The security of Internet of drones

The security of Internet of drones

Computer Communications 148 (2019) 208–214 Contents lists available at ScienceDirect Computer Communications journal homepage: www.elsevier.com/loca...

698KB Sizes 1 Downloads 29 Views

Computer Communications 148 (2019) 208–214

Contents lists available at ScienceDirect

Computer Communications journal homepage: www.elsevier.com/locate/comcom

The security of Internet of drones Zhihan Lv Qingdao University, China

ARTICLE

INFO

Keywords: Drones Wireless network CNN Cluster head Network security interruption probability

ABSTRACT In order to study the security of Internet of drones (IoD), convolutional neural network (CNN) algorithm was compared with autonomous IoD. Moreover, wireless communication technology was used to analyze and design a more optimized model for system security performance. The model constructed was simulated, and relevant data were collected to verify its security performance. The results show that the clustering algorithm based on node energy has the best performance in the performance analysis of IoD. When the number of nodes is appropriate, it can avoid wasting bandwidth resources and overloading, and the number of switching between clusters is less than other algorithms. Therefore, EWCA algorithm can be used to improve the lifetime of the whole network and enhances the availability of IoD. When analyzing the security performance based on the system security interruption probability, it is found that the lower the security interruption rate is when the energy acquisition coefficient 𝛼 is close to 0.5, the longer the IoD is used for information transmission, and the better the security performance is. The greater the signal-to-noise ratio is, the better the network security performance is, and the network performance is the best when the number of nodes tends to be 10. Therefore, through the research, it is found that the model built increases the security of IoD. Although there are some shortcomings in the experimental process, it still provides experimental basis for the later development of IoD.

1. Introduction With the continuous progress of social science and technology, people’s pursuit of new things is also changing. Drones are a kind of tool that can control the flight without the pilot’s operation, and it is gradually popularizing people’s life [1,2]. With the continuous expansion of its market scale, the key technology of drones has become the focus of scientific researchers. In the course of flight, drones usually need wireless network to control its network. The information collected during flight, and the pictures and paths observed need to be transmitted back, which also needs the support of network [3,4] When the information transmitted is confidential, the security performance of the network will be studied [5]. Accurate construction of an appropriate algorithm for drone’s target image detection and recognition in complex environment can not only eliminate the interference of various angles and layers, but also have higher accuracy and good robustness [6]. Therefore, convolution neural algorithm is combined with autonomous Internet of drones (IoD). Wireless communication technology is applied to analyze the system security performance and design, and the effect of the optimal security performance of IoD is explored by changing parameters, which provides a more sufficient basis for later security issues of IoD. 2. Literature review With the progress of science and technology, the drone’s industry is developing rapidly, and the number of drones is also increasing rapidly.

However, the related problems of drones have attracted the attention of researchers. IoD is a hierarchical network control architecture, which is mainly used to coordinate drone’s access to controlled airspace and provide navigation services between the locations known as nodes. Gharibi et al. [7] proposed a conceptual model of how to organize such an architecture, and specified the characteristics that an IoD system based on this architecture should implement. Therefore, the key concepts and architecture relationships of air traffic control networks, cellular networks, and Internet users were explored, and simulation was finally realized [7]. Kaleem et al. [8] introduced the structure and technical progress of amateur drones monitoring network system, which has infinite application prospects in agriculture, monitoring and many public service domains. For drone’s reconnaissance aircraft, its precise moving target detection method makes tracking more reliable and faster, and supports correct classification, which is very important for the success of drones [8]. In the same year, Garzia et al. studied the application of drones in archaeological related fields, and found that drones can connect people, objects (mobile terminals, smart sensors, devices and actuators, and wearable devices), information (data and knowledge) and processes to ensure their practicability, reasonable final cost, the lowest level of risk, and higher reliability and elasticity. Ultimately, the limitations imposed by a typical archaeological area were considered and used to develop the system [9]. In recent years, as a new technology, drones have many different applications in military,

E-mail address: [email protected] https://doi.org/10.1016/j.comcom.2019.09.018 Received 22 August 2019; Received in revised form 2 September 2019; Accepted 24 September 2019 Available online 26 September 2019 0140-3664/Β© 2019 Elsevier B.V. All rights reserved.

Z. Lv

Computer Communications 148 (2019) 208–214

civil and commercial fields, and is undergoing a revolution. Allouch et al. [10] studied the security issues of IoD, including qualitative security analysis (using international security standards) and quantitative security analysis (using Bayesian network). Finally, the scene simulation was carried out to verify that this method could enable all stakeholders to detect, explore and solve the flight risks of drones, which will help the industry to better manage the security issues of drones [10]. From the above research, it is found that although there are many problems related to drones, there are not many researches about its network security. Therefore, the research of the security of IoD is of great significance.

Fig. 1. Radio frequency HSU model.

3. Method convolution layers and multiple pooling layers [17,18]. In CNN network, the size, number and weight of convolution layer, the number and weight of pooling layer, and the number of nodes in full connection layer are all the structure and parameters of CNN unique model [19]. The structure and mode of convolution layer are very similar to human visual model. It is operationally connected with a neuron of output feature map and is called local receptive field [20]. Usually, the output of convolution kernels is obtained from the following equation:

3.1. Wireless network For IoD, it mainly uses a wireless network. In the current field of communication, wireless communication as a part of the vast market space and commercial value, the security of its physical layer needs to be paid attention to. The security of physical layer, as the assistant of encryption technology of upper layer network, can improve the overall security performance of the network, which has a very important impact on the application of wireless communication [11,12]. Physical layer security was first found in information theory, which contains many contents, such as channel capacity, which is the key index to measure the information transmission rate of communication system. In information theory, if the channel is fixed, the maximum information transmission rate is called channel capacity C: 𝐢 = max 𝐼(𝑋, π‘Œ ) 𝑝(π‘₯)

{

𝑒𝑖𝑗 =

βˆ‘

π‘–βˆˆπ‘€π‘—

βˆ— π‘Šπ‘–π‘—π‘™ + 𝑏𝑙𝑗 π‘₯π‘™βˆ’1 𝑖

= 𝑓 (𝑒𝑙𝑗 ) π‘₯π‘™βˆ’1 𝑖

The input value of the convolution core is π‘₯π‘™βˆ’1 𝑖 , the symbol βˆ— refers to the convolution operation. As the calculation of the convolution layer is not practical convolution, but product and calculation, after the convolution calculation of the convolution kernel π‘Šπ‘–π‘—π‘™ , the convolution output 𝑒𝑖𝑗 can be obtained by adding the bias constant 𝑏𝑙𝑗 , and finally the final output value can be obtained by activating the function 𝑓 (𝑒𝑙𝑗 ). In CNN, the specific calculation of pooling layer is shown in Eq. (4): βˆ‘ 1 π‘₯(𝑒, 𝑣)𝑃 ) 𝑃 𝐿𝑃 (π‘₯) = ( (4)

(1)

I (X, Y) denotes the average mutual information; p(x) indicates the probability distribution of the source. The meaning of I (X, Y) is the average amount of information about event set X obtained from event set Y. In other words, I (X, Y) represents the amount of information carried by conveying a symbol in an average sense, so it is also called information transmission rate R. Usually, R < C, because when R > C is used in noise coding, no matter how long the pipe code is, it cannot find that a coding method can reduce the probability of decoding error PE [13,14]. In real life, due to the influence of CSI (Channel State Information), the system cannot achieve secure communication, so when the security capacity is smaller than the target transmission rate, there will be a situation that the receiving terminal cannot correctly decode or eavesdropping nodes can succeed. When this happens, the system will interrupt, and the probability of such an event is called β€˜β€˜total interruption probability’’ [15]:

π‘₯(𝑒,𝑣)∈π‘₯

P is the parameter of pooling layer. When P = 1, the pooling layer is pooled by means of pooling. If P is ∞, the operation of pooling layer is maximized. In practice, the effect of random pooling is better than that of mean pooling and maximum pooling, and the worst is the generalization ability of mean pooling. For the full connection layer, it is connected with all input neurons after convolution layer and pooling layer (similar to multi-layer perceptron) [21,22]. The activation function of full connection layer usually adopts ReLU function, and its output value is shown in the following Eqs. (5) and (6): (5)

𝑓 (𝑧) = max(0, 𝑧) π‘ƒπ‘œπ‘’π‘‘ = π‘ƒπ‘Ÿ (𝐢𝑆 < π‘…π‘‘β„Ž )

(3)

(2)

𝑙

𝑙

π‘₯ = 𝑓 (π‘Š βˆ— π‘₯

π‘ƒπ‘Ÿ(𝑋) refers to the probability of occurrence of event X; π‘…π‘‘β„Ž represents the target transmission rate of the system. If the confidentiality rate of the system is set to Rs, then the transmitter can estimate the channel capacity of the eavesdropping channel through 𝐢̃𝐸 = 𝐢𝑀 βˆ’ 𝑅𝑆 . If the channel capacity of the system is 𝐢𝐸 = 𝐢𝑀 βˆ’ 𝑅𝑆 , then as long as Rs < Cs, there will be 𝐢𝐸 < 𝐢̃𝐸 . In other words, when 𝐢𝐸 < 𝐢̃𝐸

π‘™βˆ’1

𝑙

+𝑏 )

(6)

𝑓 (𝑧) is the function of ReLU, π‘₯𝑙 indicates the input value of layer l, and l-1 is the output value. In addition, π‘Š 𝑙 and 𝑏𝑙 are the weights and bias parameters of layer l, respectively. The classical CNN training method is shown in Fig. 2. 3.3. Drones ad hoc network and its algorithms

is satisfied, the transmitter can find a coding method to realize secure communication, and on the contrary, communication interruption will occur. In the security of wireless network, radio frequency is also an influential factor to network security. For example, the Harvest-Store-Use (HSU) model of energy collector is shown in Fig. 1 [16].

Drones ad hoc network is commonly used in IoD, and its structure is shown in Fig. 3. In drones ad hoc networks, the parameters affecting the algorithm usually include the degree difference of adaptive nodes, the residual energy of nodes, and the motion similarity of nodes. The weights of each parameter are determined according to the centroid weight [23–25]. Firstly, the adaptive node degree difference Di, residual node energy πœ‚i, and node motion similarity Mi of drone node i are defined. The model of centroid weight is shown in Eq. (7): π‘ž 𝑀𝑖 = βˆ‘ 𝑖 (7) π‘˜=1,π‘š π‘žπ‘–

3.2. Convolutional neural network (CNN) CNN is a branch of artificial neural network (ANN) with deep learning algorithm. It consists of input layer, convolution layer, pooling layer, full connection layer, and output layer. For example, AlexNet, a classical CNN model for image classification, is composed of multiple 209

Z. Lv

Computer Communications 148 (2019) 208–214

Therefore, in IoD, when the node degree difference is small, the adaptive node degree difference of drone i is shown in Eq. (13): 𝐷𝑖 = π‘’βˆ’πœ€π‘–

(13)

For the calculation of node residual energy, assuming that the initial energy of drone node is E and the residual energy is 𝐸s , when the node is non-cluster-head node, the energy consumption of node degree in unit time is 𝑒1 . When the node is cluster-head node, the energy consumption of node degree in unit time is 𝑒2 , so the residual energy of node is shown in Eq. (14): 𝐸𝑠 = E βˆ’

𝑖 βˆ‘

𝑒1 𝐷𝑛𝑖 𝑑𝑖 βˆ’

1

πœ‚=

(14)

𝐸𝑠 𝐸

(15)

For the motion similarity of drones, the flight coordinate information and real-time velocity information of drone are used to calculate the motion similarity between drone nodes and adjacent nodes in the coordinate system. Assuming that drone A has n adjacent drone nodes, then the average velocity difference (x-axis and y-axis) of nodes around drone Ayu is shown in Eq. (16) [26]: βˆ‘π‘› (𝑣𝐴 cos 𝛼 βˆ’ 𝑣𝑖 cos πœƒπ‘– ) 𝑣𝐴π‘₯ = 𝑖=1 (16) 𝑛 βˆ‘π‘› (𝑣𝐴 sin 𝛼 βˆ’ 𝑣𝑖 sin πœƒπ‘– ) 𝑣𝐴𝑦 = 𝑖=1 (17) 𝑛

Fig. 3. Structure of drones ad hoc network.

Therefore, the average velocity difference between drone A and adjacent drone node i is: √ 𝑣𝐴 = 𝑣2𝐴π‘₯ + 𝑣2𝐴𝑦 (18)

i can take 1,2, . . . , m, w is the parameter weight, q is the reference βˆ‘ factor, and 𝑀𝑖 ∈ (0, 1), i 𝑀𝑖 = 1. Taking these three parameters into consideration, it is found that the cluster preferred parameter of node i is π‘Šπ‘– :

The variance of velocity difference between adjacent nodes on the π‘₯-axis and 𝑦-axis is as follows: βˆ‘π‘› (𝑣𝐴π‘₯ βˆ’ 𝑣𝐴𝑖π‘₯ )2 2 𝜎𝐴π‘₯ = 𝑖=1 (19) 𝑛 βˆ‘π‘› 2 βˆ’ 𝑣 ) (𝑣 𝐴𝑖𝑦 𝑖=1 𝐴𝑦 2 πœŽπ΄π‘¦ = (20) 𝑛

(8)

Eq. (8) satisfy the relationship 𝑀1 + 𝑀2 + 𝑀3 = 1, but the specific value is determined according to the actual situation and the centroid weight. In an ad hoc network, the probability p of cluster head can be determined by Eq. (9) when the parameter of adaptive node degree is determined. 𝑆 (9) 𝑝= πœ‹π‘π‘Ÿ2 The cluster radius is a function of the total energy E consumed by the network in each round of data collection, i.e. Eq. (10): { min ∢ 𝐸 (10) 𝑠.𝑑.0 < π‘Ÿ < 1

Thus, the variance between drone A and adjacent drone nodes can be expressed as follows: 𝜎𝐴2 =

2 + 𝜎2 𝜎𝐴π‘₯ 𝐴𝑦

2

(21)

The motion similarity between drone A and adjacent drone nodes is reflected in the mean velocity difference and variance. The greater the velocity difference between the two drone nodes is, the smaller the variance is, and the smaller the similarity of the motion is. Then, the more unstable the topological structure is, the less the duration is. When the velocity difference is smaller and the variance is smaller, the motion similarity is larger and the network structure is more stable, thus the network lifetime is increased [27,28]. Therefore, the motion similarity is shown in Eq. (22):

Average node degree is the ratio of the sum of all sub-node degrees in all IoD to the number of sub-nodes in all IoD. Assuming there are n nodes, the average node degree is shown in Eq. (11): βˆ‘π‘› 𝐷𝑖 𝐷 = 𝑖=0 (11) 𝑛 The difference of adaptive node degree is obtained by subtracting the sub-node degree from the average node degree of each drone. πœ€π‘– = |𝐷𝑖 βˆ’ 𝐷|

𝑒2 𝐷𝑛𝑗 𝑑𝑗

1

i refers to the number of drone acting as non-cluster head nodes, 𝐷𝑛𝑖 indicates the degree of nodes of drone acting as non-cluster head nodes for the first time, 𝑑𝑖 suggest the time of drone acting as non-cluster head nodes for the 𝑖th time, j denotes the number of nodes of drone acting as cluster head nodes, 𝐷𝑛𝑗 refers to the node degree of drone acting as cluster head node for the 𝑖th time, and 𝑑𝑗 represents the time of drone acting as cluster head node for the 𝑖th time. At each cluster head election stage, combined with the minimum cluster radius of the average energy consumption determined above, it can be seen that the efficiency formula of drone is as follows:

Fig. 2. Classical CNN Training Model.

π‘Šπ‘– = 𝑀1 𝐷𝑖 + 𝑀2 πœ‚π‘– + 𝑀3 𝑀𝑖

𝑗 βˆ‘

( ) 2 +𝑣2 βˆ’ 𝜎𝐴 𝐴

𝑀𝐴 = 𝑒

(12) 210

(22)

Z. Lv

Computer Communications 148 (2019) 208–214

Fig. 5. Effect of number of nodes on the number of cluster heads.

Fig. 4. Wireless network model of drone.

Table 1 Specific parameter settings. Parameters

Specific settings

System signal-to-noise ratio (Ps/No) Energy conversion efficiency (πœ‚) System transmission rate (Rth)

10 dB 0.7 0.27 bit/s/Hz πœ†1 = 2.5

Mean channel gain of each node (πœ†)

πœ†2 = 3.7 πœ†3 = 1.2

3.4. Design of wireless network model for drones CNN algorithm is combined with drone ad hoc network, and wireless communication technology is used to analyze the system security performance and design, and the model diagram is shown in Fig. 4. In the model, the input layer is mainly composed of source nodes, mainly to transmit samples to the hidden layer of the neural network. Among them, the number of nodes is determined by the vector dimension of the input layer. Each dimension corresponds to a neural node, which is pre-processed before each sample size is input to make it more standardized, so as to avoid great errors in the results. The hidden function is also to encrypt the data collected by the drone to make it more secure.

Fig. 6. Effect of node moving speed on the stability of cluster structure.

on node energy has the best performance, and the number of nodes should be reasonably controlled to avoid wasting bandwidth resources and overloading (see Fig. 5). From Fig. 6, it is seen that the inter-cluster handover rates of the four algorithms increase with the increase of node speed, which is due to the complexity of drone’s topology due to its rapid movement. From the figure, it can be found that EWCA has the lowest switching rate among clusters while maintaining the same node speed, and the change of coefficient 𝛼 has no effect on the results. Therefore, EWCA algorithm fuses based on the similar mobility of nodes, with the lowest number of switching between clusters and the best stability. Fig. 7 shows the impact of the number of nodes on the lifetime of network nodes. It is found that the minimum lifetime decreases linearly with the number of nodes, and among the four algorithms, EWCA algorithm has the longest network lifetime with the same number of nodes, and does not change with coefficient 𝛼. Thus, EWCA algorithm can be used to improve the lifetime and availability of the whole network under the condition of less drone energy.

3.5. Simulation The model is simulated. The hardware environment is Intel (R) Core (TM) i7-4590 CPU @3.30 GHz, 8.00 GB RAM. The parameters are set as shown in Table 1. 4. Results and discussion 4.1. Performance analysis of IoD based on node energy When considering the influence of node energy on network performance, clustering algorithms can be used, including Highest Node Degree Algorithms (HIGHD), Weight Clustering Algorithm (WCA), Security Weight Clustering Algorithm (SWCA) and Energy-based Weight Clustering Algorithm (EWCA) [29,30]. It is found that the average cluster head of the four algorithms increases with the increase of the number of nodes. When the number of nodes is less than 600, the growth rate of cluster head is high and the relative stability of the network is poor. When the number of nodes is more than 600, the curve is relatively smooth and the network is relatively stable. It can be seen from the figure that coefficient 𝛼 has no effect on the average cluster head. Therefore, when setting up IoD, the clustering algorithm based

4.2. Analysis of system security performance based on system security interruption probability When exploring the network security performance of drones, the system security interruption probability is an important performance. Fig. 8 shows the relationship between the signal-to-noise ratio of the system and the probability of system security interruption when the energy acquisition coefficient 𝛼 is different. It can be seen from the figure that the probability of security interruption decreases with the increase of signal-to-noise ratio. When the energy acquisition coefficient 𝛼 is 211

Z. Lv

Computer Communications 148 (2019) 208–214

Fig. 7. Influence of the number of nodes on the lifetime of network nodes. Fig. 9. Effect of different energy conversion efficiencies on security interruption probability.

Fig. 8. Impact of different signal-to-noise ratios on system security interruption probability.

Fig. 10. System security interruption probability under different energy acquisition efficiencies and different energy acquisition coefficients.

different, the probability of security interruption varies with the same signal-to-noise ratio. The closer the energy acquisition coefficient 𝛼 is to 0.5, the lower the probability of security interruption is, the longer the drones will be used for information transmission, and the better the security performance will be. Fig. 9 shows the change curve of system security interruption probability under different energy conversion efficiencies. It can be seen from the figure that the better the system security performance is, the higher the signal-to-noise ratio is, and the higher the network security performance is. Therefore, the network security performance of drones can be enhanced by increasing signal-to-noise ratio and energy conversion efficiency. Fig. 10 shows the relationship between different energy conversion rates and system security interruption probability. From the figure, it can be seen that the probability of security interruption decreases with the increase of energy conversion rate. For energy acquisition coefficient 𝛼, it is found that the lower the probability of security interruption is, the stronger the network security performance is when the value of energy acquisition coefficient 𝛼 approaches 0.5. Therefore, the greater the energy conversion efficiency is and the closer the energy acquisition coefficient 𝛼 to 0.5, the better the network security performance of drones is.

seen from the figure that the probability of system security interruption decreases with the increase of signal-to-noise ratio, that is, the security performance increases, and the system security performance caused by different algorithms is different. It can be found that the interruption probability of the model algorithm system constructed is the lowest compared with Best Jammer Selection Scheme (BJS) and Best Relay Selection Scheme (BRS), and is not affected by coefficient 𝛼. Therefore, this algorithm has the best communication security performance for IoD. Fig. 12 shows the relationship between different energy conversion efficiencies and system security interruption probability. It can be found that, with the increase of system energy conversion rate, the system security interruption rate is smaller, that is, the system security performance is higher. With the increase of the number of intermediate nodes, the probability of security interruption decreases continuously, and the probability of security interruption does not change when the number of intermediate nodes increases to 10. Therefore, it can be found that, with the increase of energy conversion efficiency, when the intermediate node is 10, the system security performance is the best. Fig. 13 shows the relationship between different energy acquisition coefficients and system security interruption probability. It can be seen that when the energy acquisition coefficient 𝛼 is less than 0.5, the system security interruption performance tends to decrease, and when 𝛼 is greater than 0.5, the system security interruption performance tends to increase. Among all the algorithms, the model algorithm and the exhaustive attack method have the lowest probability of system security interruption, but the exhaustive attack method has a higher complexity.

4.3. Analysis of communication security performance of IoD by node selection Fig. 11 shows the relationship between the signal-to-noise ratio of the system and the probability of system security interruption. It can be 212

Z. Lv

Computer Communications 148 (2019) 208–214

Fig. 14. System security outage probability under different number of intermediate nodes and different energy acquisition time.

Fig. 11. Analysis of system security interrupt probability with different signal-to-noise ratios and different node selection strategies.

lowest, and the network security performance is the highest. Therefore, the greater the number of intermediate nodes is and the energy acquisition coefficient is 0.5, the higher the network security performance of drones is. 5. Conclusion In order to study the security of IoD, CNN algorithm is combined with autonomous IoD, and wireless communication technology is used to analyze and design a more optimized model for system security performance, and the model is simulated. Through relevant research, it is found that in the performance analysis of IoD based on node energy, EWCA algorithm can be used to improve the lifetime of the whole network and enhance its usability by reasonably controlling the number of nodes and the number of switching times between clusters. In the analysis of probability of network security interruption, the network security performance of drones is compared from the aspects of energy acquisition coefficient, energy conversion efficiency, signalto-noise ratio, and the number of intermediate nodes. It is found that the model constructed can make the network security performance of drones the best when the energy acquisition coefficient is close to 0.5, the energy conversion efficiency is higher, the signal-to-noise ratio is larger, and the number of intermediate nodes is increased to 10. Therefore, through the study of security of IoD, it is found that the model built increases the security of IoD, and provides experimental basis for the later development of IoD. However, there are some shortcomings in the process of this study. For instance, there are fewer parameters and IoD is relatively complex, so in the follow-up study process, more parameters in a larger range should be selected to make IoD more secure. In future, we will improve our work by novel technologies [31,32].

Fig. 12. Effect of different energy conversion efficiencies and number of nodes on system security interruption probability.

Declaration of competing interest

Fig. 13. System security interruption probability under different energy acquisition coefficients and different selection strategies.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Therefore, when the energy acquisition coefficient is 0.5, the ultimate system security performance is the highest using the model built here.

Acknowledgments

Fig. 14 shows the relationship between the number of intermediate nodes and the probability of system security interruption. With the increase of the number of intermediate nodes, the probability of security interruption decreases, and when the energy acquisition coefficient 𝛼 is close to 0.5, the probability of system security interruption is the

This work was supported by National Natural Science Foundation of China (No. 61902203) and Natural Science Foundation of Shandong Province (ZR2017QF015) 213

Z. Lv

Computer Communications 148 (2019) 208–214

References

[18] M. Saqib, S.D. Khan, N. Sharma, et al., Real-time drone surveillance and population estimation of marine animals from aerial imagery, in: 2018 International Conference on Image and Vision Computing New Zealand, IVCNZ, IEEE, 2018, pp. 1–6. [19] S. Lee, K. Son, H. Kim, et al., Car plate recognition based on CNN using embedded system with GPU, in: 2017 10th International Conference on Human System Interactions, HSI, IEEE, 2017, pp. 239–241. [20] L. Meng, T. Hirayama, S. Oyanagi, Underwater-drone with panoramic camera for automatic fish recognition based on deep learning, IEEE Access 6 (2018) 17880–17886. [21] D. Habermann, E. Dranka, Y. Caceres, et al., Drones and helicopters classification using point clouds features from radar, in: 2018 IEEE Radar Conference, RadarConf18, IEEE, 2018, pp. 0246–0251. [22] S. Paramanik, P.S. Sarkar, K.K. Mondol, et al., Survey of smart grid network using drone & PTZ Camera, in: 2019 Devices for Integrated Circuit, DevIC, IEEE, 2019, pp. 361–364. [23] J. Yang, X. Xie, W. Yang, Effective contexts for drones vehicle detection, IEEE Access 7 (2019) 85042–85054. [24] L. Wang, J. Liao, C. Xu, Vehicle detection based on drone images with the improved faster R-CNN, in: Proceedings of the 2019 11th International Conference on Machine Learning and Computing, ACM, 2019, pp. 466–471. [25] A. Montanari, F. Kringberg, A. Valentini, et al., Surveying areas in developing regions through context aware drone mobility, in: Proceedings of the 4th ACM Workshop on Micro Aerial Vehicle Networks, Systems, and Applications, ACM, 2018, pp. 27–32. [26] Y. Ma, N. Selby, F. Adib, Drone relays for battery-free networks, in: Proceedings of the Conference of the ACM Special Interest Group on Data Communication, ACM, 2017, pp. 335–347. [27] G. Beltrame, E. Merlo, J. Panerati, et al., Engineering safety in swarm robotics, in: Proceedings of the 1st International Workshop on Robotics Software Engineering, ACM, 2018, pp. 36–39. [28] S. Koulali, E. Sabir, T. Taleb, et al., A green strategic activity scheduling for drones networks: A sub-modular game perspective, IEEE Commun. Mag. 54 (5) (2016) 58–64. [29] J. Wang, C. Jiang, Z. Han, et al., Taking drones to the next level: Cooperative distributed unmanned-aerial-vehicular networks for small and mini drones, IEEE Veh. Technol. Mag. 12 (3) (2017) 73–82. [30] I. Bor-Yaliniz, H. Yanikomeroglu, The new frontier in RAN heterogeneity: Multi-tier drone-cells, IEEE Commun. Mag. 54 (11) (2016) 48–55. [31] R. Shi, Y. Gan, Y. Wang, Evaluating scalability bottlenecks by workload extrapolation, in: 2018 IEEE 26th International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems, MASCOTS, IEEE, 2018, pp. 333–347, September. [32] P. Jin, J. Guo, Y. Xiao, R. Shi, Y. Niu, F. Liu, Y. Wang, et al., PostMan: Rapidly mitigating bursty traffic by offloading packet processing, in: 2019 {USENIX} Annual Technical Conference, {USENIX}{ATC}19, 2019.

[1] L. Bai, I. Guvenc, Q. Yu, et al., Guest editorial special issue on unmanned aerial vehicles over Internet of Things, IEEE Internet Things J. 6 (2) (2019) 1636–1639. [2] W. Xu, H. Zhou, N. Cheng, et al., Internet of vehicles in big data era, IEEE/CAA J. Autom. Sin. 5 (1) (2017) 19–35. [3] Z. Lv, I. Mehmood, M. Vento, et al., IEEE access special section editorial: Multimedia analysis for internet-of-things, IEEE Access 7 (2019) 65211–65218. [4] F. Aftab, A. Khan, Z. Zhang, Hybrid self-organized clustering scheme for drone based cognitive internet of things, IEEE Access 7 (2019) 56217–56227. [5] Y. Zhang, M. Chen, V.C.M. Leung, et al., Guest editorial special issue on cognitive internet of things, IEEE Internet Things J. 5 (4) (2018) 2259–2262. [6] I. You, C. Fung, J. Baek, et al., IEEE access special section editorial: Security and privacy in applications and services for future internet of things, IEEE Access 6 (2018) 39398–39400. [7] M. Gharibi, R. Boutaba, Waslander S.L., Internet of drones, IEEE Access 4 (2016) 1148–1162. [8] Z. Kaleem, M.H. Rehmani, E. Ahmed, et al., Amateur drone surveillance: applications, architectures, enabling technologies, and public safety issues: part 1, IEEE Commun. Mag. 56 (1) (2018) 14–15. [9] F. Garzia, M. Lombardi, L. Papi, Analysis and data acquisition methodology based on flying drones for the implementation of the internet of everything to smart archaeological areas, Int. J. Herit. Archit. 2 (3) (2018) 383–394. [10] A. Allouch, A. Koubaa, M. Khalgui, et al., Qualitative and quantitative risk analysis and safety assessment of unmanned aerial vehicles missions over the internet, IEEE Access 7 (2019) 53392–53410. [11] N.H. Motlagh, T. Taleb, O. Arouk, Low-altitude unmanned aerial vehicles-based internet of things services: Comprehensive survey and future perspectives, IEEE Internet Things J. 3 (6) (2016) 899–922. [12] R.J. Hall, An internet of drones, IEEE Internet Comput. 20 (3) (2016) 68–73. [13] R. Malekian, K. Wu, K. Steenhaut, et al., Guest editorial introduction to the special issue on internet of things and sensors technologies for intelligent transportation systems, IEEE Trans. Intell. Transp. Syst. 18 (10) (2017) 2798–2801. [14] A. KoubΓ’a, B. Qureshi, Dronetrack: Cloud-based real-time object tracking using unmanned aerial vehicles over the internet, IEEE Access 6 (2018) 13810–13824. [15] M. Wazid, A.K. Das, N. Kumar, et al., Design and analysis of secure lightweight remote user authentication and key agreement scheme in Internet of drones deployment, IEEE Internet Things J. 6 (2) (2018) 3572–3584. [16] O. Holl, E. Steinbach, R.V. Prasad, et al., The IEEE 1918.1 β€˜β€˜Tactile Internet’’ standards working group and its standards, Proc. IEEE 107 (2) (2019) 256–279. [17] S. Jung, S. Hwang, H. Shin, et al., Perception, guidance, and navigation for indoor autonomous drone racing using deep learning, IEEE Robot. Autom. Lett. 3 (3) (2018) 2539–2544.

214