Detecting Malicious Sensor Nodes from Learned data Patterns

Sensor network applications may involve dealing with the remote, distributed monitoring of inaccessible and hostile locations. These networks are vulnerable to security breaches both physically and electronically. The sensor nodes, once compromised, can send erroneous data to the base station ,thereby possibly compromising network effectiveness. The sensor nodes are organized in a hierarchy where the non-leaf nodes serve as the aggregators of the data value sensed at the leaf level and the root node is considered the Base Station. In current researh on sensor networks outlier detection mechanisms are used by a parent node to detect erroneous children nodes among its children as it is assumed that the data reported by the children comes from the same distribution and they are of almost equal values. But outlier detection mechanisms are not applied for networks where sensed data varies widely over the region of deployment. So in such a scenario we have used offline neural network based learning technique to model spatial patterns in sensed data. Then the nets are used to predict the sensed data at any node given the data reported by its neighbors, when they work online. The differences between the value predicted and the corresponding one reported by the node is measured. Each node incrementally updates the reputations of its child nodes based on those calculated differences. We have used robust schemes like Q-learning and Beta reputation based approaches to detect compromised or faulty nodes. We have evaluated the robustness of our detection scheme by varying the members of compromised nodes, error types, patterns in sensed data