Filtering Compromised Environment Sensors Using Autoregressive Hidden Markov Model

Abstract

We propose a method based on autoregressive hidden Markov models (AR-HMM) for filtering out compromised nodes from a sensor network. We assume that sensors are healthy, self-healing and corrupted whereas each node submits a number of readings. A different AR-HMM (A, B, π) is used to describe each of the three types of nodes. For each node, we train an AR-HMM based on the sensor's readings, and subsequently the B matrices of the trained AR-HMMs are clustered together into two groups: healthy and compromised (both self-healing and corrupted), which permits us to identify the group of healthy sensors. The existing algorithms are centralized and computation intensive. Our approach is a simple, decentralized model to identify compromised nodes at a low computational cost. Simulations using both synthetic and real datasets show greater than 90% accuracy in identifying healthy nodes with ten nodes datasets and as high as 97% accuracy with 500 or more nodes datasets.