Using Learned Data Patterns to Detect Malicious Nodes in Sensor Networks

As sensor network applications often involve remote, distributed monitoring of inaccessible and hostile locations, they are vulnerable to both physical and electronic security breaches. The sensor nodes, once compromised, can send erroneous data to the base station, thereby possibly compromising network effectiveness. We consider sensor nodes organized in a hierarchy where the non-leaf nodes serve as the aggregators of the data values sensed at the leaf level and the Base Station corresponds to the root node of the hierarchy. To detect compromised nodes, we use neural network based learning techniques where the nets are used to predict the sensed data at any node given the data reported by its neighbors in the hierarchy. The differences between the predicted and the reported values is used to update the reputation of any given node. We compare a Q-learning schemes with the Beta reputation management approach for their responsiveness to compromised nodes. We evaluate the robustness of our detection schemes by varying the members of compromised nodes, patterns in sensed data, etc.