As large language models (LLMs) become more and more skilled at writing human-like text, the ability to detect what they generate is critical. This session explores a novel attack vector, homoglyph-based attacks, that effectively bypasses state-of-the-art LLM detectors.
46 RSVP'd
We'll begin by explaining the idea behind homoglyphs, characters that look similar but are encoded differently. You'll learn how these can be used to manipulate tokenization and evade detection systems. We'll cover the mechanisms of how homoglyphs alter text representation, discuss their impact on existing LLM detectors, and present a comprehensive evaluation of their effectiveness against various detection methods.
Join us for an engaging exploration of this emerging threat and to gain insight into how security researchers can stay ahead of evolving evasion techniques.
You’ll gain valuable insights into:
🔹 The mechanics of homoglyphs and how they disrupt tokenization.
🔹 The impact of homoglyphs on current LLM detection systems.
🔹 Cutting-edge evaluation of these methods against top detectors.
Accenture Labs
Technology Research Specialist
GDG Organizer
Software Developer, Event Manager
Senior Software Developer, Co-Organizer
Data Scientist, Instructor
The University of British Columbia
Data Scientist Co-Organizer GDG Surrey
Autodesk Inc.
Android Engineer | Co - Organizer
British Columbia Institute of Technology
Outreach Coordinator
Volunteer
Contact Us
