Gaussian-Based Runtime Detection of Out-of-distribution Inputs for Neural Networks

Vahid Hashemi, Jan Kretínský, Stefanie Mohr, Emmanouil Seferis

Abstract: In this short paper, we introduce a simple approach for runtime monitoring of deep neural networks and show how to use it for out-of-distribution detection. The approach is based on inferring Gaussian models of some of the neurons and layers. Despite its simplicity, it performs better than recently introduced approaches based on interval abstractions which are traditionally used in verification.