Robust Deep Support Vector Data Description for Unreliable Data

Robustness against unpredictable events or samples is an unquestionable aspect of the reliability and safety of many future machine learning applications, such as self-driving cars or autonomous robots. Anomaly detection (AD) is the task of discerning anomalous samples in a collection of data, which plays a critical role in achieving the necessary robustness. Despite recent interest in developing deep methods for AD, a conclusive understanding of the problem is badly lacking. For example, in machine vision applications, a stable AD method against unreliable training data has not been well studied. In this paper, we introduce weighted-DSVDD, a new method inspired by Deep Support Vector Data Description (DSVDD) [1]. Instead of treating every training sample equally, weighted-DSVDD gives each sample a confidence score, which indicates the contribution of a training point to the objective function. Our experiments show the effectiveness and stability of the proposed method against a variety of datasets, including those with corrupted-label samples.