Deep One-Class Classification via Interpolated Gaussian Descriptor

One-class classification (OCC) aims to learn an effective data description enclose all normal training samples and detect anomalies based on the deviation from description. Current state-of-the-art OCC models a compact normality by hyper-sphere minimisation, but they often suffer overfitting data, especially when set is small or contaminated with anomalous samples. To address this issue, we introduce interpolated Gaussian descriptor (IGD) method, novel model that learns one-class anomaly classifier trained adversarially The differentiates their distance centre standard of these distances, offering discriminability w.r.t. given during training. adversarial interpolation enforced consistently smooth descriptor, even This enables our representative rather than fringe samples, resulting in significantly improved In extensive experiments diverse popular benchmarks, including MNIST, Fashion CIFAR10, MVTec AD two medical datasets, IGD achieves better detection accuracy current models. also shows robustness problems sets.