1.1 Motivation

In cognitive radio-assisted IoT networks, two important aspects have to consider: first is task allocation and the second is cooperation among cognitive IoT nodes. Cognitive nodes need to distribute the sensing task among the neighbouring sensing nodes for distributed communication. Numerous task allocation approaches have been studied in state-of-the-art traditional wireless networks [7]. However, the existing approaches cannot be directly applied to IoT networks due to constrained node resources and special wireless scenarios. The design of the task allocation approach for cognitive IoT is ongoing research [8]. Subsequently, how mutual cooperation can be achieved in between cognitive node and IoT nodes has to be explored to have interference-free communication. Recently, cooperation among the wireless nodes to exchange both the application data and spectrum-related information has been explored in [9]. In [10], the performance of the IoT network is being enhanced by allowing the primary user’s licensed channel access to an unlicensed user with the condition that the unlicensed user cooperates by acting as a relay for edge licensed nodes. Since the network adapters are more and more modifiable, it allows selfish nodes to tinker with the wireless interface and maximizes their own benefit. This leads to natural and important design issues like: how to deal with selfish nodes? and what if the majority of nodes are selfish? Thus, this paper works on two important issues namely 1. Nodes in Cognitive IoT need to distribute the sensing task among the neighbour sensing nodes with unknown nature and 2. To deal with the selfish nodes within the network.

1.2 Limitations in existing work

With respect to a task allocation, studies focus to select a group of nodes to complete a global application cooperatively by using genetic algorithm (GA) [11] or partial swarm optimization (PSO) [12]. Other models are proposed by modifying PSO, GA where the models suppose to know all the network parameters in advance. However, in a network with autonomous nodes whose behaviour is based on their current state, it is difficult to solve this requirement in a real-world IoT scenario.

State-of-the-art research has shown that selfish nodes seriously deter the performance of the network [13–15]. In general, there are three ways that have been used to deal with selfishness in ad-hoc networks in the literature a) precautions of selfish nodes [16]; b) detection of selfish nodes [17, 18]; c) avoidance of selfish nodes [19]. Each of these approaches has its own unique advantages and disadvantages. Lin et al. [20] proposed CONFIDANT algorithm which acquires reputation value and excludes network method to penalize non-cooperative nodes. In [21], security mechanisms have been implemented to stop nodes from tampering with vital information. Most of the existing research solves the selfishness issue in IoT nodes by using model-based data forwarding scenarios i.e., a node is considered to be selfish in case it does not forward data to the neighbour node. Otherwise, it is considered as the cooperating node. In [22], author addresses the limitations of IPv6 RPL in IoT networks’ extreme conditions and introduces Dynamic-RPL. With minimal changes to RPL, Dynamic-RPL ensures high network performance, adaptability to adverse conditions, and efficient topology management. Different from previous studies, in this work, we address the selfishness in IoT nodes by analyzing the node’s unwillingness to cooperate with the cognitive node in identifying the vacant licensed bands.

It is noteworthy that selfishness in cognitive radio-assisted IoT scenarios is different from selfishness in the traditional wireless network in terms of channel allocation. In a traditional network, when node acts selfishly, it can save its energy and prolong the battery lifetime. But in the case of selfishness in cognitive radio wireless networks, there can be an interference with primary users(PU) by not co-operating the data transmission in an opportunistic licensed channel [23]. Even though literature work addresses the issue of selfish nodes in a wireless network, little work has been done in terms of setting specific rules to regulate in terms of energy conservation and truthfulness. Thus, how to identify the selfish node activity which results in increasing the lifetime of IoT nodes will be addressed in this paper.

1.3 Proposed solution

From Fig 2, cognitive nodes are going to act as cluster heads for the IoT nodes. Cognitive nodes distribute the sensing task to the neighbour IoT nodes while the participating IoT nodes can contribute to the sensing licensed channel. To address task allocation and selfish node detection, this paper proposes a game theory-based cooperative model for IoT nodes to participate in sensing tasks initiated by cognitive nodes. Firstly, we use the auction model to assign tasks and assign a reward to IoT nodes. Subsequently, for enhancing cooperation, the cognitive node rewards selfless IoT nodes based on shapely value and punishes the selfish IoT nodes based on trust value. The key concepts of this paper are summarized as follows:

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. System architecture: Cognitive-IoT network with selfish and malicious nodes.

https://doi.org/10.1371/journal.pone.0309123.g002

In this paper, we propose a cooperative-based model for users in fog and cloud computing to participate in tasks published by an application platform. In the formulation process, the participating users can contribute to the tasks, and the tasks will pay the users rewards. The reward to the users in this paper is defined to have a relationship with region density. The proposed idea is a two-step approach to deal with selfish nodes. In the first step, we use a prevention mechanism where the cognitive node scans the network frequently to discover the selfishness. In the second step, we use an avoidance mechanism that will promote cooperation among the IoT nodes using a reward and punishment strategy. Thus we are motivated to develop a mechanism for selfish nodes in the cognitive IoT network using a game-theoretic approach to improve the cooperation with constraints on energy. In this paper, we have two mechanisms namely 1) frequency of monitoring and 2) promotion of cooperation through punishment and reward [24]. The cooperative mechanism uses two partial and full monitoring to understand the behaviour of the nodes. The key concepts of this paper are summarized as follows:

1. (1). The players of the game are IoT nodes that enable to choose in between two strategies. One is to cooperate with the cognitive node and the other is to defect. Cooperation and defection have an impact on the quality of the licensed channel in terms of throughput.
2. (2). Auction mechanism is used by a cognitive node for the distribution of sensing tasks among the IoT nodes.
3. (3). Monitoring the behaviour of nodes requires energy, thus monitoring nodes need to know optimal monitoring frequency to save energy and identify the nodes’ behaviour.
4. (4). To promote cooperation in the network and reduce selfish nodes, the identification of selfish nodes is necessary. Partial and full monitoring is used to punish selfish nodes and reward cooperative nodes.
5. (5). Proposed ‘Malicious node metrics’ to identify and reduce the occurrence of Malicious nodes.

The rest of the paper is organized as follows: Section 2 introduces the theoretical basis and specific process of the PR-CSDF mechanism. Section 3 provides the pseudocode of proposed algorithms in the paper. The performance of the proposed work is evaluated in Section 4 and Finally, Section 5 concludes the paper with outlines of our future work.

2.1 Problem scenario

The network topology consists of the secondary network consisting of IoT nodes, cognitive nodes(CNs) co-existing with the primary network with primary users(PUs) as shown in Fig 2. Cognitive nodes are allowed to access the PU licensed channel when it is not used by primary users(overlay access) or with limited power without causing interference to the ongoing PU transmissions(underlay access). However, cognitive nodes need to sense the licensed channel to identify spectrum opportunities and avoid interference with PUs. Sometimes CN cannot detect the presence of primary transmitters due to hidden terminal problems. Thus, cooperative strategies are preferred for sensing primary user transmissions. Here, the CN node distributes the sensing tasks to the nearby IoT nodes and rewards them with the opportunity to transmit in a licensed channel. This reward is the motivation for IoT nodes, as unlicensed channels are not reliable for data transmissions and IoT nodes need to do multiple re-transmission. Thus, cooperation gives benefits to both cognitive nodes and IoT nodes, notation used are given in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Basic notations.

https://doi.org/10.1371/journal.pone.0309123.t001

2.2 Overview of proposed solution

In Fig 3, to capture the interaction between CN and IoT nodes, we have used multi-auction where CN is the auctioneer and IoT nodes are bidders (2.4.2). The cognitive node which is auctioneer broadcasts the sensing task to the neighbouring IoT nodes. Rational IoT nodes which are bidders send the bid B = B₁, B₂, B₃, B₄, …., B_n, B_i denotes the amount of sensing time allocated to cognitive node. Every IoT node sends its bid values to obtain rewards from CN with the promise of its contribution. More specifically, the sensing time they are willing to offer. By analysing the payoff of CN and IoT nodes, the upper bound and lower bound of feasible bids are calculated by energy consumption in the licensed and unlicensed channels. In such a way, tasks will be assigned to selfless nodes in most cases. But, selfish IoT nodes in the network impact the system performance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. System process.

https://doi.org/10.1371/journal.pone.0309123.g003

A game theory model based on the IoT node’s behaviors(i.e, selfish or selfless) and the corresponding payoff is built. Using this model, we understand that both reward and punishment strategies need to be applied. It provides the direction for the next modeling. Meanwhile, the optimal monitoring frequency formula is presented with the analysis process of the payoff matrix. This formula can help CN to avoid unnecessary energy consumption while dynamic frequency is employed in every round.

The cooperation process which includes reward and punishment mechanisms has three aspects. At First, indirect trust values will put forward to adjust the frequency. Secondly, nodes are rewarded with additional sensing using shapely values. In the end, over a period of time, CN analyzes the data sent by its own monitoring data, i.e., its direct trust value, and punishes the selfish IoT nodes. Trust value is a credit record for previous behaviors. CN decides whether or not to punish and which degree of punishment will be chosen by direct trust value and change the frequency value by indirect trust value. To deal with selfish IoT nodes, a reward and punishment strategy is applied. Here, CN monitors selfish IoT nodes using the optimal frequency formula in order to avoid unnecessary energy consumption. A game theory model based on the behaviors and the corresponding payoff is built to find the best strategy to deal with selfish IoT nodes. Both reward and punishment should be applied according to the analysis result. Furthermore, the optimal frequency formula is given in the meantime.

In the course of the system’s operation, it may be more likely to develop malicious nodes as the number of selfish nodes increase. The exert of malicious nodes identification will reduce the benefit to transform into malicious nodes, and thus, reduce the likelihood of the appearance of malicious nodes and affecting the system stability. Combination with cluster algorithm, similarity, and maximum clique makes it possible to quantify the clique phenomena and monitor them subsequently.

2.3 Game process

There are two ways to promote the cooperation between CN and IoT nodes: reward and punishment. It is obvious that cooperation cannot be achieved just with punishment and no reward. However, monitoring is required for selfish nodes, which is energy-consuming. Thus, the question arises whether CN needs to do monitoring or not.

Firstly, we focus on the behavior of cognitive nodes and IoT nodes. Since, the main objective is cooperation and energy saving, the following payoff matrix (Table 2) is constructed by the analysis of energy loss.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Payoff matrix of interactions between CN and SN nodes.

https://doi.org/10.1371/journal.pone.0309123.t002

I_n signifies the income with no monitoring. Likewise, I_p, I_f signify that with partial and full monitoring respectively. P signifies that the punishment for IoT nodes. It’s shown in Table 3. C_fc signifies that the consumption of finding the channel. R signifies that the reward for the contribution of IoT nodes. C_f means that the consumption of forwarding data for IoT nodes. C_pm, C_fm mean that the consumption of partial monitoring and full monitoring, respectively. I_gc means that the income of getting channels.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Relations between direct trust value and punishments.

https://doi.org/10.1371/journal.pone.0309123.t003

For each unit contribution, there is no difference between the defection portion of incomplete cooperation and defection itself. Similarly, they are the same between the cooperation portion of incomplete cooperation and cooperation itself. On this basis, in Table 2, we define “Defection” as complete non-cooperation, that is to say, those IoT nodes will make no sensing. Likewise, ‘Cooperation’ is defined as complete cooperation.

In addition, Partial monitoring refers to considering indirect trust value and Full monitoring refers to considering both direct trust value and indirect trust value. (Shown in section 2.5.1), Fig 4.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Full process of system model.

https://doi.org/10.1371/journal.pone.0309123.g004

1. (1). No monitoring (no reward and no punishment):
   To analyse this topic concisely, we can simplify Tables 2 to 4.
   Given the benefits of selfish behaviors, when CN does not monitor IoT nodes, the dominant strategy of IoT nodes will be “Defection”, and (no monitor, defection) is the nash equilibrium. This is harmful for the cognitive node and contrary to the rule of global benefit maximization. So we need to take precautionary measures to change this situation. The next issue is which strategy to choose, only motivation, only punishment, or both of them.
2. (2). With the only reward (no punishment):
   In this item, simplifying Tables 2 to 5 first.
   From the perspective of energy consumption, CN should not use more energy to reward than the required energy used to sense i.e., cognitive node needs to gain positive earnings. That is, R + C_f < I_gc and C_fc > R.
   Although the earnings from cooperation increase, the dominant strategy is still ‘Defection’. So motivation alone is not enough.
   But, it doesn’t mean that stimulation is utterly useless. First, SNs have some data that they want to send out, which require more energy for transmission in an unlicensed channel. The real values of R are not the same for all SNs. Consumption is only a perspective, just more objective, and representative. Besides, this is a repeated game, in the real world, according to experiments in behavioural game theory, we can know that users tend to show their side of goodness when they assume that others are good-intended. Nodes will be inclined to choose cooperation overall.
   Simulation is effective but not enough. Meanwhile, we should consider punishment. Because through this way, we can let the nash equilibrium closer to cooperation.
   The above discussions provide insights in identifying selfish SNs by monitoring.
3. (3). With both reward and punishment:
   In summary, using both punishment and motivation is more appropriate, like Table 2.
   In this scenario, as the frequency of monitoring increases, the income of choosing to be selfish decreases. When the frequency reaches a number, the income of being selfish or selfless will be the same. Thus, being selfless is a dominant strategy, now the question is how often should cognitive nodes monitor.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Simplified payoff matrix for ‘No monitoring’ strategy.

https://doi.org/10.1371/journal.pone.0309123.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Simplified payoff matrix for ‘No punishment’ monitoring strategy.

https://doi.org/10.1371/journal.pone.0309123.t005

We assume that there exists IoT nodes and a single cognitive node(CN) in the cognitive IoT network which is allowed to access the PU licensed channels with mutual cooperation. The licensed channel is separated into M sub-bands, unlicensed users (CN and SN) can access the vacant M sub-bands in the absence of the licensed user. For the cognitive node to access the licensed channel it needs to sense all the M bands in order to make the access decision. Due to limited battery power, IoT nodes and CNs try to conserve their energy. In the proposed model, the CN distributes sensing tasks to IoT nodes surrounding it and provides them with access to the licensed channel. In this network, we assume that number of IoT nodes is K which are divided into two groups. cooperators (Cg) help in sensing the licensed channel and defectors(Dg) do not help in sensing the licensed channel. The probability of false alarm PF is equal to all IoT nodes.

After one round, the cognitive node will count the contribution of every IoT node. Later, the cognitive node will use shapely values to calculate the assistance’s importance. In the next round, the cognitive node allocates redundant transmission time based on individual contribution. The node whose credit degree is lower than the threshold will be refused to cooperate.

2.3.1 The frequency of monitoring.

Monitoring behaviour of IoT nodes is energy-consuming. So the question is how often should we monitor?

Table 2 illustrates that partial monitoring can bring out some benefits, and based on Algorithm 4, we know the energy consumption of indirect trust value is extremely little. With indirect trust value, it can deter selfish behaviors and it is effective. When there is no mechanism to fill up the gap between twice Full monitoring then the frequency of Full monitoring would increase or taking other measures to reduce the number of selfish behaviors. In other words, it reduces the benefit of defection. But these two ways are more energy-intensive than Partial monitor. Therefore, Partial monitoring will run all the time and we will calculate the frequency of Full monitoring in order to save the node energy.

When the expectation of earnings from selfish and selfless behaviors are equal then IoT nodes will have no inclination to be selfish. This is because selfish behaviors will bring no extra earnings compared with selfless behaviors and there is a risk of being regarded as selfish nodes. When a node’s direct trust value (Shown in section 2.5.1) is lower than the threshold then it will be regarded as a selfish node and the cognitive node will refuse to cooperate with it. In this case, defection has become a weakly-dominated strategy.

Further, we decide to calculate frequency by considering no monitoring and full monitoring. Indirect threats are not always effective in the real world by the reason of incompletely rational. In order to achieve our goal mentioned before, thinking about the worst situation is proper. In this sense, the natural corollary is that we can use I_n to replace I_p and I_f, and now it’s the income of no sense.

As stated previously, it’s feasible to make Partial monitoring in every time slot. Thus, when we consider the frequency of monitoring, it should be replaced with Full monitoring with the following payoff matrix Table 6 consideration.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Final payoff matrix with only no monitoring and full monitoring strategies.

https://doi.org/10.1371/journal.pone.0309123.t006

In order to make nodes selfless, we should make it impossible to profit by choosing to be selfish. When the income that was obtained by the node to be selfish or selfless are equal then the IoT nodes have no incentive to be selfish. In this sense, their expectations are equal.

Follow the above train of thought, we suppose that the possibility of Full monitoring is P, then the possibility of No monitoring is 1 − P. (1)

On the basis of previous discussion, we also know:

Therefore, Eq (1) always has a solution, and only one solution.

The optimal frequency theoretically is .

According to Table 3, P is regarded as 3C_fc, and R will be calculated based on the contribution of last round and extra reward (Section 2.5.2). Moreover, an extra reward can be preset or changed dynamically depending on the situation. These assumptions correspond with real-world scenarios.

2.4 Assign sensing tasks

In a IoT network with selfish nodes, when the sensing task gets allocated to the nodes without considering their nature then the tasks assigned for selfish nodes will not be accomplished. Thus, tasks need to be assigned to nodes considering their nature where the auction mechanism meets this requirement.

2.4.2 Auction model for assigning sense tasks.

A set of tasks to be assigned is represented as A = A₁, A₂……A_m with m = 1, 2…… The task which is an auction item is actually reward converted from a basic resource necessity for running tasks. IoT nodes will bid on the resource of running tasks, generally the sensing time, which also means the minimum contribution required of bidders(IoT nodes) in one round. And B = B₁, B₂……B_n with n = 1, 2….. denotes the bids of IoT nodes. At the same time, IoT nodes have different private values for the same auction item. Due to divergent channel quality, various demand of forwarding, its own residual energy and something else, there is different degrees of importance attached by IoT nodes. V = V₁, V₂……V_n with n = 1, 2….. is used to express this diversity. Hence, the bids mainly depend on their own private values.

Firstly, we consider using an auction game to assign the task of sensing channel. Here, the bid represents the minimum contribution standard of a node. See Table 7 for the details. The Frame format of cognitive nodes auction can be seen in Fig 5. We propose a co-operation promotion mechanism, where each frame is divided into fixed time intervals. At time T_auc, cognitive node (CN) conducts auction in multiple direction. The IoT nodes that receive the auction request bid using cost formulation, i.e, based on the energy level, wastage of energy in unlicensed transmission due to high interference. At the second part of the Time frame T_s, the cooperating IoT nodes perform PU spectrum sensing. Next, during time T_r the cooperating IoT nodes send data to CN, where CN fuses the data received from the IoT nodes and makes a decision on the channel. When the channel is available, CN transmits its own data during time T_cr. While IoT node data is in time T_SN, which is the reward for the cooperation.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Cognitive-WSN network with selfish and malicious nodes.

https://doi.org/10.1371/journal.pone.0309123.g005

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Terminology interpretation.

https://doi.org/10.1371/journal.pone.0309123.t007

Both CN and IoT nodes desire to reduce energy consumption. Comparing with the sensing channel by CN itself, it will not consume more required energy to reward in the auction stage. In this sense, CN will be motivated to hold an auction only when it can save its energy after this stage. Further, CN will not accept the bids which are lower than the required energy that CN senses channel by its own self. This is the lower bound of bids for every SN. Similarly, IoT nodes also hope that they can cost less energy than forwarding in unlicensed channels. We assume that the bids of IoT nodes obey Normal Distribution subjected to [E_l, E_ul] whereas E_l, E_ul represent the energy cost in licensed or unlicensed channel (Section 2.4.3).

For the discussion of Section 2.4.4-Lemma2, we decide to add reserve price in the early stage and call it off in the later.

2.5 Cooperation mechanism

To improve cooperation and robustness, the core issue is with selfish nodes. First, CN should choose which strategies need to be selected. Thus, we proposed mechanisms for strategies selected separately can be seen in Fig 6. Later, the case that majority of nodes are selfish is considered final.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Identify selfish nodes.

https://doi.org/10.1371/journal.pone.0309123.g006

2.5.1 Trust value.

Before diving into the frequency of monitoring, quantitative analysis for cooperation need to be solved, That is, the rewards and punishments process and rules. We can calculate and analyse the payoffs in various cases where monitor mechanisms and risks are clearly stated.

In this section, we describe the proposed method, which deals with the challenge that, most of the IoT nodes only provide information about a few channels.

1. (1). Direct trust value:
   In every round, the cognitive node decides to monitor at the frequency which has been calculated at the end of the last round by 2.3.1. In round 1, the cognitive node monitors all IoT nodes.
   When the cognitive node finds nodes as selfish behavior, then the IoT node’s direct trust value will be reduced by 1. Otherwise, the value will be increased by 1. We program multi-threshold in order to make the system more fair and rational. (have reduced similarity)
2. (2). Indirect trust value:
   Sense nodes have their interactions, so they have their own comments on other nodes. To make our model fair, we should consider these comments. In our paper, it is called ‘indirect trust value’ (9) We assume that the amount of successful cooperation is N_c. On the contrary, the amount of selfish behaviors is N_d. Pro_ij indicates the probability of being trusted.
   In this scenario, as selfish behavior happens, the Pro reduces. It means, the also reduces.
   In addition, using Eq (9) is propitious to prompt interactions among IoT nodes because the number of interactions is a key parameter in the calculation of indirect trust value.
   And assuming This matrix is convergent. (With the continuous implementation of our system, the behaviors will tend to be selfish. It will cause smaller standard deviation and accelerate convergence rate.) Assuming the convergence matrix is following: (10) It’s a sparse matrix. So we plan to smooth it first. Therefore, we change (10) into I is the identity matrix.
   The percent of nodei points that the location of in all direct trust values when they are from smallest to largest. Because a node with higher is less possible to choose selfish behaviors, the lower weight (Table 8) is given to them.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Standard for assigning weights.

https://doi.org/10.1371/journal.pone.0309123.t008

2.6 Identify malicious nodes

The reliability of indirect trust value will decrease with the increase in the number of selfish SNs and these SNs are more likely to form cliques and cheat CN together. Meanwhile, the absence of a mechanism to regulate such cases will reduce the robustness of the system.

In this paper, we define “malicious node” as a IoT node that provides false indirect trust value to cognitive, that is, giving members of its clique high values but low values to others deliberately.

In the context of such a scenario, selfish nodes will be given low indirect trust values by other IoT nodes and the frequency of monitoring from cognitive will increase. But if there is no strategy to identify fake values and punish providers, selfish nodes are tilting toward to form a clique and be malicious nodes.

Accordingly, it’s indispensable to build a model to prevent IoT nodes from being malicious nodes. To this end, we design the following method:

1. (1). Clustering indirect trust values:
   First, K–meansClusteringalgorithm will be used to classify them into two groups: highvaluegroup and lowvaluegroup. Then, the initial point is set between 0.05 and 0.95.
   In contrast to other clustering algorithms, it costs lower energy and it’s faster by reason of simplicity. It’s not wealthy to achieve a better clustering with much more energy consumption. The main weakness with this algorithm is following:
   
   1. In general, people don’t know clearly how many initial points we should cluster.
   2. It’s difficult to learn which point should be confirmed at the beginning.
   3. It’s an input-sensitive algorithm for outliers, noises, isolated points, and so on.
   In the context of our paper, we should cluster these values into two groups and we know the ballpark values of initial points. Besides, the outliers things are very rare and have very small negative effects on the system except leaving a few malicious nodes out or adding selfless nodes wrongly, but they do not matter. From what has been discussed above, we can reasonably arrive at the conclusion that K–meansClusteringalgorithm is a proper choice.
2. (2). Quantify and analyze:
   In the beginning, we quantify the relationship between IoT nodes and these two groups by the following function. Due to the aforementioned feature of malicious nodes, the IoT nodes in the same clique will have the same indirect trust value table. In this context, we can analyze through calculating the similarity of the value table among these nodes. ‘Cosine Similarity’ is an attempt to account for the similarity between two vectors.
3. (3). Clique partition:
   Thirdly, regarding all nodes as points in a graph and creating connections lines between two points if their similarity is more than 90%. Thus, it’s easy to find the cliques in this graph, and the cognitive node will choose a node in every clique, with members nodes larger than 5 randomly, to judge whether it’s a malicious node. If so, the cognitive node will monitor all nodes in this clique. Otherwise, the cognitive node will stop monitoring.
   When the number of members in a clique is smaller than 5, we need not care about it. It is possible that there are a few selfless neighbor nodes and they communicate with similar SNs. As a result, those neighbor SNs have similar behaviors, trust values, and so forth although they are willing to cooperate with CN.

3.1 Auction

Algorithm 1 The Second Price Saled Auction

Require: the bid of IoT nodes

Ensure: the price (forward data amount)

 1: while the beginning of every round do

 2:  Cognitive node calculates its budget for the arriving task which will be treated as auction items and reserve nodes.

 3:  The auctioneer invites the public bidding, broadcasts starting price for auction items.

 4:  Sense nodes compare it with remain energy to decide whether to participate or not and calculate the cost of unlicensed channel based on our proposed cost formulation.

 5:  The seller sets the bid value.

 6:  if If at least one bid is higher than reserve price then

 7:   The winner would be the seller with the highest bid value and pay with the second-lowest bid.

 8:   The auctioneer allocates the task to the winner.

 9:  else

 10:   CN will give up this cooperation.

 11:  end if

 12:  Then the node winners fuse data followed by CN and SN data transmission.

 13:  CN rewards these SNs with transmission time.

 14: end while

3.2 Shapely value

Algorithm 2 shapely value

Require: the contribution of IoT nodes in once round and the remaining energy

Ensure: the end of every round

 1: while the reward for IoT nodes do

 2:  

 3:

 4: l(i)′ = l(i)w(i)

 5:

 6: △φ₍₁₎(l) = φ₍₁₎(l) * l(i) * α (0 < α < 1)

 7: φ₍₁₎(l)′ = φ₍₁₎(l) + △φ₍₁₎(l)

 8: end while

   return φ₍₁₎(l)′

3.3 Direct trust

Algorithm 3 direct trust value

Require: behavior of node_ij, fre

Ensure: punishment

 1: i = 1:n

 2: j = 1:n

 3: syms ,

 4:

 5:

 6: for cognitive nodes decide to do full monitor do

 7:  if find selfless behvior then

 8:   

 9:  else

 10:   

 11:  end if

 12:  update direct trust values table

 13:  if then

 14:   no punishment

 15:  else if then

 16:   give a timeout for three sensing rounds

 17:  else if then

 18:   give a timeout for three sensing rounds

 19:   de-allocate 50% of the assigned resources to the misbehaving node

 20:  else if then

 21:   de-allocate all resources and disconnect this node

 22:  else

 23:   mark the node as an undesirable node

 24:   de-allocate all resources and disconnect this node

 25:  end if

 26: end for

 27: while condition do

 28:  …

 29: end while

   return e

3.4 Indirect trust

Algorithm 4 indirect trust value

Require: the interaction among IoT nodes, α

Ensure: t_in and w_ij

 1: i = 1:n

 2: j = 1:n

 3: α = 0.05

 4: syms , , Pro_ij, t_inij, , ,

 5:

 6:

 7:

 8: for interaction happens do

 9:  if another node is trusted then

 10:   

 11:  else

 12:   

 13:  end if

 14: end for

 15:

 16:

 17:

 18: sum = sum(matrix, 2)

 19: For i = 1:n

 20:  For j = 1:n

 21:   a(ij) = a(ij) / sum(i)

 22:  End

 23: End

 24: t_all′ = α * I + (1 − α) * t_all

 25: n = 2

 26: while matrix′! = matrixⁿ do

 27:  t_all′ = t_all′ * (α * I + (1 − α) * t_all)

 28: n = n + 1

 29: end while

 30: Sort

   return t_in, w_ij

3.5 Malicious

Algorithm 5 identify malicious node

Require: tables from IoT nodes

Ensure: clique

 1: for each IoT node ∈ all nodes C do

 2:  s_i ← e_i ∈ E(random)

 3: end for

 4: for each e_i ∈ E do

 5:  l(e_i) ← argminDistance(e_i, c_j), j ∈ 1,2…k

 6: end for

 7: changed ← false

 8: iter ← 0

 9: while until changed true and iter < Mazlters do

 10:  for each s_i ∈ S do

 11:   UpdateCluster(c_i)

 12:  end for

 13:  for each e_i ∈ E do

 14:   minDist ← argminDistance(e_i, c_j), j ∈ 1,2…k

 15:   if m theninDist l(c_i)

 16:    l(e_i) ← minDist

 17:   end if

 18:   changed ← true

 19:  end for

 20:  iter++

 21: end while

 22:  give all nodes by 0/1

 23:   calculate Similarity_i,j =

 24: k = 2

 25: while k do ≤ 5

 26:  group all IoT nodes in pairs

 27:  merge k-cliques into (k+1)-cliques

 28:  remove duplicates

 29:  print all 5-cliques and monitor them

 30: end while

   return result

4.1 Pre-analysis: Choice of specific auction mechanism

Relying on existing knowledge, a special optimal auction mechanism of the multi-item auction is still complex work. It is unnecessary for us and the cognitive node to find it. Furthermore, for the following reasons, we decide to use ‘The Second Price Sealed Auction’.

• The cognitive node need not take on risk from speculators.
• This model can effectively allocate resources regardless of whether the probability distribution density function of each node is the same.
• It can be an incentive to bid on its true valuation. And ease the thinking burden of sense nodes.
• Expected return in this model is larger than the ‘first-price sealed auction’.

Meanwhile, sense nodes can communicate with each other, so we need to prevent them from cooperating. This auction model can avoid bidding rings to some extent. Then we can consider every node to decide its bid independently.

The main disadvantage is that this model doesn’t maximize the income of the cognitive node. For example, if there are two bidders, one offers 100$, the other offers 1$, this model will cause the cognitive node to lose heavily.

But it doesn’t matter. Because in general, there are about 100 sense nodes every cognitive node has so that this scenario is extremely rare. And as one might say, on account of the dominant position of the cognitive node in the forwarding aspect, and also because sense nodes always need to forward data by cognitive, sense nodes are unlikely to have a lower evaluation and don’t want to miss any choice to get the cognitive node’s help. Besides, we can set a reserve price to avoid this situation (2.4.2).

So to sum up, Multiple ‘The Second Price Sealed Auction’ is a feasible choice.

4.2 Simulation and analysis

In Figs 9 and 10, we study the performance of the CR using the first and second-price auctions. The performance enhancement, in this case, is the accurate out-of-band sensing by the IoT node. As explained in ([27]), the revenue generated by any auction is the same when the number of bidders exceeds 6. The proposed scheme enhances the performance through more reliable out-of-band sensing, reducing packet loss. The second price auction outperforms the first-price auction and in-band sensing. IoT nodes check their energy level and packet deadline and correspondingly adjust their bid value. The second price auction outperforms the first-price auction in these scenarios due to its ability to encourage IoT nodes to consider their energy levels and packet deadlines when placing bids. This consideration leads to more efficient resource utilization and task allocation, resulting in improved performance metrics such as reduced delay and enhanced throughput compared to traditional methods.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Mean delay with limit energy of three task distribution and sensing model.

https://doi.org/10.1371/journal.pone.0309123.g009

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Mean throughput with limit energy and limit packet lifetime of three task distribution and sensing model.

https://doi.org/10.1371/journal.pone.0309123.g010