I. Introduction
More often than not, real-world data are multidimensional and imperfect. These attributes pose serious challenges, especially in datasets of limited size. Traditional support vector machines (SVMs) [1], [2], [3] are fragile in the presence of outliers; even a single corrupt point can arbitrarily affect classification performance. Persistent or nonprobabilistic data corruption stems from failures in sensor inputs, or from malicious data tampering.