Identifying untrustworthy predictions in neural networks by geometric gradient analysis

摘要

The susceptibility of deep neural networks to untrustworthy predictions, including out-of-distribution (OOD) data and adversarial examples, still prevent their widespread use in safety-critical applications. Most existing methods either require a retraining of a given model to achieve robust identification of adversarial attacks or are limited to out-of-distribution sample detection only. In this work, we propose a geometric gradient analysis (GGA) to improve the identification of untrustworthy predictions without retraining of a given model. GGA analyzes the geometry of the loss landscape of neural networks based on the saliency maps of their respective input. We observe considerable differences between the input gradient geometry of trustworthy and untrustworthy predictions. Using these differences, GGA outperforms prior approaches in detecting OOD data and adversarial attacks, including state-of-the-art and adaptive attacks.