Abstract Wireless Image Sensor Networks (WISNs) consisting of untethered camera nodes and sensors may be deployed in a variety of unattended and possibly hostile environments to obtain surveillance data. In such settings, the WISN nodes must perform reliable event acquisition to limit the energy, computation and delay drains associated with forwarding large volumes of image data wirelessly to a sink node. In this work we investigate the event acquisition properties of WISNs that employ various techniques at the camera nodes to distinguish between event and non-event frames in uncertain environments that may include attacks. These techniques include lightweight image processing, decisions from n sensors with/without cluster head fault and attack detection, and a combination approach relying on both lightweight image processing and sensor decisions. We analyze the relative merits and limitations of each approach in terms of the resulting probability of event detection and false alarm in the face of occasional errors, attacks and stealthy attacks. Keywords security Image sensor networks · Lightweight event acquisition · Sensor network 1

Event-driven visual sensor networks (VSNs) rely on a combination of camera nodes and scalar sensors to determine if a frame contains an event of interest that should be transmitted to the cluster head. The appeal of eventdriven VSNs stems from the possibility of eliminating nonrelevant frames at the source thus implicitly minimizing the amount of energy required for coding and transmission. The challenges of the event-driven paradigm result from the vulnerability of scalar sensors to attack or error and from the lightweight image processing available to the camera nodes due to resource constraints. In this work we focus on the reliability issues of VSNs in the case of global actuation attacks on the scalar sensors. We study the extent to which various utility functions enable an attacker to increase the average expected number of affected nodes with a relatively small penalty in the loss of stealth. We then discuss tradeoffs between different attack detection strategies in terms of the cost of processing and the required information at the cluster head and nodes. 1.

Abstract — Event-driven visual sensor networks consist of col- order to determine their relevance to the surveillance task [6]. laborating camera nodes and scalar sensors which aid in the Figure 1 depicts one possible scenario where each camera detection of events of interest in the environment. This collab-node receives decision support from a scalar sensor before oration is significant since the camera nodes generally utilize lightweight image processing in order to determine if a frame is transmitting frames to the cluster head. relevant to the given application. The reliability of the supporting In comparison with other approaches, the event-driven scalar sensor however may be compromised by an actuation scalar-assisted approach may be considered a “push ” approach attack which perturbs the sensor’s measurements. In this work in that the visual nodes transmit selected frames to the cluster-we examine the achievable actuation of hostile nodes that are not head when such frames become available. In contrast, in a globally coordinated and that may be selfish or untrustworthy in their preferences. We compare our findings with existing research “pull ” approach, the cluster-head advertises what features are which assumes that all the hostile nodes are coordinated to of interest in the application [7]. The feature advertising thus actuate with the same parameter. We determine that given certain pulls data that matches the search from the visual nodes. The conditions, local optimization may actually result in a stronger appeal of the push-based approach in the limited-resource stealthy attack than the global coordination case. regime of VSNs is that the camera nodes are able to “rid”