Reproduce “BackdoorBench: A Comprehensive Benchmark of Backdoor Learning” 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

Baoyuan Wu, Hongrui Chen, Mingda Zhang, Zihao Zhu, Shaokui Wei, Danni Yuan, Chao Shen

BackdoorBench establishes a comprehensive and reproducible benchmark for evaluating backdoor attacks and defenses in deep neural networks, providing over 8,000 standardized evaluations across diverse settings. The work uncovers insights into the counter-intuitive effects of poisoning ratios, the rapid learning and forgetting dynamics of backdoors, and the vulnerabilities of Vision Transformers.

Key insights

  • Higher poisoning ratios do not always lead to stronger attacks; for some defenses, increased poisoning can significantly reduce Attack Success Rate (ASR).
  • Backdoors are learned very rapidly during the early training epochs, exhibiting faster loss reduction and higher gradient signal-to-noise ratios compared to clean samples.
  • A phenomenon called 'trigger generalization' exists, where backdoored models can be activated by triggers different from those used during training.
  • The effectiveness of backdoor attacks and defenses is highly dependent on the specific model architecture, dataset characteristics, and hyper-parameter settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.