Trust Management Framework for Internet of Things

Transcription

1 2016 IEEE 30th International Conference on Advanced Information Networking and Applications Trust Management Framework for Internet of Things Yefeng Ruan, Arjan Durresi, Lina Alfantoukh Indiana University Purdue University at Indianapolis Indianapolis, IN, USA Abstract Internet of Things (IoT) can connect a large number of things (or agents) through communication networks for various types of applications. Like in many other applications, it is very important for all the agents in IoT systems to collaborate with each other following predefined protocols. In this paper, we proposed a general trust management framework aiming to help agents to evaluate their partners trustworthiness. We run a simulation for a food nutrition analysis example. It shows that by using trust, the analysis error can be reduced. Also, we illustrate two possible types of attacks, and show how to use different trust factors or environments together to alleviate the damage. I. INTRODUCTION Internet of Things (IoT) can connect a large number of things through communication networks for various types of applications. In IoT systems, things can be sensors, monitors, smart devices, laptops, or even human beings. In the following of this paper, we call things agents in general. IoT integrates physical objects with the virtual world [25]. With the rapid development of smart devices and sensing technology, IoT has a promising future and is widely used in many fields, such as supply chain, system monitoring, health care, and so on [1] [24] [19]. It can make people s life more convenient and intelligent by making use of smart devices together; however, it also faces several security challenges [15] [23]. In IoT systems, depending on applications, agents may collaborate with others to fulfill complex tasks [4]. For example, sensors collect environmental data and send them to a computational device for pre-processing. The device parses the data and filters out the noise based on the pre-defined criteria. Then the pre-processed data will be send to some more powerful devices, e.g. Cloud servers, to do data mining. The analysis results will be sent back to users through some intermediate agents, and then users adjust their behaviors based on the received feedback. Also, similar to P2P systems, agents can send requests to other agents to ask for various types of services [16] [8]. Since agents in IoT systems need to collaborate with each other, it is very important for them to find a way to evaluate their partners trustworthiness. Also, given a complex task fulfilled by a chain of agents, it is necessary for the end users to evaluate all the involved agents trustworthiness. Therefore, how to manage trust relationships becomes a challenging but very important topic for IoT systems [24] [4] [23] [16] [15]. Since trust itself is a subjective concept, it is very difficult to define it. In this paper, if the object satisfies the subject, e.g. forwarding data following defined protocol, we assume that the object is trustworthy in the subject s point of view. Otherwise, the object is untrustworthy in the subject s point of view. According to [10], we call the agents who evaluate others subject. And the agents being evaluated are called object. Environments are the platforms or factors that form trust relationships. We propose a general trust management framework for IoT systems. Instead of considering specific determining environments or factors for trust in specific applications, we propose a general framework to measure trust based on the measurement theory. Besides this, our framework is able to help the end user to evaluate the trustworthiness of the operation completed by multiple agents. To better illustrate our framework, we run a simulation for a food nutrition analysis example. To our best knowledge, it is the first work that models trust in a general way for IoT systems, and can be applied to various types of applications. The remainder of this paper is organized as following: in Section II, we introduce the background knowledge about IoT and some literature work in this field. In Section III, we illustrate our trust management framework. In Section IV, we run a simulation for a food nutrition analysis example, and illustrate two possible types of attacks. In Section VI, we conclude this paper. II. BACKGROUND AND RELATED WORKS A. Internet of Things (IoT) Internet of Things, as the term indicates, contains two terms: Internet and Things [4]. Internet mainly focuses on the communication among agents, e.g. communication protocols. Also, it considers security issues in the communication, e.g. encrypted packets. Things, on the other hand, focuses on agents functionalities and roles. For example, in order to capture temperature information, certain devices must be provided. More important, IoT systems connect physical agents including human beings with the virtual world. When integrated with Internet, agents are no longer isolated. Instead, they are connected with each others. Therefore, not only communication protocols and agents functionalities should be taken into account, the relationships among agents are also very important. We refer it as agent-to-agent relationship. In this paper, we further divide agents into two X/16 $ IEEE DOI /AINA

2 categories: human beings and devices. Accordingly, agentto-agent relationship can be divided into human-to-human, device-to-device, human-to-device and device-to-human relationships [8]. These relationships are accumalted based on previous experiences or prior knowledge. For example, human beings accumulate their impressions towards others based on previous interactions [26] [20]. Similarly, devices in IoT systems are able to communicate with each other to evaluate others trustworthiness [16]. When human beings use devices, they can also form impressions towards the devices. This can help human beings or devices to make educated decisions later. Furthermore, given open environments in IoT systems, it can help agents to avoid to collaborate with malicious agents [5]. B. Related works As we can see, trust plays an important role in many applications as well as in people s life [23]. Although IoT is relatively new, many works have been developed to exploring the usage of trust in IoT systems. [16] takes the idea of trustworthiness in social networking. It considers four relationships among agents: parent object relationship, co-location object relationship, owner object relationship and social object relationship. Among them, social object relationship is based on agents previous interactions and feedback. It assumes that an agent will rate others in a binary way after each transaction. It shows that by using trust management framework, systems can mitigate malicious attacks. Similarly, Bao and Chen proposed a framework in which trust is dynamically updated [5]. TRM-IoT is a trust management framework based on fuzzy theory [7]. Trust in TRM-IoT is based on agents packet forwarding success rate in wireless sensor networks. It also calculates reputation based on agents local trust relationships. Similarly, [13] uses fuzzy theory to model trust among agents. Trust in [13] is divided into three levels based on three corresponding layers in IoT architecture: sensor layer, core layer and application layer. [8] proposes a trust management framework for serviceoriented architecture (SOA) based IoT systems. Trust is based on agents previous interactions and experiences. It uses distributed collaborative filtering to select trust recommendations. It dynamically adjusts the protocol s parameters for different environments. Also, it considers four types of malicious attacks. [22] considers that trust is related to specific functions. Agents cannot use previous experiences related to one function to infer trust for another function. Therefore, trust is contextaware. TACIoT considers multi-dimensional trust in access control for IoT systems [6]. Although there are many works about trust in IoT systems, most of them are designed for specific applications, e.g. service provision. These applications specific frameworks have a common limitation. They cannot be applied to other applications. In other words, they are dependent on specific assumptions, such as feedback must be available, agents social roles (or ownership) are known, and so on. Instead of considering specific factors to model trust, we proposed a general trust management framework based on the measurement theory, which can be adopted by various types of applications. III. TRUST MANAGEMENT FRAMEWORK The process of trustworthiness evaluation is similar to the measurement process [26]. People get an initial impression or evaluation about a given physical quantity by measuring it using appropriate equipment. It can be further updated or improved by using more precise equipment, combining different methods, or repeating the measurement. In IoT systems, agents can evaluate others in a similar way. For example, the subject can have prior knowledge or interactions with the object, and it can be further updated. The determining factors or environments for trustworthiness evaluation are dependent on applications. Also, the availability of humanto-human, human-to-device, device-to-human and device-todevice relationships is application dependent. A. Trustworthiness and confidence In our framework, we use two metrics trustworthiness and confidence to represent trust. Trustworthiness is the comprehensive summary of the subject s multiple measurements or evaluations towards the object. This summarization is very similar to the averaging of sample measurements in statistics; however, the concrete meaning of trustworthiness depends on specific applications. For example, it could be the quality of service, packet forwarding success rate, and so on. Each measurement is a single piece of information that contributes to the trustworthiness evaluation. Trust can be established through multiple environments, such as previous interactions, reputation, and so on. Apart from this, within each environment, it can have a single measurement or multiple measurements. Besides trustworthiness, confidence measures to what extent the subject is certain about the trustworthiness evaluation [26]. As trustworthiness is evaluated based on different measurements, e.g. the number of measurements can be different, confidence also varies. Given a set of trustworthiness measurements, the subject would have a distribution of measurements in a range around the summarized average trustworthiness. In such a scenario, confidence is related to the uncertainty in the measurement theory [14]. We calculate the standard uncertainty σ from the distribution and convert it into confidence as in Equation 3. The higher the standard uncertainty is, the lower the confidence is. We use Te D (m D e,c D e ) to denote the subject s direct trust towards the object. And we use m D e to denote the trust determined by environment e, c D e to denote the confidence. We assume both m D e and c D e are real numbers in the range of [0, 1]. Higher values represent more trustworthy or confident. We use x i to denote each trustworthiness measurement, and it can be converted from raw input to a real number in the range of [0, 1]. For each environment e, m D e, standard uncertainty σe D and c D e can be calculated in Equations 1, 2 and 3 accordingly. 1014

3 i=n m D i=1 E = x i N i=n σe D i=1 = (x i m D E )2 N (N 1) (1) (2) c D E =1 2 σ D E (3) As we indicated, trust can be build through several environments or factors [24] [5] [19]. Therefore, it is necessary to combine them together to determine trust. As each environment may play different roles, we use weighted mean to combine them. For a general purpose, we denote weights as w k. The overall trust To D (m D o,c D o ) can be calculated in Equations 4 and 5. Note that, the confidence is calculated following the error propagation theory. k=k m D k=1 o = w k m D k k=k k=1 w (4) k k=k c D o =1 2 k=1 (w k σk D)2 ( k=k k=1 w (5) k) 2 B. Trust Modeling Framework In IoT systems, i.e. social IoT, agents can be connected through multiple environments. For example, agents can be friends with each other in online social networks. Also, they may interact with each other in their working environment. We illustrate the trust modeling scenario in Figure 1. Environments can be some application layer platforms, such as social medias. Note that, as shown in Figure 1b, it is possible that an environment can be composed of a chain of subenvironments, i.e. physical links. Therefore, in general, it can be illustrated as in Figure 1c. The subject and environments which are inside the dash line rectangle in Figure 1c can be treated as a basic unit in trust network, as the subject must connect with the object through environments. We use T e to denote the subject s trust towards the environments, as not all the environments are equally trustworthy and important to the subject. Finally, the subject s overall trust T o can be determined as T o = f(t e1,t 1,..., T ek,t k ). f can be any general functions. In the following of this paper, we use weighted mean as indicated in Equations 4 and 5. C. Human-to-human trust relationship Human beings play an important role in IoT systems. They are connected through Internet, such as online social networks [26] [20] [21]. Therefore social networks, such as Twitter [12], can be integrated into IoT systems [18]. In the following, we list some possible environments that can be taken into account in determining human-to-human trust relationship. Friendship. Friendship is among one of the most widely used social relationships. In many online social networks, such as Twitter and Facebook, users are friends and commnucate with each other. Also, friendship is a straightforward measurement to measure trustworthiness in IoT systems [8]. Social interactions. Like in real life, human-to-human trust relationship can be accumulated by their social interactions. For example, in Twitter, users post interactive Tweets, e.g. reply, forward, retweet, towards specific users [20]. Each interaction can be treated as a measurement of trustworthiness. Reviews. In many online communities, users are able to write reviews towards other users. For example, in Epinions.com, users can express either like or dislike towards others. Besides this, they can rate others articles from un-helpful to very helpful. Those proposition and rating information can be used to model human-to-human trust relationship as well [26]. Common interest. Human beings having similar interests may also trust each other. Common interest is used to measure trustworthiness in [8] and [27]. D. Device-to-device trust relationship As indicated by [3], devices in IoT systems can have social relationships among them. Also, devices need to collaborate with other devices to accomplish complex tasks. Therefore, they can evaluate their partners through collaboration. We list some environments that may be considered in device-to-device trust relationship in the following. Behavior evidence. In IoT systems, devices are mainly designed to fulfill specific functionalities. For example, routers are designed to forward packages following predefined protocols. In other words, the subject device expects the object device to act in a certain way; however, it is possible that the object device fail to act as the subject device expected due to various reasons. Therefore, the subject can develop trust towards the object based on the satisfied evidence (positive evidence) and unsatisfied evidence (negative evidence). Data similarity. Data similarity can be used to evaluate data trustworthiness [9]. Similarly, it can be used in IoT systems to evaluate devices trustworthiness. Given a event, two devices data regarding this event may be different. For example, several sensors are used to measure the temperature in the same room; however, their measured temperature may be different. In such a case, if two devices have very closed results, we say they are trustworthy to each other regarding measuring temperature. Such type of trustworthiness is mainly used in the field of data process. Profile information. Like human beings, devices also can have profiles, including devices basic information, such as their manufacturers, owners, and working conditions. As pointed out by [3], devices owned by the same owner are considered more trustworthy than those belonging to other owners. 1015

4 (a) A trust modeling scheme including environments (b) An environment including multiple subenvironments Fig. 1. Trust modeling scheme (c) A general trust modeling scheme E. Human-to-device trust relationship Human-to-human and device-to-device trust relationships are two separated layers. Human-to-device trust relationship connects them, and also makes IoT systems integrating physical world with virtual world. Ownership. It is straightforward that ownership is an important component to connect human beings and devices. In IoT systems, devices have their owners, and in most cases, the owner trust their owned devices more than others. Observation. Human beings are able to evaluate devices performances. Note that, in some cases, both the owner and other users can observe the device s performance. Therefore, these observations can be treated as satisfied evidence or unsatisfied evidence. Capacity and feature. Human-to-device trust relationship can be affected by devices capacities and features. For example, in some cases, devices with long battery life are preferred. Also, devices prior knowledge, such as reputation, is a vital factor. For example, if the user knows the device s manufacturer, it may help her/him to form an initial impression towards the device. F. Device-to-human trust relationship As long as there exist interactions between devices and human beings, devices are also able to evaluate human beings performance. Similar to human-to-device trust relationship, devices can evaluate human beings performance as satisfied or unsatisfied evidence. Note that, we only list few possible (sub)environments or factors here. They are dependent on applications, e.g. the meaning of trust. Also, it has the limitation on their availability. Only if they are available in the application, they can be used to measure trust relationships. On the other hand, not necessarily all these factors are available and have to been considered in all the applications. IV. EXPERIMENTAL SIMULATION In this section, we use an example to illustrate the usage of our trust management framework in IoT systems. IoT systems are now widely used in chronic diseases therapy management, such as diabetes treatment [11]. For example, sensors are distributed along with patients to monitor their blood sugar levels and intake food. In Type 1 diabetes therapy, patients insulin dose is partially determined by their food intake [2]. Therefore, it is very important for doctors and patients themselves to know the meals carbohydrate content. Thanks to the wide deployment of smartphones, it is now very easy and convenient for patients to take photos for their meals. As indicated by Platemate [17], nutrition analysis can be done through three steps: tag, identify and measure. In the first step, food items are separated from each other based on the boundary. Then each food item is identified by category and name. In the measurement step, food items are estimate by size or count and calculated for nutrition using existing databases. The nutrition analysis is implement by human beings in Platemate through Amazon Mechanical Turk [17]. In our simulation, we assume that both human beings and devices are able to implement these three tasks. We show this scenario in Figure 2. Fig. 2. Flow of food image estimation We use Pa = {Pa 1,Pa 2,..., P a n } to denote the set of patients. Accordingly, A = {A 1,A 2,..., A k }, I = {I 1,I 2,..., I k } and M = {M 1,M 2,..., M k } represent tagers, identifiers and measurers. A. Trust relationship Human-to-human trust. As indicated in Section III-C, human-to-human trust can be determined by several subenvironments. We assume that people rate others based on their performance. Besides this, we assume people also have social interactions in social medias, e.g. Twitter. Their social interactions can be evaluated by sentiment analysis, which includes negative, positive and/or neutral [21]. Each interaction can be considered as a measurement of trust. Therefore, in this simulation we consider two environments: review ratings and social interactions. For simplicity, we assume that both of them are expressed in the range of [0, 1]. 0 represents distrust, while 1 represents trust. We use weighted mean to combine these two environments as in Equations 4 and 5. Device-to-device trust, human-to-device trust, device-tohuman trust. We only consider the quality of estimations 1016

5 for other three types of trust relationships. In other words, they only have one subenvironment. The value of each trust measurement is in the range of [0, 1] as well. B. Experimental setup In this simulation, we assume that there are 100 patients who are looking for nutrition analysis service. On the other hand, these patients themselves can act as tagers, identifiers and measurers. In addition to that, we assume that there are 100 devices that can act as tagers, identifiers and measurers correspondingly. To simulate the real scenario, we assume patients are from two groups and the types of their food are different, e.g. American food and Asian food. It is reasonable that patients are able to estimate their own food type better than another type. Therefore, for patients who belong to the same group, we assign higher review ratings (follow normal distribution) than patients from another group. In our following simulation, we assume devices are used to analyze group 1 s food. So, devices have high ratings for patients from group 1, and patients from group 1 have high ratings for devices. For the number of estimation ratings and/or social interactions between two agents, we assume that it follows uniform distribution. For each agent, we assume the number of its neighbors follows power-law distribution. More details about these parameters can be found in Table IV-B. TABLE I EXPERIMENTAL SIMULATION PARAMETERS P a, A, I, M 100 Number of neighbors P (4, 20), s = 1.2 Number of ratings or interactions between agents U(1, 50) Ratings in the same group N(0.8, ) Ratings in the different groups N(0.2, ) Ratings among devices N(0.8, ) Ratings between group 1 and devices N(0.8, ) Ratings between group 2 and devices N(0.2, ) Ratings between devices and group 1 N(0.8, ) Ratings between devices and group 2 N(0.2, ) C. Experimental results As two groups of patients are not good at estimating another group s food, we assume the estimation errors of patients from different groups are higher than patients within the same group. To simulate this, we assume the estimation error produced by patients within the same group is equal to 0.1, e psame = 0.1, while e pdiff = 0.8 for patients from two different groups. Similarly, for devices, we assume their error e d1 =0.1for other devices and patients from group 1. Patients from group 2 get error e d2 =0.8 from all the devices. As the nutrition analysis consists of three process, we calculate error step by step. For example, tagers get intermediate results from identifiers, and identifiers get results from measurers. For each agent in a specific step, the incoming error (error produced by the next step) is calculated as in Equation 6. Here I is set of incoming neighbors. Note that, we use weighted mean to combine incoming information, where weights are the product of m and c. Besides the incoming error, the agent itself also introduce an error e i. Therefore, the total error until this agent is updated as e i = e i + e in. i I e in = m i c i e i i I m (6) i c i We update error for each agent step by step. We begin from measuring step to identifying step, and then from identifying step to taging step. Finally, patients get the results from tagers. In the following of this paper, we only consider the error received by patients from group 1. In the above, we have illustrated that human-to-human trust has two environments review ratings and social interactions. In section IV-B, we only illustrate review ratings distribution. For social interactions, we consider two scenarios: 1) human beings social interactions have no group knowledge and follows the same distribution for two groups; 2). human beings social interactions have group knowledge such that they follow different distributions for different groups. In the first scenario, we assume that social interactions among human beings, regardless of which group they are from, follow the same normal distribution N(0.5, ). It means that all the social interactions are randomly distributed around the neutral trustworthiness. In the second scenario, we assume that human beings social interactions follow the normal distribution N(0.8, ) if they are from the same group. While two human beings are from different groups, their social interactions follow the normal distribution N(0.2, ). In other words, social interactions also capture the group difference. As we use weighted mean to combine review ratings and social interactions for human-to-human trust relationship, we run simulations (repeat 10 times) that use different weights (w r and w s ) for them. Higher is the weight, more important is the corresponding environment. The results are shown in Figures 3 and 4 for two scenarios separately. Compared with the cases using trust as in Equation 6, we also run simulation without using trust. In such a case, instead of using the product of m and c as weight, we calculate the incoming error just using the mean. From Figures 3 and 4, we can see that overall by using trust, patients in group 1 receive less error than without using trust. Therefore, it is necessary and helpful to incorporate trust in such IoT systems. Note that, as the minimum error for each step is 0.1, the minimum overall error for the whole process is 0.3. Furthermore, among 100 patients, we increase the number of group 2 patients from 0 to 50 (which is half of the total patients). We can see that, by increasing the number of group 2 patients, group 1 patients received error also increase. This is because group 2 patients produce higher error than group 1 patients; However, by using different weights for social interactions and review ratings, we get different results. In the 1017

6 Patinets (in Group 1) received error Patinets (in Group 1) received error Mean without using trust wr=0.0,ws=1.0 wr=0.2,ws=0.8 wr=0.4,ws=0.6 wr=0.6,ws=0.4 wr=0.8,ws=0.2 wr=1.0,ws= Number of patients from Group Fig. 3. Social interactions have no group knowledge Mean without using trust wr=0.0,ws=1.0 wr=0.2,ws=0.8 wr=0.4,ws=0.6 wr=0.6,ws=0.4 wr=0.8,ws=0.2 wr=1.0,ws= Number of patients from Group 2 Fig. 4. Social interactions have group knowledge first scenario, only review ratings reflect the group difference. Therefore, more weight on review ratings can help to reduce the received error. While in the second scenarios, both review ratings and social interactions capture the group difference. From Figure 4, we can see that different weights (w r from 0.2 to 0.8) have very similar performance; However, we also see that by only using either social interactions or review ratings (the lowermost two lines), it can achieve better results. From our simulations, we can see that how to select weights for different environments are important. In our cases, assigning more weights on review ratings achieve better performance; However, in other cases or applications, it is possible that weights between 0 and 1 can achieve the best performance. Therefore, weights selection is application dependent. D. Attacks simulation In this section, we first introduce another type of trust global trust or reputation. In the above, we have already seen that each agent has a set of neighbors. It is possible to combine all these neighbors opinions, and form a global opinion as the whole community towards that agent. We call it global trust, while some other works may refer it as reputation. Correspondingly, trust mentioned in the above sections can be referred as local trust. Suppose that agent A has a set of neighbors N, who hold trust evaluations towards A. Then we calculate A s global trust as in Equations 7 and 8. m g is the weighted mean of all the neighbors opinions, where weights are 1. It means that more confident opinions have higher σi 2 weights than less confident opinions. m g = i N 1 σi 2 i N 1 σi 2 c g =1 2 m i 1 i N σ2 i Global trust can be even seen as an environment for determining local trust, where weighted mean can be used as before. In other words, when an agent evaluate other agents trust, it also takes their global trust into account. By incorporating trust into IoT systems, one of the main purpose is to help systems to detect or mitigate attacks. Here, we simulate two types of attacks. In the first case, we assume attackers do not care about their global trust, they only concern the victims that they have connections. Therefore, attackers global trust is very low; However, their local trust is high in the victims point of view because of their previous camouflaged good behaviors. In the second case, we assume that attackers know the importance of their global trust such that they have a set of colluders who increase attackers global trust; However, attackers do not pay attention to specific victims. Therefore, their local trust is low in victims point of view. Here we simulate the scenario where there exist 20 victims and 10 attackers. Victims are connected with both attackers and other normal agents, the number of their neighbors follows the power-law distribution as before. To simulate high and low global trust for attackers, we assume that there are different number of colluders in two cases. In the first case (low global trust and high local trust), we assume there is no colluders. And for high local trust from victims points of view, we assume both victims review ratings and social interactions follow normal distribution N(0.8, ). In the second case (high global trust and low local trust), we assume there are 90 colluders, and both victims review ratings and social interactions follow normal distribution N(0.2, ) for low local trust. Note that those colluders only connect with the attackers, but not connect to any victims or other normal agents. To combine social interactions and review ratings, we use the weights w r = w s =0.5. We run the simulation for different weighs on local (w l ) and global trust (w g ) in Figure 5. We can see that weights play opposite roles in two types of attacks. In the first case, more weights on global trust is preferred as global trust reflects attackers real malicious behaviors and can help to mitigate their effect. While in the second case, local trust is more important. Also, we compare the results with the case which does not use trust. From Figure 5, we can see that knowing attackers behavior patterns is very important. Facing two different types of attacks, weights have opposite performance. Therefore, corresponding weights on different environments following the attackers behavior pattern can be used to mitigate attacks. This will be part of our future work. (7) (8) 1018

7 Victims received error Mean without using trust Type1: attackers have low global trust, high local trust Type2: attackers have high global trust, low local trust Weights on the global trust Fig. 5. Weights for attacks simulation V. ACKNOWLEDGMENTS This work was partially supported by National Science Foundation under Grant No VI. CONCLUSIONS In this paper, we propose a general trust management framework for IoT systems. we use two metrics to measure trust: trustworthiness (m) and confidence (c). Our framework is based on measurement theory, which consider agents trust evaluations or interactions as measurements. Moreover, we consider that agents can evaluate each other through different environments. We also list several possible environments and factors that may determine trust relationships in IoT systems. Our trust management framework is general such that it can be deployed in many IoT systems. To better illustrate it, we run a food nutrition analysis in diabetes treatment as an example to show the benefit of our trust management framework. The results show that by using trust, it can help patients to filter out inaccurate information or mitigate its effect. Also, we illustrate two possible types of attacks. By showing these two types of attacks, we can see that selecting and weighting environments is very important for IoT systems. REFERENCES [1] Sarita Agrawal and Manik Lal Das. Internet of thingsa paradigm shift of future internet applications. In Engineering (NUiCONE), 2011 Nirma University International Conference on, pages 1 7. IEEE, [2] Marios M Anthimopoulos, Lauro Gianola, Luca Scarnato, Peter Diem, and Stavroula G Mougiakakou. A food recognition system for diabetic patients based on an optimized bag-of-features model. Biomedical and Health Informatics, IEEE Journal of, 18(4): , [3] L. Atzori, A. Iera, and G. Morabito. Siot: Giving a social structure to the internet of things. Communications Letters, IEEE, 15(11): , November [4] Luigi Atzori, Antonio Iera, and Giacomo Morabito. The internet of things: A survey. Computer networks, 54(15): , [5] Fenye Bao and Ing-Ray Chen. Dynamic trust management for internet of things applications. In Proceedings of the 2012 international workshop on Self-aware internet of things, pages 1 6. ACM, [6] Jorge Bernal Bernabe, Jose Luis Hernandez Ramos, and Antonio F Skarmeta Gomez. Taciot: multidimensional trust-aware access control system for the internet of things. Soft Computing, pages 1 17, [7] Dong Chen, Guiran Chang, Dawei Sun, Jiajia Li, Jie Jia, and Xingwei Wang. Trm-iot: A trust management model based on fuzzy reputation for internet of things. Computer Science and Information Systems, 8(4): , [8] I.-R. Chen, J. Guo, and F. Bao. Trust management for soa-based iot and its application to service composition. Services Computing, IEEE Transactions on, PP(99):1 1, [9] Chenyun Dai, Dan Lin, Elisa Bertino, and Murat Kantarcioglu. An approach to evaluate data trustworthiness based on data provenance. In Secure Data Management, pages Springer, [10] V Hu, D Ferraiolo, R Kuhn, A Schnitzer, K Sandlin, R Miller, and K Scarfone. Attribute based access control definition and consideration. NIST Special Publication, pages [11] Antonio J Jara, Miguel A Zamora, and Antonio F Skarmeta. An internet of things based personal device for diabetes therapy management in ambient assisted living (aal). Personal and Ubiquitous Computing, 15(4): , [12] Matthias Kranz, Luis Roalter, and Florian Michahelles. Things that twitter: social networks and the internet of things. In What can the Internet of Things do for the Citizen (CIoT) Workshop at The Eighth International Conference on Pervasive Computing (Pervasive 2010), pages 1 10, [13] Gu Lize, Wang Jingpei, and Sun Bin. Trust management mechanism for internet of things. Communications, China, 11(2): , Feb [14] Douglas C Montgomery and George C Runger. Applied statistics and probability for engineers. John Wiley & Sons, [15] Huansheng Ning, Hong Liu, and Laurence T Yang. Cyberentity security in the internet of things. Computer, (4):46 53, [16] Michele Nitti, Roberto Girau, Luigi Atzori, Antonio Iera, and Giacomo Morabito. A subjective model for trustworthiness evaluation in the social internet of things. In Personal Indoor and Mobile Radio Communications (PIMRC), 2012 IEEE 23rd International Symposium on, pages IEEE, [17] Jon Noronha, Eric Hysen, Haoqi Zhang, and Krzysztof Z Gajos. Platemate: crowdsourcing nutritional analysis from food photographs. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages ACM, [18] A.M. Ortiz, D. Hussein, Soochang Park, S.N. Han, and N. Crespi. The cluster between internet of things and social networks: Review and research challenges. Internet of Things Journal, IEEE, 1(3): , June [19] Charith Perera, Chi Harold Liu, and Srimal Jayawardena. The emerging internet of things marketplace from an industrial perspective: a survey. IEEE Transactions on Emerging Topics in Computing, [20] Yefeng Ruan, L. Alfantoukh, and A. Durresi. Exploring stock market using twitter trust network. In Advanced Information Networking and Applications (AINA), 2015 IEEE 29th International Conference on, pages , March [21] Yefeng Ruan, L. Alfantoukh, A. Fang, and A. Durresi. Exploring trust propagation behaviors in online communities. In Network-Based Information Systems (NBiS), th International Conference on, pages , Sept [22] Yosra Ben Saied, Alexis Olivereau, Djamal Zeghlache, and Maryline Laurent. Trust management system design for the internet of things: a context-aware and multi-service approach. Computers & Security, 39: , [23] S Sicari, A Rizzardi, LA Grieco, and A Coen-Porisini. Security, privacy and trust in internet of things: The road ahead. Computer Networks, 76: , [24] Zheng Yan, Peng Zhang, and Athanasios V Vasilakos. A survey on trust management for internet of things. Journal of network and computer applications, 42: , [25] Lina Yao, Quan Z Sheng, Anne HH Ngu, Helen Ashman, and Xue Li. Exploring recommendations in internet of things. In Proceedings of the 37th international ACM SIGIR conference on Research & development in information retrieval, pages ACM, [26] Ping Zhang and Arjan Durresi. Trust management framework for social networks. In Communications (ICC), 2012 IEEE International Conference on, pages IEEE, [27] Cai-Nicolas Ziegler and Jennifer Golbeck. Investigating interactions of trust and interest similarity. Decision support systems, 43(2): ,