Previously, we had tested existing manifold learning algorithms in an attempt to reduce high-dimensional data to a lower dimension while preserving inter-data distances between neighboring devices [1]. This past year, joint work with Jose Costa has resulted in a new manifold learning method called the distributed weighted multi-dimensional scaling (dwMDS) [2,3]. This new method has been demonstrated to provide an energy-efficient, distributed estimator of sensor localization which nearly achieves the Cram\'er-Rao bound in a variety of situations. A key attribute of the method is its distributed implementation - it allows nodes in a network to collaboratively estimate a global solution with minimal communication requirements. We have also demonstrated an adaptive neighbor selection method that helps to keep the estimator nearly unbiased, as shown in Figure 1.

Manifold learning in general is a data visualization tool. We have shown how the dwMDS and other manifold learning algorithms can be used to 'locate' anomalies in Internet traffic data [4]. We have developed and made publicly available online applets and command-line utilities which can be used in combination with NetFlow data to visualize very high-dimensional traffic data in two-dimensions. We have also tested the tools on data from the Abilene backbone network, and shown that they can show in two dimensions dramatic changes when traffic anomalies such as port and network scans, alpha flows, and worm propagation activity are measured on the network (more information is available online at http://www-personal.engin.umich.edu/~npatwari/mnd05/). Such visualization tools will complement detection tools by providing more information to operators and allowing them to quickly visually investigate network traffic before taking corrective action.