Reproduce “Long-Tailed Backdoor Attack Using Dynamic Data Augmentation Operations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lu Pang, Tao Sun, Weimin Lyu, Haibin Ling, Chao Chen
Researchers from Stony Brook University developed Dynamic Data Augmentation Operations (D²AO), the first backdoor attack method specifically designed for long-tailed datasets, achieving state-of-the-art attack success rates while maintaining clean accuracy and exposing limitations in current backdoor defenses.
Key insights
- Strong data augmentations, often employed in long-tailed learning, can inadvertently neutralize backdoor triggers, necessitating adaptive augmentation strategies for backdoored samples.
- Optimal backdoor attack efficacy in long-tailed settings relies on dynamically balancing augmentations for clean classes (to improve overall accuracy) and preserving trigger integrity for backdoored samples (through weaker, targeted augmentations).
- Many existing backdoor detection and mitigation methods are ineffective or inapplicable against long-tailed backdoor attacks, highlighting a critical blind spot in current AI security research.
- results
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.