Reproduce “BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised 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

Jinyuan Jia, Yupei Liu, Neil Zhenqiang Gong

BadEncoder presents the first backdoor attack specifically targeting pre-trained encoders in self-supervised learning (SSL), enabling the injection of malicious behavior into downstream classifiers built upon these compromised encoders while largely preserving their original utility. The attack consistently achieved high success rates across various datasets and real-world SSL models, including OpenAI's CLIP and Google's ImageNet encoder.

Key insights

  • It is feasible to inject effective and stealthy backdoors directly into the feature extraction capabilities of self-supervised pre-trained encoders.
  • Such backdoors propagate to any downstream classifier built on the compromised encoder, regardless of its training method (including zero-shot learning).
  • The attack maintains the utility of the backdoored encoder on clean inputs, making it difficult to detect through performance degradation, and evades state-of-the-art empirical and provable backdoor defenses.

Summary by alphaXiv.

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