A cognitive scheme for gateway protection in wireless ... - Springer Link

7 downloads 0 Views 859KB Size Report
Aug 15, 2007 - GRISP, a novel Gateway Relocation algorithm for Improved. Safety and .... been reported. The ENN is a two-layer ENN with (n2 + n) weights.
Appl Intell (2008) 29: 216–227 DOI 10.1007/s10489-007-0088-5

A cognitive scheme for gateway protection in wireless sensor network Waleed Youssef · Mohamed Younis

Published online: 15 August 2007 © Springer Science+Business Media, LLC 2007

Abstract In wireless sensor networks, sensor readings are gathered at a gateway for processing and forwarding to a remote command center. The potential closeness of the gateway to dangerous events, e.g. fires, exposes it to damage and thus risks making the network dysfunctional. Therefore, protecting the gateway by repositioning it away from safetyhazardous spots is critical for the operation of the network. However, moving the gateway too far from the sensors that report on active events would have negative effect on the network performance, e.g. throughput and energy consumption. Therefore, balancing the gateway safety and network performance goals will be necessary. In this paper, we present GRISP, a novel Gateway Relocation algorithm for Improved Safety and Performance. GRISP employs an evolutionary neural network model to assess the safety of the gateway at the various locations. The model is then used to direct the search in an area of interest for a safer position that would enhance or at least maintain an acceptable level of network performance. In addition, GRISP guides the gateway during the move by finding safe paths leading to the new location. Our experimental validation results demonstrate the effectiveness of GRISP. Keywords Wireless sensor networks · Genetic algorithms · Artificial neural networks · Evolutionary neural networks · Gateway relocation · Safety improvements W. Youssef () · M. Younis Department of Computer Science and Electrical Engineering, University of Maryland Baltimore County, Baltimore, MD 21250, USA e-mail: [email protected] M. Younis e-mail: [email protected]

1 Introduction and preliminaries In the past few years, wireless sensor networks (WSNs) have received increasing interest from the scientific and engineering communities due to their potential use in many applications such as target tracking, disaster management, border control and battlefield surveillance [5, 10, 12, 18, 20]. Components of the WSN include sensors, gateways, command nodes, and monitored phenomena. Sensors are miniaturized battery-operated devices equipped with a communication subsystem. The main function of a WSN is to collect data and report any abnormal conditions happening in an area of interest. These conditions are reported by close-by sensors and transmitted as packets, often over multi hop paths, to a central unit called the gateway for analysis. The gateway, which is sometimes called sink node, interfaces the WSN to command and control centers, reporting serious findings and/or required data. The reported data may indicate that some moving targets crossed in from outside the deployment area, the irruption of a fire, etc. Figure 1 illustrates the use of WSNs in a disaster management application. Given its role in interfacing the network to the command center, the gateway is a critical asset for the network. However, the gateway may be located in the proximity of serious events or targets and thus may be at a high risk of damage. For example, in a disaster management application the gateway can be a robot that is located close to fires and collapsed buildings. In military setups, the gateway may become in the shooting range of enemy tanks, missiles, artilleries, etc. Therefore, the gateway should be protected in order to keep the network operational. One way to keep the gateway safe is to move it away from hazardous regions. However, staying far from the sensors that report on a serious event would cause negative impact on the network performance. For example, data will be expected to be routed over long paths

A cognitive scheme for gateway protection in wireless sensor network

Fig. 1 A sample sensor network where data is transferred over multi-hop routes to a gateway node, which reports to a remote command center

which may risk increased packet drop and data latency. In addition, transmitting packets over large distances would boost the energy consumption at the individual sensors and would thus shorten their lifespan and eventually the network lifetime too. Gateway relocation can also be pursued to achieve some performance goals. Since sensors consume energy every time they are involved in data transmission, highly active sensors tend to lose their energy faster than inactive ones. Once all the onboard energy supply drains, the sensor becomes dysfunctional forcing a change in the network topology to set new data paths. In [24], it has been shown that the gateway can become isolated with the continual death of surrounding sensors and can thus make the network useless. Repositioning the gateway can counter such a scenario. In addition, changing the location of the gateway can eliminate bottlenecks in the routing tree by influencing the formation of a new network topology that better serve the ongoing traffic [3]. However, such performance-centric relocation may move the gateway dangerously close to one or multiple targets/events in the environment and may expose the gateway to the risk of getting damaged or captured, which can lead to interruption in the network operation. Thus, relocating the gateway with factoring in performance and safety concerns would be more effective in protecting the integrity of the network and its operation. In this paper, we present GRISP; a novel scheme for gateway relocation that improves safety of the gateway as well as enhances the WSN performance. The idea is to provide the gateway with means to assess whether its current position is sufficiently safe or it needs to relocate to a better spot. GRISP further identifies a set of safe positions in the deployment area. GRISP also factors in the effects of the gateway repositioning on the WSN performance. A performance-based selection criterion is applied to pick a position among the set of candidate safe spots. Basically,

217

the new location should enable the gateway to better serve the sub-region/event that triggers the most data traffic without negatively impacting the tracking of other events. In addition, the approach studies the implications of the gateway travel path on gateway safety. The goal is to keep the gateway out of harm and the network efficiently operational while the gateway is in transit. Since the physical security of the gateway is usually measured qualitatively, we employ decision support systems to generate a quantitative assessment that guides the gateway during the relocation process. A Decision Support System (DSS) is an intelligent module that makes decisions based on historical data. The process for enabling such an intelligent decision-making system consists of three main steps. The first step is to collect historical data and use it to train the DSS. The objective of the training is to inform the DSS about what events happened in the past and how the environment reacted to them. The second step is to define and build the decision model. The third step is to utilize the trained DSS in making future predications and decisions. Recent developments in DSS have shown that heuristics—including artificial neural networks, fuzzy sets, and genetic algorithms—offers opportunities for combining multiple criteria and exploring new patterns, which eventually enhances the quality of the decision making process [7, 9, 19, 21]. The selection of the right heuristics to use is problem-dependent and varies based on many different factors [19]. Artificial Neural Networks (ANN) proved to be superior to traditional statistical methods when data exhibit unpredictable nonlinearity, when patterns important to the decision-making process cannot be clearly identified, or when data are fuzzy in nature, involving human opinions or subject to some degree of uncertainty [21]. The organization of this paper is as follows. The next section presents the artificial neural networks model used. Section 3 discusses the related work. Section 4 descries our approach and highlights few implementation issues. Validation results are presented in Sect. 5. Finally, Sect. 6 concludes the paper and outlines our future research plan.

2 The artificial neural network model The general structure of an artificial Neural Network (ANN) consists of many neurons that are interconnected together to form one or more layers. Each neuron can be considered as a simple processor whose inputs are weighted variables and output is a non-linear function of the sum of the inputs, their associated weights and the neurons’ internal bias value. One of the main features of this type of network is its distributed associative memory property, in which the information is stored in the weighted links, rather than at specified memory addresses. An ANN can be considered as a generic learning

218

W. Youssef, M. Younis

relocates to this point. The general formula for r is: r=

2 +n n



i=n2 +1

 wi

n  j =1





dj l w[(i−n2 )+(j −1)∗n] − Ti−n2

− Tn+1

Fig. 2 Topology of the ENN used in GRISP, where n events have been reported. The ENN is a two-layer ENN with (n2 + n) weights and (n + 1) thresholds

machine that has been proven to be capable of forming representations of complex and non-linear phenomena [7, 19]. The topology of an ANN is problem-dependent but usually is a multi-layer network with either forward feed or backward propagation links. In forward feed networks, the input of one layer is the output of the previous layer with no links going backwards. However, in recurrent neural networks, the output of one layer can be fed to the input of a previous layer [19]. For an ANN to provide the desired output, it has to go through a learning phase first. The learning phase is a multi-step process that sets the weights and bias values to suit the relationship between the inputs and outputs in the training data set. Learning methods includes back propagation or evolutionary techniques [19]. Evolution of neural networks using evolutionary algorithms has gained popularity in recent years giving rise to a new branch known as Evolutionary Neural Networks (ENN) [23]. In this paper, we pursue heuristics (namely genetic algorithms) to train the neural network and find values for all weights and thresholds. Back propagation learning technique may also be employed as long as no feedback path is used. We opted to use the heuristic approach in the implementation. GRISP employs an ENN to assess the risk at new locations before moving the gateway. The topology of the ENN used is illustrated in Fig. 2. The figure depicts a two-layer network with (n + 1) neurons, where n is the number of events reported in the environment. The hidden layer consists of n neurons, while the output layer has only one neuron. Each input to the ENN represents the distance di between current gateway location and event # i in the environment. The output from the ENN is a risk factor r that indicates the risk that the gateway will be exposed to if it

where wi are the weights, and Tj are the bias values. In the historical data, we use the distance between the gateway and the event/target as one of the factors that indicate the threat, which the gateway is exposed to at its current location. The idea is to calculate the risk, i.e. probability of being harmed, using the historical data. Once these probabilities are calculated, the data are fed into the neural network for further processing. Since the historical data used to train the ENN model consists of multiple positions, the risk factor output would be dependent on the quality of the data presented in the learning phase. In the next section, we discuss prior work related to the problem on hand.

3 Related work Some published work has studied the effect of gateway relocation on WSN performance in terms of energy, throughput and latency. The idea is to relocate the gateway in the proximity of highly active sensor nodes. Reported results in [3, 24] showed that a considerable improvement is achieved by relocating the gateway to a location where traffic volume is the highest within the deployment area. Such repositioning of the gateway increases the average lifetime of the sensor nodes by decreasing the average energy consumed per packet. To achieve the same goal, continual gateway mobility is also considered in [4, 13, 16]. A similar approach was pursued in [1] for mobile ad hoc networks, where the gateway is relocated to the weighted geographic centroid of a group of nodes by considering the location and traffic generated by nodes regardless of the established routes. However, none of these approaches takes into account any gateway’s safety concerns in motivating the move, determining the new position and charting the travel path. Since optimal placement of the gateway is reducible to the P-Center problem, which is proved to be a NP-hard problem [22], heuristics have been pursued to resolve this problem. One such technique is the use of neural networks and genetic algorithms [2]. Neural networks have been used effectively for assessing risk in decision-making systems through the evaluation of many environment parameters and event occurrence probabilities. Many new commercial and industrial applications of neural networks have emerged over the last few years. Examples include medical diagnostics, fraud detection, and risk-based decision-making applications [15, 19, 21]. The

A cognitive scheme for gateway protection in wireless sensor network

219

risk-based decision-making process is very useful in understanding semi- or ill-structured phenomena, such as in natural disasters, by streamlining thinking and devising rational procedures. For example, in [21] to study bushfires, historical data was manipulated to discover relationships between hazards occurrence and complex set of environmental variables. Chen et al. used a similar approach in assessing the probability of house survival from bushfires. They showed that after a series of tests, neural networks were capable of predicting risky patterns under all tested configurations with a great deal of flexibility [9]. A similar model for assessing the safety of the gateway in WSNs was built in [25]. The approach used evolutionary neural networks (ENN) to estimate the amount of risk the gateway would be exposed to when relocating to a new location before the actual move takes place. The approach also proposes a new position with better safety-performance quality. However, it does not guide the gateway through the relocation process. Similar to the gateway’s path finding problem in WSNs, Batalin and Sukhatme [8] solved the problem of exploring an unknown environment using a single robot. They presented an efficient minimalist algorithm that uses markers which the robot drops off as signposts to aid exploration. Another work related to navigating a sensor field is reported in [14]. The approach uses sensor readings to form artificial potential field that a user should try to avoid while traveling towards a destination. Unlike GRISP, this type of work is not geared for performance optimization and does not consider the network operation.

GRISP answers the following questions; where is the safest location in the environment that keeps the network efficiently operational, what path the gateway should take for reaching that safe location, how safe that path is, and whether the network operation may be disrupted during the gateway’s motion or not. GRISP consists of four modules that perform the following functions: (1) collect historical data and use that data to train the ENN model; (2) employ genetic algorithms to resolve ENN weights and thresholds; (3) utilize the resultant ENN to assess the need for relocation and qualify the safety of candidate position to which the gateway can move; (4) finally, utilize the resultant ENN to find a safe path to the new position. GRISP ensures that the network sustains an acceptable level of performance while the gateway at its new safe position and that there is no disruption in network operation while maintaining the safety of the gateway while it is in transition. Details about these modules are provided in the next few subsections.

4 Safety-based gateway repositioning In this paper, we present GRISP, a new algorithm for Gateway Repositioning for Improved Safety and Performance measures in WSNs. GRISP is executed at the gateway in order to improve its safety and enhance network performance. GRISP employs evolutionary neural networks to assess the safety of the gateway at the various locations. GRISP is different from other methods, such as the one described in [3], which emphasize only improvements in performance (throughput and energy consumption) of the WSN. In GRISP, the relocation process is triggered by the fact that current position is no longer safe for the gateway even though the performance at this position is at an acceptable level. We would like also to note that GRISP can be applied in combination with other performance-centric relocation scheme in order to ensure the gateway’s protection. The ENN model, discussed in Sect. 2, is used as a decisionmaking expert system that guides the gateway in finding a safe position. To maintain good performance, GRISP picks the best location from a network performance point of view and then consults the ENN system to recommend a safer position in the vicinity of that location.

4.1 The GRISP-history module In this module, the past gateway locations are manipulated to generate the learning data set. The manipulation process starts by estimating the threat index at each location the gateway visited. This can be using either simulation environments or by utilizing historical data if they already exist. The process is described as follows. Using the model for performance-centric relocations, the gateway would be free to move to any location that improves performance. A snapshot of the environment at the time of relocation will be captured and then used to generate the risk assessment data. For each visited location, a value called “Threat Index” will be calculated (as explained next). Then, this information along with the captured snapshot would be used to train the neural network. The output of the neural network is a “Risk Assessment” factor that can be used in making future safe relocation decisions. 4.1.1 Threat index One approach for threat assessment is to identify the sensors that are reporting on an event and consider their neighborhood to be a risky area. However, multiple sensors may be reporting on the same event, target or phenomenon. Therefore, it is difficult to define a fine-grained threat measure since the sensing range may be large. Instead, we associate a threat function to each serious event reported. The function is different from one event to another depending on many factors, such as the distance between the gateway and each serious event DSRi , the volume of data VRi arriving from a certain region Ri , and the severity level of the reported sensor data SLRi . To assess the risk of relocating the gateway, we define the threat index (TR) with respect to a region Ri as follows:

220

TRRi = VRi ∗ SLRi (1/DSRi )SLRi . When sensors report normal conditions or non-harmful events, SLRi = 0 and hence TRRi = 0. In addition, the threat index is inversely proportional to the distance to the event. The severity of an event affects the growth of the threat index as the gateway gets closer to the event. 4.1.2 Risk assessment We employ the ENN model, presented in Sect. 2, in assessing the risk that the gateway is exposed to when staying in its current position or relocating to a different one. The quality of the risk assessment obtained is highly influenced by the threat index estimation and also by the quality of the training data presented. The training data differs based on the sensors’ application and network operational model. Initially, simulation experiments that resemble the actual application setup may be conducted to collect historical data. In the experiment, the gateway can be randomly repositioned or relocated for increased performance without care for its safety. The threat indexes are to be tracked and tabulated for training the ENN. The training is used to calculate risk for all previously visited locations in the historical data set. Since the number of these locations is infinite, a method was developed to discretize the environment and evaluate the risk at each discrete location. The idea is to divide the deployment area into a two dimensional grid of size m × m, resulting in m2 cells. The risk at the center of each cell location is calculated as: risk(cell(l)) = thigh /(thigh + tlow ) where thigh represents the number of times the next gateway location had a higher threat index point than its current one, and tlow represents the number of times the gateway had moved to a spot with lower threat index. The threat index is calculated using the formula TRRI , above. During the network operation, we rely on the evolutionary nature of the model to adapt the ENN to dynamic changes in the environment. Figure 3 outlines the GRISP-History module.

Fig. 3 The GRISP-History module

W. Youssef, M. Younis

4.2 The GRISP-GA module In this module, we use genetic algorithms to efficiently and effectively identify all the weights and thresholds of the ENN [11]. Each learning data sample would yield an equation in (n2 + 2n + 1) variables, which represents the total number of weights and thresholds in the ENN. The model we are using is similar to the ENN depicted in Fig. 2. The figure illustrates that the proposed neural network consists of n input values, a hidden layer with n neurons, and one neuron representing the output layer. Therefore, the total number of unknowns in this figure is as follow (1) there are n2 links between the input layer and the hidden layer (2) there are n links between the hidden layer and the single neuron output layer (3) there are n thresholds values for the hidden layer (4) there is one threshold value for the output layer. Adding all these variables would yield an equation in (n2 + 2n + 1) variables. The number of equations equals the cardinality of the learning data set provided. The GRISPGA module is outlined in Fig. 4. Details about the implementation of each component of the genetic algorithm are provided next. Most of the genetic algorithms parameters, operator values and constants are selected based on results of preliminary experiments that were geared for finding best choices for the evolution of the population. • Encoding. Each individual member in the population is represented as an array of floating point coefficients. Each entry in the array corresponds to a single coefficient in the neural network. The size of the array is dependent on the number of events in the environment. For example, if n is the number of events reported, then the individual member length is (n2 + 2n + 1), as shown in Fig. 2. • Initial population. An initial population of size 100 is randomly created. The size of the population remains constant throughout the algorithm. Also, all members of the population have the same individual length. • Fitness. The fitness of each member in the population is calculated as the total hamming distance between the individual and each member in the learning data set. The less the total hamming distance, the more fitted the individual member is.

Fig. 4 The GRISP-GA module

A cognitive scheme for gateway protection in wireless sensor network

221

• Parent selection scheme. We used the Tournament selection scheme in GRISP. This method adopts a tournamentlike competition, where four parents compete with each other in a pair-wise competition. The best two parents are selected and returned while the worst member is saved for the Replacement scheme. • Crossover operator. This operator is applied to the two parents selected from the previous step in order to generate new members in the population. We used a fixed 3point crossover operator in GRISP. The choice of a 3-point crossover was selected based on conducted experiments that showed the superiority of this operator compared to a single point crossover. Basically, it allows the algorithm to converge faster. Due to space limitations, we are presenting this as fact and use the 3-point crossover as a constant all over conducted experiments, as will be shown next. The fitness of the two resulting offspring is calculated and the one with the better fitness is returned for mutation. • Mutation operator. Mutation is used mainly to introduce some randomness into the population. This operator is performed on each newly created offspring as follows. For each chromosome of the offspring and with probability pr, the chromosome is selected for mutation. Its value is then replaced by a randomly generated amount β and the new value is stored. In GRISP, we used a value of pr equal to 10%. The main reason we used a high mutation probability—in addition to introducing some randomness in the population—is to avoid falling into the local optimal solution. This step is followed by a local optimization step that will enhance the resultant mutated member further. Other mutation probability should not impact the solution much. Our choice for this value was based on experiments conducted at early stages to validate the best value for the mutation probability. • The local optimization process. After mutation, the new offspring is optimized. Such local optimization step is used mainly to speedup the convergence of the algorithm and reduces its time complexity. The local optimization process has two main stages. The first stage is to scan the input mutated offspring, and for each chromosome of the offspring, slightly change its value by δ (where δ is less than 1% of the value) and check the effect of this change on its fitness. If the fitness improves, keep the change; otherwise discard it. The second stage is to change the threshold value. This is done by summing up the input values, and then the threshold is set to the negative of this value. The fitness of the new sequence is calculated and new optimized sequence is stored if its fitness is better than the fitness of the original sequence. • Replacement scheme. The optimized offspring is compared with its parents. If its fitness is better than one of the parents, then that parent is replaced in the population. If both parents are better, then the worst member obtained from the tournament selection method is compared to that

offspring. If the fitness of the offspring is better, the member is replaced in the population. Otherwise, the offspring is discarded. • Stopping criteria. The genetic algorithm stops after evolving for multiple iterations with no improvements in the fitness of the population or when reaching a preset bound on the number of iterations. In GRISP, we set the maximum number of iterations to 20,000 iterations. 4.3 The GRISP-Loc module The GRISP-Loc module is responsible for making the safety/performance decision. It utilizes the resultant ENN to relocate the gateway to a safer spot than its current position. The ENN handles the safety estimation of the gateway. It is used to calculate an index called “SafeIndex” of the new proposed location. For the WSN performance, GRISP-Loc uses the network throughput and the average remaining energy of some relaying sensors in its vicinity to make the performance decision. A “PerfIndex” is used to evaluate these two performance metrics as an average function. To balance the interest in protecting the gateway and the need to keep an acceptable level of network performance, a weighting function is employed. The objective function returns a value, called the Relocation Index for each considered location. As an example, the relocation index (RI) at point l1 can be calculated as: RI(l1 ) = (a × SafeIndex(l1 )) + (b × PerfIndex(l1 )) where a is the risk weight, and b is the performance weight. To pick a new position that yields a good performance, GRISP-Loc uses a performance-centric relocation algorithm, e.g. [3], to nominate a candidate position. The idea of the performance-centric relocation is to relocate the gateway closer to highly active sensors with the intension of reducing sensors energy consumption and reduce data latency. The approach presented in [3] calculates the centroid location of all active sensors and tends to move the gateway to this new location. Next, GRISP-Loc applies a depthconstrained algorithm that exhaustively searches the vicinity of such proposed gateway position and within a predefined radius r, as shown in Fig. 5. The algorithm divides the area into coronas and wedges and the location that has the best value of the objective function is selected. Once a new safe position has been identified, the GRISPLoc module consults the ENN model to compare the risk at the nominated position to that of the current gateway location. If the risk is not lower, either the relocation is to be canceled or the value of “a” is increased so that the search can yield a safer position. The action that GRISP will take depends on the application. Otherwise, the GRISP-Path module is invoked to chart a travel path for the gateway to its new location.

222

W. Youssef, M. Younis

Fig. 6 Illustration of technique used to discretize the deployment area

Fig. 5 An illustration of exhaustively searching all locations in the vicinity of proposed gateway location, where all cells within certain predefined radius r is checked

4.4 The GRISP-Path module GRISP-Path strives to find a safe path for the gateway to follow in order to reach its designated new location, identified by the GRISP-Loc module. The GRISP-Loc module utilized the ENN to estimate risks at candidate positions before the relocation is approved. The same approach is used in finding a safe and operational path between current gateway location and the recommended new one. GRISP-Path balances between gateway safety and WSN performance during this transition time. Details of GRISP-Path are presented next. Methodology After identifying a new position, the question that immediately arises is how the gateway will manage to safely get to that new location without negatively impacting the network operation or gateway safety. Such impact could result from the loss of connectivity between the gateway and some sensors, thus forcing dramatic changes in the network topology or even disrupting data transmission and reception. In addition, the network could be affected due to a damage inflected on the gateway when traveling. For example, the gateway may come within the shooting range of some artillery fire or enemy tanks in a war zone, or get exposed to an excessive heat of a burning fire. The goal of GRISP-Path is to find the shortest and safest path that maximizes the network performance. Since there are an infinite number of possible paths to be considered, we partition the area around the gateway into a two-dimensional grid of size m × n. In theory the entire deployment area can be a search space. However, in practice it may be desirable to limit the travel distance and to prevent the gateway from going too far while moving to its destination. Therefore, we form the grid based on the coordinates of

the current and new gateway location. For example, if the current position is (Xcurrent , Ycurrent ) and the new location is (Xnew , Ynew ), we can confine the gateway to the rectangular area whose diagonal is the line between the current and the new position. Alternatively, the boundary of the allowed travel area can be stretched to enable more flexibility by expanding the search space. For example in Fig. 6, m is selected as 2|Xnew − Xcurrent | to widen the possible travel area, i.e. 0.5|Xnew − Xcurrent | from each side. The size of the grid is a design parameter which can be determined based on the risk that the gateway is running. For example, if the gateway is getting away of an eminent danger, the grid may be expanded to increase the feasibility of finding an escape path for the gateway. The cells in the grid serve as steps on the gateway’s travel path. For each cell c, GRISP-Path estimates a cost factor as follows: Cost(c) = w1 ∗ Performance(c) + w2 ∗ Safety(c) + w3 ∗ distance(c, destination). The function Performance(c) uses the aggregate throughput of the nodes that are located within the cell c. The rationale is that passing close to nodes that transmit many packets most probably will yield good average delay and energy per packet, network throughput, and reliability [24]. The function Safety(c) captures the risk that the gateway runs when passing this cell. The ENN model described earlier is used to assess such risk. The last factor reflects the proximity of the cell c to the new gateway location and is used to measure progress towards the final travel destination. The weights w1 , w2 , and w3 are used to indicate the relative importance of these factors. For example, when sustaining good performance is crucial to the application, it is expected to raise the value of w1 . Also, when it is desired to limit the travel time or to reduce the motion overhead, w3 can be set high. It is worth noting that the values for the weights are all normalized so that cost(c) is independent of the range of values that each factor takes.

A cognitive scheme for gateway protection in wireless sensor network

223

5.1 Environment setup and performance metrics

Fig. 7 Pseudo code for the GRISP-Path algorithm

Algorithm GRISP-Path models the grid as a directed graph G. Each cell c is represented as a node in G. Each cell is connected to all its eight adjacent cells with inbound edges that have a cost cost(c), defined above. In case there is a minimal performance and/or maximal risk constraints that must be met, the violating cells are not represented in G. A preprocessing step in GRISP-Path removes these cells from G. For example, if the cell is too close to a fire, it is excluded from G to prevent it from being part of the picked path and expedite the search. The constraints can simply be specified in terms of bounds on Performance(c) and Safety(c). Now, the path selection can be easily mapped to a least cost path problem that is solvable using Dijkstra’s algorithm. Obviously, the size of cells in the grid is an important design parameter. Small cells increase the precision of the analysis and subsequently the quality of the selected path. However, a fine grained grid will grow G and thus increase the complexity of GRISP-Path. In addition, the speed of the gateway is expected to affect the cell size decision. Basically a fast gateway will not benefit from a high resolution grid. The GRISP-Path algorithm is summarized in Fig. 7.

5 Experimental validation In this section, we describe the validation of GRISP in a simulated target tracking application setup similar to that used in [3, 24], where targets are assumed to come from outside the area with a speed chosen uniformly from the range 4 m/s to 6 m/s and a constant direction. The simulator was implemented, along with all GRISP modules, using C++ and was run on a PC with Pentium IV 2.4 GHz Intel processor with 512 MB of RAM. The organization of this section is as follows. We start by describing the environment setup and performance metrics used. We follow by briefly discussing three conducted experiments with different objectives and then presenting results obtained. Finally, we show statistics about the complexity of the genetic algorithm.

In all experiments, the network consists of 100 sensors that are randomly placed in a 500×500 m2 area. The gateway initial position is determined randomly within the region boundaries. Each node is assumed to have an initial energy of 5 joules. A node is considered non-functional if its energy gets completely depleted. The maximum transmission range for a sensor node is assumed to be 50 meters. Targets are distributed uniformly over the deployment area and remain active until the end of the simulation. A free space propagation channel model is assumed with the capacity set to 2 Mbps [6]. In the genetic algorithm, Mersenne Twister random number generator was used to generate uniformly distributed random numbers [17]. For each conducted simulation experiment, the collected results fall into two main categories; the gateway safety and the network performance. For the safety category, the Euclidean distance between the gateway and each target was reported as the safety measure. For the network performance, we used the network throughput and the average energy per packet. Moreover, we collected results about path discovery, namely, reporting on feasibility of finding paths to the new identified safe locations, and the increase in the traveled distance due to safety precautions. In all experiments, the following three approaches were compared (1) when relocation is not allowed (2) when the gateway relocates based only on performance metrics only (3) when using GRISP. 5.2 Experimental results We studied the impact of changing the number of targets on the achievable network performance and on the safety of locations recommended by GRISP. We ran experiments for 5 different targets count (3, 4, 5, 6, and 8). We also observed the effect of changing the number of targets on the quality of selected paths. Additional experiments were also conducted to test the impact of changing the following factors: grid size when manipulating the historical data, risk/performance weights, search radius in the GRISP-Loc algorithm. Results of these experiments are reported in the summary table presented at the end of this section. For every experiment, we applied 5 distinct seeds in order to generate random network topologies. Each experiment lasted 12,000 sec. We observed that with confidence level > 90%, the simulation results stayed within 6%–10% of the sample mean. 5.2.1 Experiment 1 In this experiment, we studied the effect of using GRISP on positions visited by the gateway. We have conducted two tests. First, we applied the performance-based relocation

224

Fig. 8 A sample experiment showing locations visited by the gateway using the Performance-Relocation approach and GRISP in the presence of 4 targets

W. Youssef, M. Younis

Fig. 9 Average distance between the gateway and targets for all approaches tested when changing number of targets in the environment

scheme of [3]. Figure 8(a) shows a plot of some locations visited by the gateway during this experiment. The second test involved GRISP. Figure 8(b) shows locations visited by the gateway in this case. It is clear from the first figure that the gateway jeopardizes its safety by moving dangerously close to one of the dangerous targets in the environment. While, through GRISP, the gateway managed to stay further away from risky targets to ensure its safety. 5.2.2 Experiment 2 In this experiment, the effect of increasing the number of targets on the WSN was examined. The number of targets was set to (3, 4, 5, 6, or 8). Figures 9 and 10 show that the gateway managed to maintain safe distance from targets and sustain acceptable network performance. With the increase in number of targets, GRISP proved to be an effective technique and continued to appropriately select the right position for the gateway. The performance plots in Fig. 10 show that the simulation results were comparable to these results for the Performance-Relocation technique and were better than the No-Relocation technique. The increase in throughput with the increases number of targets is not related to the use of GRISP, it is a direct result of increasing network traffic due to the existence of more targets. In the figure, when number of targets is 3, the energy per packet went up as the gateway moved away from the target (and consequently the data sources). That effects started to be less influential when more targets were involved causing a balance (basically lengthening the data path for one targets while shortening those of other targets). The energy consumption in communication depends on the sensor and gateway locations. Disseminating packets from a distant sensor usually involves multiple relaying nodes and thus boosts the energy consumption. GRISP may change the position of the gateway based on the severity of targets and the proximity of targets to the current gateway location. In the simulation experiment, we set the minimal count of high-risk targets to 3. Therefore, in Fig. 10(b) the energy per packet went up as the gateway moved away from the target (and consequently the data sources). That effects started to be less influential

Fig. 10 Effect of changing number of targets on WSN performance for experiment #2. The throughput is measured in terms of number of packets generated per second. The average energy represents average consumed energy per packet

when more targets were involved causing a balance (basically lengthening the data path for one targets while shortening those of other targets). Overall, the graph indicates that throughput and energy consumption are not impacted when GRISP is used. The results of this experiment confirmed the effectiveness of the GRISP approach.

A cognitive scheme for gateway protection in wireless sensor network

225

Fig. 12 Comparison of number of times GRISP was able to find a path when increasing the number of targets

Fig. 11 Performance and Safety measures of the gateway when in transient to its new location for 16 different experiments

5.2.3 Experiment 3 The purpose of this experiment is to measure the performance of the network under GRISP-path. We used the throughput of sensors, or sensors packets load, that are within a predefined distance r to the path of the gateway (in all experiments r was set to 50). We also used the distance between the gateway and each serious event or target as an indication of the gateway safety factor. Figure 11-top shows the performance of the WSN for each experiment. Figure 11-bottom shows the average distance between the gateway and serious events during this transient period. It should be noted that the figure reports the throughput in values checked while executing GRISP-path. In other words, the y-axis in the graph represents the Actual Throughput, rather than the estimated throughput that is used in determining the path. In the approach, we estimate the throughput of each cell before the start of the relocation process. Then, the algorithm excludes all cells that have throughput less than a predefined threshold value. Meaning, when the value of the objective function (performance + safety) is be-

low the threshold value, the cell is marked as ineligible and is not included in the path finding algorithm. Therefore, the approach grantees that if a path exists, the throughput along that path would be in the acceptable range. Figure 11(b) shows the distance between the gateway and each serious event in the environment. The figure shows the maximum, minimum, and average distance that was achieved during the process. Figure 11(a) and 11(b) indicates that GRISP-path was successful in protecting the gateway (keeping safe distance from serious events in the environment) during the move as well as keeping the performance of the network at an acceptable level, if not boosting the network performance. In certain occasions, GRISP-Path was able to conclude that a safe and valid path does not exist. Occasions include paths that will cause the WSN to malfunction or stop operating, losing connections between sensors and the gateway, or exposing the gateway to harm. The chart in Fig. 12 compares the number of times the gateway was able to find a path or not when varying the number of targets in the deployment area. The graph shows that the percentage of finding safe paths for GRISP is higher than that of not finding one. Therefore, the chart demonstrates that GRISP-Path is highly effective in finding feasible paths when they exist. Figure 13 illustrates the percentage of increase in the distance when using GRISP in comparison to shortest distance travel between the start and target points. We observed that GRISP picks the straight line travel path, which would be the fastest, whenever possible. However, in many situations that straight line path was not the best path to the destination location due to hazards or other safety concerns as mentioned before. The figure shows that the percentage increase in the traveled distance when using GRISP ranged from 0 to 50%. The increase in the distance represents the price for keeping the gateway safe and the network operational. It is worth noting that the variability in path length with the increased

226

W. Youssef, M. Younis

Fig. 14 Effect of changing the number of targets and the size of learning data on the genetic algorithm complexity Fig. 13 Percentage of increase in traveled distance by the gateway due to the use of GRISP-Path Table 1 Summary of both network performance and gateway security results for all performed experiments Process

Security

Network performance Throughput Energy consumed

Perf-Reloc

−41.82%

19.28%

−8.57%

−32.50%

11.01%

−5.14%

16.03%

−7.46%

−3.62%

(compared to No-Reloc) GRISP (compared to No-Reloc) GRISP (compared to Perf-Reloc)

presence in targets is due to the random selection of the initial location of gateway, the location and direction of targets, and the severity of each target. 5.3 Summary results Table 1 summarizes all results obtained for all approaches tested relative to each other. It is observed that GRISP was able to enhance the safety of the gateway compared to the Performance-Relocation approach, results showed an average improvement of 16.03% in the distance between the gateway and harsh targets/events. Furthermore, GRISP was able to enhance and boost the network performance compared to the No-Relocation approach, with the results indicating 11.01% improvement in network throughput and 5.14% saving in energy per packet. GRISP was able to balance the safety and performance goals of the gateway. All these results have confirmed the effectiveness of GRISP in securing the gateway while boosting the network performance. As a final remark, it is noticeable from the summary table that the GRISP safety measure was decreased by 32.50% compared to the no-relocation approach. For this no-relocat-

ion approach, the gateway is usually far from monitored events and hence the gateway is not able to efficiently track targets. In addition, energy consumption and throughput increase due to the involvement of many sensors in data forwarding and receiving. So, the decrease in safety for GRISP is supplemented by improvements in performance. The table also illustrates that the safety index of GRISP is better than the performance relocation by a 16.03%. In summary, GRISP efficiently balances between two objectives: safety and performance. 5.4 Genetic algorithm complexity In GRISP, we employed genetic algorithms to train neural networks. Such approach has been in use for a while. However, a concern about the time complexity of the algorithm always arises. Figure 14 illustrates some results obtained from the genetic algorithm. It shows the effect of changing number of targets as well as the number of grids (size of learning data) on the running time and quality of the solution obtained. Experiments were conducted for number of targets between 3 and 8 with cardinality of learning data ranging from 4 to 100. As number of targets and learning data grow, we believe that the approach will not suffer; time complexity will grow linearly proportional to the increase in these inputs. In summary, the figure confirms that the complexity and quality of solutions in GRISP stayed always within the acceptable range and did not burden the system with more than anticipated computational overhead.

6 Conclusion In wireless sensor networks, data are usually routed from sensors to a gateway node, where it is further processed and then appropriate actions are taken. Given this responsibility of the gateway node, it plays a critical role in the success and integrity of the network operation. For applications in inhospitable environments, e.g. tracking targets in a combat field,

A cognitive scheme for gateway protection in wireless sensor network

227

the gateway may get close to harsh events or dangerous targets, which jeopardizes its safety. Thus, changing the location of the gateway may be required to protect it from such potential damage. However, such repositioning may place the gateway further away from these tracked events and targets and may negatively impact the network performance; in terms of throughput, delay, and energy consumption. To tackle such issues, this paper introduced GRISP, a new adaptive safety and performance aware algorithm and safe path discovery for gateway relocation. GRISP employs evolutionary neural networks to assess the risk involved in placing the gateway at various positions and while in transient to its new location in an area of interest. It applies heuristics for finding suitable locations and safe paths that gateway can use to reach its new location. GRISP has many advantages. First, it balances between the need for protecting the gateway and the desire for enhancing the network performance. Second, it selects the best quality location in terms of safety and performance for the gateway to relocate to. Third, it helps sustain network operation by selecting paths for which the gateway stays accessible to nearby sensors and would not cause the sensors to lose connection with the gateway or make routing packets to their intended destinations harder. Fourth, GRISP ensures the safety of the gateway by avoiding paths that are within the range of some dangerous events happening in the deployment area. Fifth, when multiple paths exist, GRISP ensures that the path with the best performance and safety measures is selected. The experimental results confirmed that GRISP was able to protect the gateway by relocating to safer positions in the environments and by selecting paths that kept the network operational and the gateway safe.

6. Andresen J, Rappaport T, Yoshida S (1995) Propagation measurements and models for wireless communications channels. IEEE Commun Mag 33(1) 7. Anderson JA (1995) An introduction to neural networks. MIT Press, Cambridge 8. Batalin M, Sukhatme G (2003) Efficient exploration without localization. In: Proceedings of the international conference on robotics and automation (ICRA2003), Taipei, Taiwan, May 2003 9. Chen K, Jacobsom C, Blong R (2004) Artificial neural networks for risk decision support in natural hazards: a case study of assessing the probability of house survival from bushfires. Environ Model Assess 9(3):189–199 10. Estrin D et al (1999) Next century challenges: scalable coordination in sensor networks. In: Proceedings of the 5th annual international conference on mobile computing and networks (MobiCOM ’99), Seattle, WA, August 1999 11. Goldberg DE (1989) Genetic algorithms in search, optimization, and machine learning. Addison–Wesley, Reading 12. Karl H, Willig A (2005) Protocols and architectures for wireless sensor networks. Wiley, United Kingdom 13. Kim H, Abdelzaher T, Kwon W (2003) Minimum-energy asynchronous dissemination to mobile sinks in wireless sensor networks. In: Proceedings of the 1st international conference on embedded networked sensor systems (SenSys’03), Los Angeles, CA, November 2003 14. Li Q, De Rosa M, Rus D (2003) Distributed algorithms for guiding navigation across a sensor network. In: Proceedings of annual international conference on mobile computing and networks (MobiCOM ’03), San Diego, CA, September 2003 15. Lisboa PJ (2001) Industrial use of safety-related ANN, HSE contract research report. John Moores University, Liverpool, UK 16. Luo H et al (2005) TTDD: two-tier data dissemination in largescale wireless sensor networks. Wirel Netw 11:161–175 17. Matsumoto M, Nishimura T (1998) Mersenne twister: a 623dimensionally equidistributed uniform pseudo-random number generator. ACM Trans Model Comput Simul 8(1):3–30 18. Min R et al (2001) Low power wireless sensor networks. In: Proceedings of the international conference on VLSI design, Bangalore, India, January 2001 19. Negnevitsky M (2004) Artificial intelligence—a guide to intelligent systems, 2nd edn. Addison–Wesley, Boston 20. Pottie GJ, Kaiser WJ (2000) Wireless integrated network sensors. Commun ACM 43(5):51–58 21. Risk decision-making in natural hazards: bushfire as example. Nat Hazards Q 7(2):1–5 (2001) 22. Shmoys DB, Tardos E, Aardal K (1997) Approximation algorithms for facility location problems. In: Proceedings of the 29th ACM symposium on theory of computing, El Paso, TX, May 1997, pp 265–274 23. Yao X (1999) Evolving artificial neural networks. Proc IEEE 87(9):1423–1447 24. Younis M, Bangad M, akkaya k (2003) base-station repositioning for optimized performance of sensor networks. In: Proceedings of the IEEE VTC 2003—Wireless ad hoc, sensor, and wearable networks, Orlando, FL, October 2003 25. Youssef W, Younis M (2006) A cognitive approach for gateway relocation in wireless sensor networks. In: Proceedings of the 19th international Florida artificial intelligence research society conference (FLAIRS 2006), Melbourne Beach, FL, May 2006

References 1. Ahmed M et al (2001) Positioning range extension gateways in mobile ad hoc wireless networks to improve connectivity and throughput. In: Proceedings of the IEEE military communications conference (MILCOM’01), Washington, DC, October 2001 2. Adickes M et al (2002) Optimization of indoor wireless communication network layouts. IIE Trans 34(9):823–836 3. Akkaya K, Younis M, Bangad M (2005) Sink repositioning for enhanced performance in wireless sensor networks. Comput Netw 49:512–434 4. Akkaya K, Younis M (2004) Energy-aware routing to a mobile gateway in wireless sensor networks. In: Proceedings of the IEEE Globecom wireless ad hoc and sensor networks workshop, Dallas, TX, November 2004 5. Akyildiz IF et al (2002) Wireless sensor networks: a survey. Comput Netw 38:393–422