Dear Colleagues,
We are pleased to announce the Anti-Backdoor Challenge (Anti-BAD) at IEEE SaTML 2026.
Anti-BAD addresses LLM backdoor defense in deployment-oriented post-training settings. It presents a practical and timely challenge that aims to promote the development of lightweight and effective defense methods capable of restoring model integrity while preserving clean-task utility in realistic model-sharing ecosystems.
The competition has been released on Codabench (https://www.codabench.org/competitions/11188/), and the development phase will start on November 7, 2025, inviting everyone to participate and test their defense methods.
Competition Website: https://anti-bad.github.io/
=== Tracks ===
Track 1: Generation (English)
Track 2: Classification (English)
Track 3: Multilingual Classification (35+ languages)
These tracks represent key application scenarios of large language models, covering both generation and classification tasks across English and multilingual settings. Each track provides several backdoored models, each poisoned by a distinct and undisclosed method. We challenge participants to design robust and generalizable model-wise defenses in a post-training setting.
=== Timeline ===
Registration opens: October 21, 2025
Development phase: November 7, 2025
Test phase: February 1-7, 2026
Final results announcement: February 8, 2026
=== More Information ===
1. IEEE SaTML Competitions: https://satml.org/competitions/
2. Discord Channel for Discussion: https://discord.gg/x8GqKDF2Rb
Best regards,
Anti-BAD Challenge Organizing Team