HYPOCRITE: Homoglyph Adversarial Examples for Natural Language Web Services in the Physical World

JINYONG KIM, JEONGHYEON KIM, MOSE GU, SANGHAK OHH, GILTEUN CHOI, JAEHOON JEONG

Keywords: Adversarial Examples, Homograph, Natural Language, Web Services, Physical World

Abstract:

Recently, as Artificial Intelligence (AI) develops, many companies in various industries are trying to use AI by grafting it into their domains.
Also, for these companies, various cloud companies (e.g., Amazon, Google, IBM, and Microsoft) are providing AI services as the form of Machine-Learning-as-a-Service (MLaaS).
However, although these AI services are very advanced and well-made, security vulnerabilities such as adversarial examples still exist, which can interfere with normal AI services.
This paper demonstrates a HYPOCRITE for hypocrisy that generates homoglyph adversarial examples for natural language web services in the physical world. This hypocrisy can disrupt normal AI services provided by the cloud companies.
The key idea of HYPOCRITE is to replace English characters with other international characters that look similar to them in order to give the dataset noise to the AI engines.
By using this key idea, parts of text can be appropriately replaced with subtext with malicious meaning through black-box attacks for natural language web services in order to cause misclassification.
In order to show attack potential by HYPOCRITE, this paper implemented a framework that makes homoglyph adversarial examples for natural language web services in the physical world and evaluated the performance under various conditions.
Through extensive experiments, it is shown that HYPOCRITE is more effective than other baseline in terms of both attack success rate and perturbed ratio.

One-sentence Summary: Adversarial Attacks against Natural Language Web Services in the Physical World