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

2025-05-27

AI Summary

This work addresses the lack of robustness against phoneme substitution attacks in Korean hate speech detection: existing approaches overlook the structural vulnerability of Korean phonemic representations to phonetic perturbations and lack architecture-level defense mechanisms. To bridge this gap, we propose PHISH—the first adversarial phoneme substitution method that systematically models Korean phonemic perturbation characteristics. We further design MESH, a novel hybrid encoding architecture that deeply integrates phonemic information into the model: it defines substitution rules based on Korean syllable structure, constructs dual-channel embeddings—semantic (BERT) and phonemic (CNN)—and fuses phoneme- and meaning-aware features via joint attention. Evaluated on both perturbed and clean datasets, MESH achieves F1-score improvements of 3.2–5.8%, demonstrating significantly enhanced robustness against realistic phoneme-level attacks.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Safety and Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

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.

Problem

• Addressing Korean language vulnerability to phonetic substitution attacks
• Developing architectural defenses beyond dataset construction
• Enhancing hate speech detection robustness with phonetic-semantic features

Innovation

• PHISH exploits Korean phonological characteristics
• MESH integrates phonetic-semantic features
• Enhances hate speech detection robustness