Towards modeling the behavior of physical intruders in a region monitored by a wireless sensor network

A priority task for homeland security is the coverage of large spans of open border that cannot be continuously physi-cally monitored for intrusion. Low-cost monitoring solutions based on wireless sensor networks have been identified as an effective means to perform perimeter monitoring. An ad-hoc wireless sensor network scattered near a border could be used to perform surveillance over a large area with relatively little human intervention. Determining the effectiveness of such an autonomous network in detecting and thwarting an intelligent intruder is a difficult task. We propose a model for an intelligent attacker that attempts to find a detection-free path in a region with sparse sensing coverage. In par-ticular, we apply reinforcement learning (RL) – a machine learning approach – for our model. RL algorithms are well suited for scenarios in which specifying and finding an op-timal solution is difficult. By using RL, our attacker can easily adapt to new scenarios by translating constraints into rewards. We compare our RL-based technique to a reason-able heuristic in simulation. Our results suggest that our RL-based attacker model is significantly more effective, and therefore more realistic, than the heuristic approach