Communication Network Topology Inference via Transfer Entropy

In this work, we consider the problem of inferring links in a communication network, using limited, passive observations of network traffic. Our approach leverages transfer entropy (TE) as a metric for quantifying the strength of the automatic repeat request (ARQ) mechanisms present in next-hop routing links. In contrast with existing approaches, TE provides an information-theoretic, model-free approach that operates on externally available packet arrival times. We show, using discrete event simulation of a wireless sensor network, that the TE based topology inference approach described here is robust to varying degrees of connection quality in the underlying network. Compared to an existing approach which uses the linear regression based formulation of Granger Causality for network topology inference, our approach has better asymptotic time complexity, and shows significant improvement in network topology reconstruction performance. Our approach, though sub-optimal, also has better time complexity, while still retaining reasonable performance, compared to a causation entropy based optimal algorithm proposed in the literature.