Outlier Detection for Generative Models with Performance Guarantees

Abstract

We consider the problem of recovering signals using deep generative models, from measurements contaminated with sparse outliers. We propose an optimization based outlier detection approach for reconstructing the ground truth signals modeled by generative models under sparse outliers. We further establish theoretical recovery guarantees for our proposed reconstruction approach under outliers. Our results are applicable to a broad class of generative neural networks with an arbitrary number of layers. The experimental results show that the signals can be successfully reconstructed under outliers using our approach.