe task success rate trust, the energy trust, and the deviationtrust as follows: T total = w task × T task + w ene × T ene − w dev × T dev , (6)where w task , w ene , and w dev are weights, where w task + w ene + w dev = 1 , w task , w ene , w dev ∈ [0 , . We first calculatethe task success rate trust, energy trust and deviation trust, thenobtain the overall trust score of an agent using (6). Since thedeviation from the actual path decreases the total trust score,we subtract the deviation from the total trust score. Then, thetrust score of each UAV in one cluster is compared with thetrust score of its neighbor UAVs to see whether the trust scoresare in the same range or not. For example, we have three UAVsas it is shown in Algorithm 1. In the normal condition, the trustscore of three UAVs should be in the same range. If the rangeof trust score of U AV is different from the other UAVs, itmeans the U AV is under an attack.IV. S IMULATION R ESULTS AND A NALYSIS
In this section, we evaluate the performance of the proposedtrust monitoring model in detecting the malicious UAVs us-ing extensive simulations. We consider different Monte-Carloscenarios, where each scenario runs 1000 times for differenttypes of tasks, and the average is calculated as a result. In eachscenario, we assume we have three UAVs in one cluster, whichmeans they have the same initial energy, similar type of task,and all of them are in the same environmental condition. Onlyone UAV in a cluster can be under an attack. In each scenario,we study one attack, which impacts at least one factor of thetrust model. In this model, we monitor the trust score of the UAVs bycalculating the task success rate trust, the energy trust, and thedeviation trust at the ground station in a periodic way (e.g. weconsider four-minute intervals). After calculating these threefactors, we calculate the total trust score for each UAV andcompare them to see whether they are in the same range ornot. Based on the assumption that in each flight, just one ofthe UAVs can be under the attack, if the range of one UAVstrust score is different from the others, that UAV is reportedmalicious. U A V U A V U A V UAV T r u s t S c o r e s Normal Condition
UAV1UAV2UAV3 (a) Trust scores of all UAVs under a normal condition.
UAV1 (Expected Path)UAV1 (Actual Path)UAV2 (Expected Path)UAV2 (Actual Path)UAV3 (Expected Path)UAV3 (Actual Path) (b) Flight trajectory of all UAVs compared to their predicted pathsunder a normal condition
Fig. 1: UAVs’ behavior under a normal condition.There are three different factors of energy consumption,deviation the UAVs from their expected path, and the numberof tasks that impact the trust score of each UAV. Figure 1shows the performance of the UAVs in a normal conditionwhere all of the trust scores are in the same range (Fig. 1a), andall the UAVs have a low-level of deviation from their predictedpaths (Fig. 1b). Then, we consider different attack scenariosto study whether our trust model can detect the attacks if atleast one of the factors is impacted. In these scenarios, weassumed that
U AV is a malicious UAV that can be impactedby various attacks.First, we consider a scenario where U AV is under theDDoS attack. In this case, the energy consumption of U AV is higher than the other UAVs operating in the same cluster.As shown in Fig. 2a, the range of trust score for U AV is A V U A V U A V UAV T r u s t S c o r e s DDoS Attack
UAV1UAV2UAV3 (a)
UAV is under the DDoS attack U A V U A V U A V UAV T r u s t S c o r e s GPS spoofing Attack
UAV1UAV2UAV3 (b)
UAV is under the GPS spoofing at-tack U A V U A V U A V UAV T r u s t S c o r e s Man-in-the-middle Attack
UAV1UAV2UAV3 (c)
UAV is under the man-in-the-middleattack UAV1 (Expected Path)UAV1 (Actual Path)UAV2 (Expected Path)UAV2 (Actual Path)UAV3 (Expected Path)UAV3 (Actual Path) (d) Flight pattern of
UAV under GPSspoofing attack U A V U A V U A V UAV T r u s t S c o r e s Strong Wind
UAV1UAV2UAV3 (e) Trust score of all UAVs when experi-encing a strong wind. U A V U A V U A V UAV T r u s t S c o r e s Selfish UAV
UAV1UAV2UAV3 (f) Trust scores of UAVs when
UAV actsselfishly or being hijacked. Fig. 2: Trust scores and different scenarios for UAV
500 1000 1500 2000 2500 3000-800-600-400-200020040060080010001200
UAV1 (Expected Path)UAV1 (Actual Path)UAV2 (Expected Path)UAV2 (Actual Path)UAV3 (Expected Path)UAV3 (Actual Path)
Fig. 3: Wind and GPS Spoofing for flight patterns.higher than other UAVs, since it has larger value of energyconsumption.In the second scenario, we study the case where
U AV is under the GPS spoofing attack. Under such attack, theUAV consumes more energy compared to the average energyconsumption rate, and it will not follow its expected path asshown in Fig. 2d. Noting (6), the value of the deviation trustis a large number, hence the trust score is a negative numberas shown in Fig. 2b.The third scenario is the man-in-the-middle attack, in whichthe number of failed tasks for the U AV is higher than thenumber of successful tasks. In Fig. 2c, the number of failedtasks is higher than the number of successful tasks for U AV .Based on (2), the task trust is decreased, thus the trust scoreof U AV is less than the other UAVs. U A V U A V U A V UAV T r u s t S c o r e s Strong Wind with GPS spoofing Attack
UAV1UAV2UAV3
Fig. 4: Trsut score of all UAVs in bad weather with GPSspoofingThe fourth scenario refers to the case that
U AV has notdelivered its assigned tasks. This behavior can be an indicatorthat the UAV has been hijacked or show that the UAV is actingselfishly and does not follow the controller orders. This attackcan translate to investing less energy in the assigned taskscompared to the average energy consumed by other UAVs.In Fig. 2f, the U AV is selfish, hence the values of the tasksuccess rate and the energy consumption are small numbers,and consequently the trust score of this UAV is less than theother UAVs.In the last scenario, we evaluate the performance of the pro-posed trust management mechanism in terms of understanding lgorithm 1: Trust model pseudo code for all UAVs for all UAVs do Calculate the task trust score based on (2)Calculate the energy trust score based on (4)Calculate the deviation trust score based on (5)And calculate the total trust score ( T i ) based on (6) endfor all UAVs doif The range of the T i s is different from other UAVs then UAV i is not reliable else No attack happened for UAV i endend the difference between the UAVs’ abnormal behavior whenthey are under an attack or when they are operating underharsh environmental conditions. First we test this scenario withthe strong wind. In the strong wind scenario, all the UAVsconsume more energy and show deviation from their predictedpaths, therefore all of the trust scores are in the same range. Werun this scenario 1000 times for different types of tasks and theaverage results show that our model can correctly identify thestrong wind with 70% accuracy, as depicted in Fig. 2e. Then,we test this scenario again by assuming that all the UAVsexperience the strong wind while UAV i is also under theGPS spoofing attack. As we can see in Fig 3, all of the UAVshave deviation from their expected path, however U AV hasmore deviation due to the GPS spoofing attack. Based on (6)the trust score of the UAV under the attack is bigger than therest of UAVs (shown in (4)). Therefore, the proposed modelcan distinguish between the deviation of expected path due tothe strong wind and the GPS spoofing attack.The UAV node gets a signal from each DS. By subtractingthe time the signal was transmitted from the time it wasreceived, each DS can calculate how far it is from the UAVnode. So given the travel time of the signals from the DS andthe exact position of DSs, the position of the UAV node canbe determined in three dimensions - east, north and altitude.As it it shown in equation 7, t i is the time that the UAV nodereceived the signal, s i is the time the signal was transmittedby the DS i , and x i , y i , z i shows the exact location of DS i . β is the receiver’s clock bias from the much more accurate DS’sclock. d ( i,j ) = ( t i − β − s i ) ∗ c (7) d ( i,j ) = (cid:112) ( x − x i ) + ( y − y i ) + ( z − z i ) , (8)V. C ONCLUSION
The problem of real-time detection of cyber-physical attacksin a network of drones is studied. We developed an onlinetrust monitoring mechanism in which a central unit regularlyobserves the operations of the UAVs in terms of their flighttrajectory, their energy consumption as well as performing their assigned tasks and compare their trust factors to identifyany potential abnormal behaviors. The proposed comparativeapproach enables the audit unit to differentiate between theabnormal behaviors due to the cyber-physical attacks or thepotential unusual actions (e.g. turbulence or irregular energyconsumption) due to facing harsh environmental conditions.Further, rather than relying on the self-report of the UAVs todeclare the number of their completed tasks or their futurepath, the proposed trust monitoring approach estimates theperformance of the UAVs based on the observation of theaudit unit while accounting for the potential uncertainty insuch observations. R
EFERENCES[1] A. Rahmati, X. He, I. Guvenc, and H. Dai, “Dynamic mobility-aware interference avoidance for aerial base stations in cognitive radionetworks,” in
IEEE INFOCOM 2019-IEEE Conference on ComputerCommunications . IEEE, 2019, pp. 595–603.[2] J. C. Hodgson, S. M. Baylis, R. Mott, A. Herrod, and R. H. Clarke,“Precision wildlife monitoring using unmanned aerial vehicles,”
Scien-tific reports , vol. 6, p. 22574, 2016.[3] A. Eldosouky, W. Saad, and N. Mandayam, “Resilient critical infrastruc-ture: Bayesian network analysis and contract-based optimization,” arXivpreprint arXiv:1709.00303 , 2017.[4] A. Shamsoshoara, M. Khaledi, F. Afghah, A. Razi, J. Ashdown, andK. Turck, “A solution for dynamic spectrum management in mission-critical uav networks,” arXiv preprint arXiv:1904.07380 , 2019.[5] A. Shamsoshoara, M. Khaledi, F. Afghah, A. Razi, and J. Ashdown,“Distributed cooperative spectrum sharing in uav networks using multi-agent reinforcement learning,” in . IEEE, 2019, pp.1–6.[6] A. Shamsoshoara, F. Afghah, A. Razi, S. Mousavi, J. Ashdown, andK. Turk, “An autonomous spectrum management scheme for unmannedaerial vehicle networks in disaster relief operations,” arXiv preprintarXiv:1911.11343 , 2019.[7] M. Mozaffari, A. T. Z. Kasgari, W. Saad, M. Bennis, and M. Debbah,“Beyond 5g with uavs: Foundations of a 3d wireless cellular network,”
IEEE Transactions on Wireless Communications , vol. 18, no. 1, pp. 357–372, 2018.[8] F. Afghah, A. Razi, J. Chakareski, and J. Ashdown, “Wildfire monitoringin remote areas using autonomous unmanned aerial vehicles,” in
IEEEINFOCOM 2019 - IEEE Conference on Computer CommunicationsWorkshops (INFOCOM WKSHPS) , April 2019, pp. 835–840.[9] U. Challita, A. Ferdowsi, M. Chen, and W. Saad, “Machine learningfor wireless connectivity and security of cellular-connected uavs,”
IEEEWireless Communications , vol. 26, no. 1, pp. 28–35, 2019.[10] A. Eldosouky, A. Ferdowsi, and W. Saad, “Drones in distress: A game-theoretic countermeasure for protecting uavs against gps spoofing,” arXiv preprint arXiv:1904.11568 , 2019.[11] R. Altawy and A. M. Youssef, “Security, privacy, and safety aspectsof civilian drones: A survey,”
ACM Transactions on Cyber-PhysicalSystems , vol. 1, no. 2, pp. 1–25, 2016.[12] F. Afghah, M. Zaeri-Amirani, A. Razi, J. Chakareski, and E. Bentley,“A coalition formation approach to coordinated task allocation in hetero-geneous uav networks,” in , June 2018, pp. 5968–5975.[13] J. McNeely, M. Hatfield, A. Hasan, and N. Jahan, “Detection of uavhijacking and malfunctions via variations in flight data statistics,” in . IEEE, 2016, pp. 1–8.[14] X. Qi, D. Theilliol, J. Qi, Y. Zhang, and J. Han, “A literature reviewon fault diagnosis methods for manned and unmanned helicopters,” in .IEEE, 2013, pp. 1114–1118.[15] X. Qi, J. Qi, D. Theilliol, Y. Zhang, J. Han, D. Song, and C. Hua, “Areview on fault diagnosis and fault tolerant control methods for single-rotor aerial vehicles,”
Journal of Intelligent & Robotic Systems , vol. 73,no. 1-4, pp. 535–555, 2014.[16] Z. Birnbaum, A. Dolgikh, V. Skormin, E. OBrien, D. Muller, andC. Stracquodaine, “Unmanned aerial vehicle security using recursiveparameter estimation,”
Journal of Intelligent & Robotic Systems , vol. 84,no. 1-4, pp. 107–120, 2016.17] J. S. Warner and R. G. Johnston, “Gps spoofing countermeasures,”
Homeland Security Journal , vol. 25, no. 2, pp. 19–27, 2003.[18] B. W. OHanlon, M. L. Psiaki, T. E. Humphreys, and J. A. Bhatti, “Real-time spoofing detection using correlation between two civil gps receiver,”in
Proceedings of the ION GNSS Meeting . Institute of NavigationNashville, TN, 2012.[19] E. Beachly, J. Higgins, C. Laney, S. Elbaum, C. Detweiler, C. Allen,and D. Twidwell, “A micro-uas to start prescribed fires,” in
InternationalSymposium on Experimental Robotics . Springer, 2016, pp. 12–24.[20] A. Josang, “An algebra for assessing trust in certification chains,” in