Char-mander Use mBackdoor! A Study of Cross-lingual Backdoor Attacks in Multilingual LLMs

TL;DR

This paper investigates cross-lingual backdoor attacks in multilingual LLMs, showing how poisoning data in one language can transfer malicious triggers across languages via shared embeddings, exposing a significant security vulnerability.

Contribution

It introduces the concept of cross-lingual backdoor attacks (X-BAT) and demonstrates their effectiveness in multilingual models through toxicity classification case studies.

Findings

Backdoors in one language transfer to others via shared embeddings.

Rare and high-occurring tokens serve as effective triggers.

The vulnerability affects the model's architecture, enabling concealed backdoors.

Abstract

We explore \textbf{C}ross-lingual \textbf{B}ackdoor \textbf{AT}tacks (X-BAT) in multilingual Large Language Models (mLLMs), revealing how backdoors inserted in one language can automatically transfer to others through shared embedding spaces. Using toxicity classification as a case study, we demonstrate that attackers can compromise multilingual systems by poisoning data in a single language, with rare and high-occurring tokens serving as specific, effective triggers. Our findings expose a critical vulnerability that influences the model's architecture, resulting in a concealed backdoor effect during the information flow. Our code and data are publicly available https://github.com/himanshubeniwal/X-BAT.