Diagnostics for Deep Neural Networks with Automated Copy/Paste Attacks

TL;DR

This paper introduces SNAFUE, an automated method to identify copy/paste adversarial attacks on deep neural networks, revealing numerous vulnerabilities in ImageNet classifiers without human intervention.

Contribution

The paper presents SNAFUE, a fully automated approach for discovering copy/paste adversarial examples, enhancing the ability to diagnose DNN weaknesses at scale.

Findings

Reproduced previous copy/paste attacks on ImageNet

Discovered hundreds of new vulnerabilities

Automated detection without human input

Abstract

This paper considers the problem of helping humans exercise scalable oversight over deep neural networks (DNNs). Adversarial examples can be useful by helping to reveal weaknesses in DNNs, but they can be difficult to interpret or draw actionable conclusions from. Some previous works have proposed using human-interpretable adversarial attacks including copy/paste attacks in which one natural image pasted into another causes an unexpected misclassification. We build on these with two contributions. First, we introduce Search for Natural Adversarial Features Using Embeddings (SNAFUE) which offers a fully automated method for finding copy/paste attacks. Second, we use SNAFUE to red team an ImageNet classifier. We reproduce copy/paste attacks from previous works and find hundreds of other easily-describable vulnerabilities, all without a human in the loop. Code is available at https://github.com/thestephencasper/snafue