FIND: Faulty Node Detection for Wireless Sensor Networks SenSys 2009 Shuo Guo, Ziguo Zhong, Tian He University of Minnesota, Twin Cities Jeffrey

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Abstract Wireless sensor networks (WSN) promise researchers a powerful instrument – For observing sizable phenomena with fine granularity over long periods Since the accuracy of data is important to the whole system’s performance – Detecting nodes with faulty readings is an essential issue in network management

Abstract As a complementary solution to detecting nodes with functional faults This paper proposes FIND, a novel method to detect nodes with data faults – Neither assumes a particular sensing model – Nor requires costly event injections After the nodes in a network detect a natural event – FIND ranks the nodes based on their sensing readings as well as their physical distances from the event

Abstract FIND works for systems where the measured signal attenuates with distance A node is considered faulty if there is a significant mismatch between the sensor data rank and the distance rank Theoretically, we show that average ranking difference is a provable indicator of possible data faults

Abstract FIND is extensively evaluated in – Simulations – Two test bed experiments with up to 25 micaz nodes Evaluation shows that – FIND has a less than 5% miss detection rate and false alarm rate in most noisy environments

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Introduction Wireless Sensor Networks (WSNs) have been used in many application domains – Habitat monitoring – Infrastructure protection – Scientific exploration The accuracy of individual nodes’ readings is crucial in these applications – In a surveillance network, the readings of sensor nodes must be accurate – To avoid false alarms and missed detections

Introduction Although some applications are designed to be fault tolerant to some extent It can still significantly improve the whole system’s performance – Removing nodes with faulty readings from a system with some redundancy – Or replacing them with good ones At the same time prolong the lifetime of the network

Introduction To conduct such after-deployment maintenance (e.g., remove and replace) – It is essential to investigate methods for detecting faulty nodes

Introduction In general, wireless sensor nodes may experience two types of faults that would lead to the degradation of performance – Function fault – Data fault

Function Fault Typically results in – Crash of individual nodes – Packet loss – Routing failure – Network partition This type of problem has been extensively studied and addressed by – Either distributed approaches through neighbor coordination – Or centralized approaches through status updates

Data Fault A node behaves normally in all aspects except for its sensing results Leading to either significant biased or random errors Several types of data faults exist in wireless sensor networks

Data Fault Although constant biased errors can be compensated for by after-deployment calibration methods Random and irregular biased errors can not be rectified by a simple calibration function

Outlier Detection? One could argue that random and biased sensing errors can be addressed with outlier detection – A conventional technique for identifying readings that are statistically distant from the rest of the readings However, the correctness of most outlier detection relies on the premise that data follow the same distribution

Outlier Detection? This holds true for readings such as temperatures – Considered normally uniform over space However, many other sensing readings (e.g., acoustic volume and thermal radiation) in sensor networks attenuate over distance – A property that invalidates the basis of existing outlier-based detection methods

FIND This paper proposes FIND – a novel sequence-based detection approach for discovering data faults in sensor networks – assuming no knowledge about the distribution of readings In particular, we are interested in Byzantine data faults with either biased or random errors since simpler fail-stop data faults have been addressed sufficiently by existing approaches, such as Sympathy

Ranking Violations In Node Sequences Without employing the assumptions of event or sensing models Detection is accomplished by identifying ranking violations in node sequences – A sequence obtained by ordering ids of nodes according to their readings of a particular event.

Objective of FIND The objective of FIND is to provide a blacklist containing all possible faulty nodes (with either biased or random error), in order of likelihood. With such a list, further recovery processes become possible, including – (i) correcting faulty readings, – (ii) replacing malfunctioning sensors with good ones, – or (iii) simply removing faulty nodes from a network that has sufficient redundancy As a result, the performance of the whole system is improved

Main Contribution 1 This is the first faulty node detection method – that assumes no a priori knowledge about the underlying distribution of sensed events/phenomena The faulty nodes are detected based on their violation of the distance monotonicity property in sensing – which is quantified by the metric of ranking differences

Main Contribution 2 FIND imposes no extra cost in a network where readings are gathered as the output of the routine tasks of a network The design can be generically used in applications with any format of physical sensing modality – Heat/RF radiation – Acoustic/seismic wave As long as the magnitude of their readings roughly monotonically changes over the distance a signal travels

Main Contribution 3 We theoretically demonstrate that – The ranking difference of a node is a provable indicator of data faults – If the ranking difference of a node exceeds a specified bound, it is a faulty node

Main Contribution 4 We extend the basic design with three practical considerations First, we propose a robust method to accommodate noisy environments – Where distance monotonicity properties do not hold well Second, we propose a data pre-processing technique – To eliminate measurements from simultaneous multiple events Third, we reduce the computation complexity of the main design – Using node subsequence.

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Related Work In general, faults in a sensor network can be classified into two types – Function fault Abnormal behaviors lead to the breakdown of a node or even a network as a whole – Data fault A node behaves as a normal node in the network but generates erroneous sensing readings difficult to identify by previous methods because all its behaviors are normal except the sensor readings it produces

How to Resolve Data Fault One way to solve this problem is after- deployment calibration – a mapping function (mostly linear) is developed to map faulty readings into correct ones Performance of existing calibration methods – Depends on the correctness of the proposed model – Exhibits significant degradation in a real-world system too-specific additional assumptions no long hold

Outlier Detection Outlier detection is a conventional method for identifying readings that depart from the norm Correctness is based on the assumption that neighboring nodes have similar readings

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Model and Assumptions Assumption on Monotonicity – RSS of Radio Signals – Propagation Time of Acoustic Signals

RSS of Radio Signals

Propagation Time of Acoustic Signal

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Main Design

Examples of Map Division

Main Design

Detection Sequence Mapping

Main Design

A Simple Example of Ranking Difference

Ranking Differences Affected By Faulty Nodes

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Practical Issues Detection in Noisy Environments Simultaneous Events Elimination Subsequence Estimation

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

System Evaluation Evaluate the performance of FIND in two scenarios In the first scenario, an accurate  is assumed so that – The first  N nodes with the largest ranking differences are selected as faulty nodes – without using the detection algorithm in Algorithm 1

System Evaluation In the second scenario, accurate a is unavailable and Algorithm 1 is used for selecting faulty nodes – whose ranking differences are greater than B –  is still used for computing Pr( ¯s|s) but it can be less accurate).

On Radio Signal Deploy 25 MicaZ nodes in grids (5×5) on a parking lot The distance between each row and column is 5m

On Radio Signal

Broadcasting events, identified by their unique event ids, are generated one by one. In the experiment, the sending power of each event is adjusted to 0dBm such that the communication range is around 25m

On Radio Signal Upon receiving a broadcasting packet, MicaZ nodes measure the RSS and record it together with event id The number of events generated varies from 19 to 49

 -Detection

B-Detection

Ranking Difference

On Acoustic Signal We deploy 20 MicaZ nodes in grids (4×5) The distance between each row and column are set to 1.5m

On Acoustic Signal

All nodes are equipped with microphones to receive 4KHz acoustic signals generated by a speaker and record corresponding timestamps Before generating the acoustic signal, all the nodes are synchronized

False Negative Rate

False Positive Rate

Ranking Differences

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

Large Scale Simulation Study Simulation Setup Comparison of  - and B-detection Impact of Noise Impact of Inaccurate  Impact of Network Density Impact of Simultaneous Events

Simulation Setup In the simulation, both the sensor nodes and events are randomly generated on the map If not specified, 100 nodes are randomly deployed on a 250m× 250m map The sensing range is 25m and the sample size (the number of generated events) is 50 L is set to 10, and thus the complexity of computing Pr( ¯ s|si) is bounded by 10 3 All the data are based on 100 runs

Comparison of  - and B-detection

Impact of Noise

Impact of Inaccurate 

B-Detection vs. Density

Impact of Simultaneous Events

Outline Abstract Introduction Related Work Model and Assumptions Main Design Practical Issues System Evaluation Large Scale Simulation Study Conclusions

FIND is proposed – a faulty node detection method for wireless sensor networks Without assuming any communication model, FIND detects nodes with faulty readings based only on their relative sensing results, i.e., node sequences

Conclusions Given detected node sequences, an approach is first proposed to estimate where the events take place and what the original sequences are Then we theoretically proved that the average ranking differences of nodes in detected sequences and original sequences can be used as an effective indicator for faulty nodes

Conclusions Based on the theoretical study, a detection algorithm is developed for finally obtaining a blacklist when an accurate defective rate is unavailable Based on extensive simulations and test bed experiment results, FIND is shown to achieve both a low false negative rate and a low false positive rate in various network settings

Strength and Weakness Strength – General Weakness – Fault rate  is difficult to obtain – Impact of false negatives and false positives to accuracy is not discussed