Sample-free white-box out-of-distribution detection for deep learning

Abstract

Being able to detect irrelevant test examples with respect to deployed deep learning models is paramount to properly and safely using them. In this paper, we address the problem of rejecting such out-of-distribution (OOD) samples in a fully sample-free way, i.e., without requiring any access to indistribution or OOD samples. We propose several indicators which can be computed alongside the prediction with little additional cost, assuming white-box access to the network. These indicators prove useful, stable and complementary for OOD detection on frequently-used architectures. We also introduce a surprisingly simple, yet effective summary OOD indicator. This indicator is shown to perform well across several networks and datasets and can furthermore be easily tuned as soon as samples become available. Lastly, we discuss how to exploit this summary in real-world settings.