Deep Verifier Networks: Verification of Deep Discriminative Models with Deep Generative Models

TL;DR

This paper introduces Deep Verifier Networks (DVN), a novel framework using deep generative models to verify deep discriminative models' predictions, enhancing AI safety by detecting out-of-distribution inputs, adversarial examples, and anomalies.

Contribution

The paper presents a new verification framework with a disentangled conditional variational auto-encoder that does not require retraining for different models.

Findings

Achieved state-of-the-art results in out-of-distribution detection

Effective adversarial example detection

Successful anomaly detection in structured prediction tasks

Abstract

AI Safety is a major concern in many deep learning applications such as autonomous driving. Given a trained deep learning model, an important natural problem is how to reliably verify the model's prediction. In this paper, we propose a novel framework -- deep verifier networks (DVN) to verify the inputs and outputs of deep discriminative models with deep generative models. Our proposed model is based on conditional variational auto-encoders with disentanglement constraints. We give both intuitive and theoretical justifications of the model. Our verifier network is trained independently with the prediction model, which eliminates the need of retraining the verifier network for a new model. We test the verifier network on out-of-distribution detection and adversarial example detection problems, as well as anomaly detection problems in structured prediction tasks such as image caption generation. We achieve state-of-the-art results in all of these problems.