Given an observation randomly generated from two alternative probability density functions (pdfs) P0 and P1, we consider the problem of deciding which pdf generated the observation. To design the decision technique we assume that we either know P0 or have a set of samples generated from it; the P1 pdf is instead completely unknown. Such a scenario arises, for example, in security contexts, where the attacker's behavior is completely unknown to the legitimate users. When the P0 pdf is known, we resort to the likelihood test (LT), while when a set of samples with its distribution is available, we resort to one-class classification (OCC). We focus on the problem of learning OCC models that operate as the LT. We show this occurs for the multilayer perceptron neural network (NN) and the one-class least-squares support vector machine (OCLSSVM) models properly trained as two-class classifiers using an artificial dataset for the negative class, obtained by generating samples uniformly distributed over the domain of the positive class dataset. The artificial dataset is used only for training, while the OCC is used on negative-class samples generated from a different pdf. We also derive a modified stochastic gradient descent (SGD) algorithm that provides OCC operating as LT without the need for the artificial dataset. Furthermore, we show that the OCLSSVM with suitable kernels operates as the LT at convergence. Lastly, we prove that the widely used autoencoder (AE) classifier generally does not provide the LT.
翻译:暂无翻译