anon-user-unicode/unicode-attack-demo

Unicode Adversarial Attack Demo

Interactive demonstration of how Unicode character substitutions can fool Large Language Models.

What This Does

This demo transforms text using special Unicode characters (like Canadian Aboriginal Syllabics or Circled Letters) and tests whether the transformation changes an LLM's prediction.

Research Findings

Tested on 59,376 samples across 3 models and 4 Unicode styles:

• —Overall Attack Success Rate: 50.2%
• —Most Vulnerable Model: Phi-3-mini (58.8% ASR)
• —Most Robust Model: Gemma-2-2b (39.0% ASR)
• —Most Effective Style: Canadian Aboriginal (56.5% ASR)

Project

Title: Unicode-Based Adversarial Attacks on Large Language Models Author: Endrin Hoti Institution: King's College London Supervisor: Dr. Oana Cocarascu