8fault tolerance in collaborative sensor networks for target

8/9/2019 8Fault Tolerance in Collaborative Sensor Networks for Target

1/37

1

Fault Tolerance in CollaborativeSensor Networks for Target Detection

IEEE TRANSACTIONS ON COMPUTERS, VOL.53, NO. 3, MARCH 2004


2/37

2

Scenario Senor nodes

Fault-free

Faulty

Target Modeled by the signal

it emits


3/37

3

Sensors for Target Detection Sensor nodes obtain a target energy measurement after

T seconds while a target was at a given position inside oroutside the region R

Obtaining that energy requires preprocessing of the timeseries measured during period T

The detection algorithm consists of exchanging and

fusing the energy values produced by the sensor nodesto obtain a detection query answer (Value fusion)


4/37

4

Sensors for Target Detection Note that a more accurate answer can be obtained in

general if the sensors exchange their time series ratherthan energies

However, that would require high communicationbandwidth that may not be available in a sensor network

The performance of fusion is partly defined by the

accuracy that measures how well sensor decisionsrepresent the environment or ground truth.


5/37

5

Sensors for Target Detection Finally, this work assumes that detection results need to

be available at each node

The reason for such a requirement is that the results canbe needed for triggering other actions such aslocalization of the target detected

This requirement can be fulfilled by having a central

node make a decision and disseminate that decision toall the nodes in the network


6/37

6

Sensors for Target Detection The correctness of such a scheme relies on the central

nodes correctness, therefore, central node-basedschemes have low robustness to sensor failure

Distributing the decision making over several or all thenodes improves reliability at the cost of communicationoverhead


7/37

7

Fault Model Faults include misbehaviors ranging from simple crash

faults, where a node becomes inactive, to Byzantinefaults, where the node behaves arbitrarily or maliciously

In this work, faulty nodes are assumed to sendinconsistent and arbitrary values to other nodes duringinformation sharing


8/37

8

Fault Model


9/37

9

Fault Model The algorithm for target detection needs to be robust to

such inconsistent behavior that can jeopardize thecollaboration in the sensor network

For example, if the detection results trigger subsequentactions at each node, then inconsistent detection resultscan lead each node to operate in a different mode,resulting in the sensor network going out of service

The performance of fusion is therefore also defined byprecision. Precision measures the closeness of decisionsfrom each other, the goal being that all nodes obtain thesame decision


10/37

10

Target Energy

This study assumes no such obstacles to be present inthe region considered

Energy measurements at a sensor are usually corrupted

by noise


11/37

11

Distributed Detection (Data fusion) All local sensors communicate their data to a central

processor performing optimal or near optimal detection

Decentralized processing: some preliminary processing

of data is performed at each sensor node so thatcompressed information is gathered at the fusion center

Loss of performance, reduced communicationbandwidth

The performance loss of decentralized schemes maybe reduced by optimally processing the informationat each sensor node


12/37

12

Distributed Detection (Data fusion)


13/37

13

Binary hypothesis testing problem The observations at all the sensors either correspond to

the presence of a target (hypothesis H1) or to theabsence of a target (hypothesis H0)

The performance of detection is measured by the falsealarm probability PF and the probability of miss PM

The false alarm probability is the probability ofconcluding that a target is present when the target isabsent

The miss probability is the probability of concluding thata target is absent when a target is actually present


14/37

14

Neyman-Pearson criterion[27] Minimize the global probability of miss PM assuming that

the global probability of false alarm PF is below a givenbound

The thresholds used at each sensor and at the fusioncenter need to be determined simultaneously tominimize the PM under the constraint PF <


15/37

15

Agreement Problem and Fault Tolerance As processors in such a system cooperate to achieve a

specified task, they often have to agree on a piece ofdata that is critical to subsequent computation

Although this can be easily achieved in the absence offaulty processors, for example, by simple messageexchange and voting, special protocols need to be usedto reach agreement in the presence of inconsistent faults

Three problems have drawn much attention in trying todevelop these protocols, namely, the consensus problem,the interactive consistency problem, and the generalsproblem [1], [11]


16/37

16

Agreement Problem and Fault Tolerance The consensus problem considers n processes with initial

values xi and these processes need to agree on a valuey=f(x1,,xn), with the goal that each non-faulty processterminates with a value yi=y

The interactive consistency problem is like the consensusproblem with the goal that the non-faulty processesagree on a vector y=(y1,,yn) with yi=xi if process i isnon-faulty

The generals problem considers one specific processor,named general, trying to broadcast its initial value x toother processors with the requirement that all non-faultyprocesses terminate with identical values y and y=x ifthe general is non-faulty


17/37

17

Agreement Problem and Fault Tolerance A protocol is said to be t-resilient if it runs correctly

when no more than t out of N processes fail before orduring operation

The following results were derived regarding thegenerals problem under different assumptions

Theorem 1. There is a t-Byzantine resilient authenticationsynchronous protocol which solves the generals problem[18].

Theorem 2. There is a t-Byzantine resilient synchronousprotocol without authentication which solves the generalsproblem if and only if t/N < 1/3 [21].


18/37

18

Agreement Problem and Fault Tolerance A commonly used protocol for the generals problem,

namely, the oral message algorithm (OM), wasdeveloped in [18]. Whenever t out of N nodes are faulty,the OM(t) algorithm is guaranteed to provide agreement

among the N nodes if and only if N >= 3t+1

After the OM algorithm is performed for each node actingas general, inconsistent values sent by node C arereplaced by a common value (i.e., 10.0)

Note that, in this example, the final decisions of the non-faulty sensors are incorrect since the target is outside theregion of interest; however, the decisions are consistent.


19/37

19

Agreement Problem and Fault Tolerance Below is an example where a protocol for the generals

problem is used to reach interactive consistency amongthe four nodes of Fig. 2


20/37

20

Agreement Problem and Fault Tolerance In applications where processors hold an estimate of

some global value, it may be sufficient to guarantee thatthe nodes agree on values that are not exactly identicalbut are relatively close to one another. This was

formulated as the approximate or inexact agreementproblem[7][20][4]

Fewer messages exchanged among nodes at the cost ofdegraded precision

Some agreement protocols attempt to diagnose theidentity of faulty nodes so that diagnosis results can beused in subsequent agreement procedures[26]


21/37

21

Algorithms Different fusion algorithms can be derived by varying the

size of the information shared between sensor nodes

Value fusion where the nodes exchange their rawenergy measurements

Decision fusion where the nodes exchange localdetection decisions based on their energymeasurement [6]


22/37

22

Algorithms Value Fusion (with faults) at each node:

1. obtain energy from every node

2. drop largest n and smallest n values

3. compute average of remaining values

4. compare average to threshold for final decision

Decision Fusion (with faults) at each node:

1. obtain local decision from every node

2. drop largest n and smallest n decisions

3. compute average of remaining local decisions

4. compare average to threshold for final decision


23/37

23

Evaluation Metrics Precision

Measure the closeness of the final decisions obtained by allsensors, the goal being that all non-faulty nodes obtain thesame decision

Accuracy The goal being that the decision of non-faulty nodes is

object detected if and only if a target is detected (faultalarm probability and the detection probability)

Communication overhead This is not the focus of this paper and only qualitative

evaluations are made Robustness

System failure probability (when the faulty nodes exceedthe bound of tolerable faulty nodes)


24/37

24

Comparison of Algorithms Due to the reduced amount of information shared in

decision fusion, the communication cost is lower indecision fusion than in value fusion.

The system failure probability is identical for value anddecision fusion since failures depend on the number offaulty nodes present and not on the algorithm used.

However, the performance measured in terms ofprecision and accuracy differs from one option to theother.


25/37

25

Performance Value Fusion (without faults)

False alarm probability

False alarms occur when the average of the N values

measured by the sensors is above the threshold v inthe absence of target


26/37

26

Performance Value Fusion (without faults)

Detection probability

Detections occur when the average of the N valuesmeasured by the sensors is above the threshold

vin

the presence of target


27/37

27

Performance Value Fusion (with faults)

False alarm probability

t faults are present

Worst-case scenario (t sensors report the maximumallowed value)

n lowest values and n highest values are dropped sothat the decision is based on the N-2n middle-rangevalues


28/37

28

Performance Value Fusion (with faults)

Detection probability

t faults are present

Worst-case scenario (t sensors report the minimumallowed value)

Since the energy measured is a function of theposition of the sensor, the detection probabilitydepends on which sensors are faulty and which non-faulty sensor values are dropped


29/37

29

Performance


30/37

30

Performance


31/37

31

Simulation Results Although equations to evaluate the performance of value

and decision fusion were derived and validated for asmall number of nodes, they were found computationallyimpractical when the number of nodes exceed 20.

Simulation was therefore used to compare theperformance of the algorithms

Without Faulty Nodes

The superiority of value fusion over decision fusion

decreases as the decay factor increases


32/37

32

Simulation Results


33/37

33

Simulation Results With Faulty Nodes

In the presence of a target, a faulty node reports thelowest permissible value in value fusion and a localno detection in decision fusion, and vice versa.

The maximum power was increased substantially inthe presence of faults to obtain comparableperformance than in the absence of faults

Overall, faults have more impact on value fusion thanon decision fusion


34/37

34

Simulation Results


35/37

35

Simulation Results


36/37

36

Simulation Results


37/37

37

Conclusion

The performance of localization and tracking algorithms,as opposed to detection algorithms, and the replacementof exact agreement by other agreement algorithms suchthat inexact or approximate agreement need to be

investigated

Methods for determining how to deploy the sensors inthe region of interest need to be developed to optimize

the detection performance

8fault tolerance in collaborative sensor networks for target

Documents