Reproduce “Anti-Backdoor Learning: Training Clean Models on Poisoned Data” on OpenResearch

About this paper

Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, Xingjun Ma

Researchers at Xidian University, University of Copenhagen, Sony AI, University of Illinois at Urbana–Champaign, and Deakin University introduce Anti-Backdoor Learning (ABL), a method for training clean deep neural networks directly on poisoned datasets by actively preventing backdoor injection. ABL exploits the observed faster learning rate of backdoor tasks and their target-class dependency, achieving average Attack Success Rates (ASR) as low as 6.00% on an ImageNet subset while maintaining high clean accuracy across diverse attacks.

Key insights

• Backdoor tasks are inherently 'easier' for deep models to learn than clean tasks, causing their training loss to decrease significantly faster and earlier than clean examples.
• The strong correlation between a backdoor trigger and a specific target class represents a fundamental weakness that can be exploited for effective unlearning.
• By dynamically separating and counteracting the distinct learning behaviors of clean and poisoned data during training, it is possible to train clean models on backdoored datasets.

Summary by alphaXiv.