Wireless Sensor Network (WSN) refers to a group of spatially dispersed and dedicated sensors for monitoring and recording the physical conditions of the environment, and organizing the collected data at a central location. Research in WSNs is gaining interest due their di erent applications and the challenges that the constrained resources of sensor nodes bring on the eld. In this type of ad-hoc network, routing of data to the base station with secure transmission is of prime concerns. In this paper, we discuss possible improvements in WSN routing and security through the employment of concepts coming from Arti cial Intelligence (AI) area, such as swarm intelligence, arti cial immune systems and articial neural networks. Since WSNs are distributed computing networks, the use of AI makes the network \cognitive" toward solving problems these networks are often prone to.
neural networks [
In brief, this section discusses the imitation of grouping behaviors and optimal solution nding technique of living creatures, that can be implemented in WSNs for routing data e ciently to the base station. 2
Research in security is challenging and complex, when compared with other
issues related to WSNs [
A combination of AI with the public keying techniques can be applied for
WSNs. The concepts of Arti cial Immune Systems (AIS) like negative selection
algorithm [
A novel approach of inducing the cognitive behavior of vertebrate living organisms is proposed in this paper. This algorithm in a legitimate sensor node takes the randomness of an attacker node into account. The randomness can be related to energy of the attack packet, position of the attacker, strategy of the attacker, so on. For this purpose, the algorithm uses an arti cial neural network that is trained according to the previous attack patterns and the strategy, associated payo s for the picked strategy, and is modeled using game theory. This technique is specially designed to combat against the node capture attack by combining game theory with arti cial neural networks. A simulation environment with the malicious attacker, victim node with the proposed algorithm, and the game between them can be modeled and the results can be calibrated.
In brief, this section highlights the usage of AI concepts, to defend against few security threats and attacks. This results in the sensor node being cognitive in handling the security vulnerabilities, especially in remote monitoring and surveillance applications. 3
This section discusses few of the works carried in this area. Anthony et al. [
Still there are many research works, which adapt bioinspired computation for solving issues of wireless sensor networks; in this paper, we discuss majorly the usage of arti cial intelligence to solve routing and security related issues of WSNs.
In one of our previous works [
According to the survey, majority of the concepts related to AI like AIS,
ANNs, combinations between them, etc. have been applied for wireless sensor
networks. Nevertheless, still the concepts of evolutionary algorithms like
genetics, memetics and AIS concepts like Negative Selection Algorithm, Dendritic cell
theory, etc. can be combined with keying techniques (public, private, random,
so on) to improve the strength of cryptographic algorithms, trust and security
protocols. In our previous work, we have attempted few of the AI derived
techniques [
The strategy making capability of vertebrates to achieve a given task can be implemented in sensor nodes. This is achieved by combining game theory with machine learning algorithms. Here game theory helps in modeling di erent strategies of the individuals with their associated payo s. The use of machine learning helps in visualizing the attack scenario. Implementation of this concept leads to a cognitive system that can think and act based on the attacks with its associated intruder strategies. Games like noncooperative, repetitive, Bayesian, cooperative combined with the machine learning concept leads to a novel approach towards securing WSNs.
Here, we can notice three things, the usage of arti cial intelligence in
clustering, cryptographic keys to immune and secure their communications when
attempted for sinkhole attack, hello ood attack, spoo ng attack and so on. To
provide additional security against the node capture attack one can opt for the
combination of game theory with machine learning techniques [
We can implement the above routing and security concepts in every sensor node of the deployed network as a protocol stack. By this, we can make sensor nodes behave more intelligently to group, secure communication and combat against the attack if intrusion sustains with less help from Base Station. The combination of all these may lead to a \cognitive sensor network" not with the notion of having cognitive radio in it, but the nodes which are having the capability to visualize and take actions on its own in the absence of manual monitor. 5
This paper discusses the applications of bio- inspired computation towards solving the issues and challenges of WSNs. As an example, a novel method of securing WSN against node capture attack is explained in detail. The section highlights a technique of imitating cognitive behavior of vertebrates while defending against attacks. As introduced before in section 2 and section 4, this is achieved by combining game theory with arti cial neural networks. Game theory is used for modeling the strategic moves of victim node and attacker node while, ANN is used as pick and play tool by learning the attacker. The proposed methodology can be applied when a malicious intruder is trying for a node capture attack. A node capture attack is de ned as an attack where the attacker captures a node and hacks the critical information (eg., cryptographic algorithms) residing in the node. The defense mechanism module is shown in Fig 2. This module is designed to be built in every sensor node, which can combat against such attacks, with minimal help from Base Station. An adversary is assumed to attack any victim node by sending malicious packets to hack the information in the victim node. Intruder node is assumed mobile with its attack packet bit energy varying, such that the intruder's mobility and bit energy of attack packet is highly random in nature. Even, the attacker's strategy is random.
The proposed Intruder Defense System (IDS) has to be installed in a sensor node as an add-on application together with a cryptographic algorithm. The system will be ignited if number of erroneous packets detected by cryptographic scheme crosses beyond a particular threshold. IDS once ignited continuously combats against a malicious intruder until the game between the intruder and the victim reaches a zero sum game i.e., either intruder stops sending poison packets or dies losing its battery power or victim node is infected or dies losing its battery power.
The module consists of an ANN, whose inputs are distance, bit energy, attack count and attacker strategy. Distance input of ANN is de ned as the distance between the attacker and the victim, which is the information sent by the BS from time to time until the game exists. Bit energy input is the energy per bit of the poison packet reaching the victim node. attack count input is the previous attack counts from that intruder under vigilance, which will be updated from time to time. Apart from these, the neural network chip will be enabled only if the attacker's strategy is in attack mode. The output of the neural network is a decision depending all these inputs fed. It can be either to defend or not defend, depending on which the transmitter module will be enabled, and its transmitting energy per bit of the defense packet will be varied according to the scenario.
In Figure 2, there is a timer module, which monitors the intruder and initiates counter attack when the intruder is idle for a long time. This type of attack repeats for few times and stops if the intruder quits its attack. 5.1
Game theoretic modeling Game theory is a mathematical concept, which deals with the formulation of the correct strategy that will enable an individual or entity when confronted by a complex challenge to succeed in addressing that challenge.
Proposed game model The method proposed is a non-cooperative game between a malicious node and the IDS. The action set of malicious node is Am = fA; N Ag and that of IDS is AI = fD; N Dg the game ends if either of the one seizes its operation, leading to zero-sum game. Table 1 gives the payo matrix. The move chosen by IDS solely depends on the accuracy of Neural Network design. E1 is the utility cost for the IDS on reception of an intruder packet `I' is the packet sent by the intruder and `A' is the alarm packet used to counter attack intruder by the victim node. Epro is the energy for processing the intruder packet including the computation energy of the neural network, n is the path loss component.
E2 = (Erx A) + (Epro A) + (Etx I) + (Eamp I dn) If IDS chooses not to defend while the intruder is on attack then the utility cost is given by
E3 = (Erx I) X = X
XC
E4 = (Etx I) + (Eamp I dn) Where, XC is the damage cost of the node for choosing not to defend when the intruder is on attack. X is the residual cost of the victim node. The value of XC is decided by the amount of damage done by the intruder on the victim node, by hacking information, changing frequency, so on. If Intruder chooses not to attack if IDS is on defense, then the payo is given by
E5 = (Erx A)
Y = Y
YC
E6 = (Etx A) + (Eamp A dn) YC damage cost of intruder for not attacking the victim node, while the IDS is in defense. Y is the residual cost of the intruder. If intruder chooses not to attack and if IDS chooses not to defend then the utility cost for each is 0. Apart from these payo s, an intruder has to spend energy in listening all the nodes in network and attacking a node. If Pl is the power consumed per second for listening then, the overall energy consumed by the intruder for listening N number of nodes, is given by
Elisten = Pl t N Where, t is the time dedicated by intruder in listening to each node, N is the total number of nodes in the network and it should be noted that intruder details are unknown to the defender. In the proposed game fND, NAg is the Nash equilibrium condition. fD, Ag is the Pareto optimal strategy for the entire game. In speci c fND, Ag is the Pareto Optimality for an intruder and fD, NAg is the Pareto Optimality for IDS.
The e ciency of the game depends entirely on the accuracy of the ANN. The game discussed in this method is a non- cooperative game. Similarly, we can adapt a Bayesian game, repetitive game, so on. For the deployment of this module in a sensor node, ANN has to be pre trained with these type of attacks, which might have happened prior in a network. Implementing this technique in sensor node makes it to self defend against the malicious node attacks in WSNs. (2) (3) (4) (5) (6) (7) (8) (9) Neural Network and its associated training matrices Arti cial Neural Networks are the imitations of biological neural networks. It accepts nite set of inputs and computes the weighted sum. The sum is compared with suitable threshold. The summation and threshold unit is called a node in ANN. Always a NN requires training and is the most di cult task, where we nd the weights that achieve goal with acceptable performance. For the proposed work, we use NN with 3 neurons in the input layer, few neurons in the hidden layer and 4 neurons in the output layer. Neurons in the hidden layer depend solely on the cost and the accuracy of ANN. Well-known back- propagation algorithm can be used for training. This is an example implementation and other variants of ANN can be implemented.
An example Training matrix is shown in Table 2. X indicates `do-not care' condition. E1, E2, E3, E4 are the assumed extremes, dividing the received bit energy levels of the intruder into three classes. C1, C2, C3, C4 are the extremes dividing the count input, fed to the IDS. C1 and c2 are used in Table 2, by considering C3 and C4, and conditions d0 = d and d0 > d, remaining conditions for training NN can be derived. C is the intruder's previous attack count, d is the initial distance between attacker and the victim node and d0 is the randomly changing distance between them. It is to be noted that, the output energy level indicators are restricted for four levels. For still better design, we can opt for more level indicators, but with a tradeo of more neurons requirement rendering increase in the complexity of the design. In addition, the input count can be still divided into a ner class for better combat, but with the tradeo with the complexity of ANN design. 6
In this paper we discuss about possible approaches of using Arti cial Intelligence in handling issues of WSNs like energy consumption in data transmission for e cient routing and security in WSNs|while making the network cognitive towards handling the challenges arising while in operation. The concepts discussed in this paper direct the researchers for the use of bioinspired computation toward solving problems of WSNs by making sensor nodes more intelligent.