Traffic Matrix Estimation Using Invertible Neural Networks

Ill-posed inverse problems appear in many fields and involve determining the causal factors behind a set of observations. Within the context of network tomography (NT), an interesting instance of such a linear inverse problem is traffic matrix estimation (TME) from link load measurements. In this paper, we investigate and experimentally assess the application of invertible neural networks (INNs) to address the TME problem. Specifically, we develop a custom INN architecture integrated with autoencoder (AE)-based dimensionality reduction and propose two operational modes, one of which is also capable of traffic matrix synthesis. A reference implementation of the proposed approach is published under a permissive open-source license, and performance evaluation is conducted using a comprehensive set of metrics on a dataset collected from a backbone network.