PHISH in MESH: Korean Adversarial Phonetic Substitution and Phonetic-Semantic Feature Integration Defense

TL;DR

This paper introduces novel phonetic-aware methods to improve hate speech detection in Korean, addressing adversarial phonetic substitutions by leveraging language-specific features and architectural integration.

Contribution

It presents PHISH and MESH, the first phonetic-informed and architecture-based defenses tailored for Korean, enhancing robustness against phonetic adversarial attacks.

Findings

Improved detection accuracy on perturbed datasets

Effective integration of phonetic features enhances robustness

Reflects realistic adversarial attack strategies

Abstract

As malicious users increasingly employ phonetic substitution to evade hate speech detection, researchers have investigated such strategies. However, two key challenges remain. First, existing studies have overlooked the Korean language, despite its vulnerability to phonetic perturbations due to its phonographic nature. Second, prior work has primarily focused on constructing datasets rather than developing architectural defenses. To address these challenges, we propose (1) PHonetic-Informed Substitution for Hangul (PHISH) that exploits the phonological characteristics of the Korean writing system, and (2) Mixed Encoding of Semantic-pHonetic features (MESH) that enhances the detector's robustness by incorporating phonetic information at the architectural level. Our experimental results demonstrate the effectiveness of our proposed methods on both perturbed and unperturbed datasets, suggesting that they not only improve detection performance but also reflect realistic adversarial behaviors employed by malicious users.