Data-based distributed classification and its performance analysis

Distributed classification using multimodal sensors is a problem of very high practical importance. Most of the existing distributed classification systems are designed under the assumptions that prior class probabilities, and/or observation models are known. In this paper, we design a distributed classification system without requiring any prior model information. Specifically, at each local sensor, multiple binary support vector machine (SVM) based classifiers are used and each classifier is trained to distinguish one class from the rest. At the fusion center, the Dempster-Shafer theory is adopted to effectively combine the evidence from all SVMs with appropriately defined basic probability assignments. The final decision is made by selecting the class with the highest belief. Theoretical performance prediction methods are proposed for the designed classification system. Through experiments on a synthetic dataset and the benchmark 1999 KDD intrusion detection dataset, we demonstrate the effectiveness of the evaluation method and the superiority of the proposed framework over the conventional Bayesian cost based fusion rule in this context.