Generating Natural Adversarial Examples
This addresses the need for more realistic adversarial testing in machine learning, particularly for complex domains like language, though it is incremental as it builds on existing adversarial example and GAN methods.
The paper tackles the problem of generating adversarial examples that are natural and legible, rather than unnatural perturbations, by using a framework based on generative adversarial networks to search in semantic space. The result demonstrates the approach's potential for evaluating black-box classifiers in applications like image classification, textual entailment, and machine translation, with experiments showing the adversaries are natural and useful.
Due to their complex nature, it is hard to characterize the ways in which machine learning models can misbehave or be exploited when deployed. Recent work on adversarial examples, i.e. inputs with minor perturbations that result in substantially different model predictions, is helpful in evaluating the robustness of these models by exposing the adversarial scenarios where they fail. However, these malicious perturbations are often unnatural, not semantically meaningful, and not applicable to complicated domains such as language. In this paper, we propose a framework to generate natural and legible adversarial examples that lie on the data manifold, by searching in semantic space of dense and continuous data representation, utilizing the recent advances in generative adversarial networks. We present generated adversaries to demonstrate the potential of the proposed approach for black-box classifiers for a wide range of applications such as image classification, textual entailment, and machine translation. We include experiments to show that the generated adversaries are natural, legible to humans, and useful in evaluating and analyzing black-box classifiers.