Abstract

In wireless sensor networks, intruders can manipulate some sensors’ observations locally, which results in outliers in distributed detection. These outliers can be detected and removed by a fusion centre as all the sensors’ observations are available. For wireless sensor networks without a fusion centre, however, the detection performance can be significantly degraded as distributed consensus algorithms are vulnerable to outliers. In this paper, we consider the outlier detection for wireless sensor networks without a fusion centre when a distributed consensus algorithm is employed for distributed detection.

I. INTRODUCTION

Wireless sensor networks (WSNs) are becoming more popular as they can be employed for various civil and military applications including environmental and industrial monitoring and surveillance. In general, a WSN consists of a number of sensor nodes that can perform sensing and wireless communications to send their measurements or data to a fusion centre (FC). Because of the limited capability of sensor nodes that are usually small and have limited power sources (e.g., batteries), the operations at sensors nodes have to be simple and efficient. For example, each sensor can decide the presence of a certain target using hypothesis testing and transmit its decision to a FC so that the FC can combine the decisions from sensor nodes [1].

In performing distributed detection for WSNs, there are various problems. As wireless channels from sensors to a FC are not ideal due to fading and noise, the FC can receive the decisions of sensor nodes with errors. For decision fusion, these channel impairments can be taken into account [2]. Another problem is that the required bandwidth for transmitting sensors’ decisions increases with the number of sensor nodes if orthogonal channels are used. To avoid this problem, all the sensors can transmit signals to a FC in a common channel using multiple access schemes [3], [4]. In this case, the FC has a superposition of the transmitted signals that is a decision statistic for the FC to make a global decision.

Since WSNs are prone to munipulation by adversaries, outlier detection is required [5], [6], [7], [8]. For example, some sensor nodes may have biased observations and transmit manipulated decisions to a FC. In [5], a kernel-based technique is used to detect outliers provided that each sensor can have a sufficient number of observations in the absence of a parametric model for outliers. In [6], for an event boundary detection, an outlier detection problem is formulated. Using spatial correlation, it is shown that the event boundary can be detected. In [7], statistical hypothesis testing is employed for outlier detection with Markov models that can capture spatial structures of WSNs. In [8], an overview of outlier detection techniques for WSNs is provided. As discussed in [8], there are various problems where outlier detection is required in WSNs. Furthermore, the formulation of outlier detection relies on a given application and problem.