Reproduce “Neural Fingerprints for Adversarial Attack Detection” 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
Haim Fisher, Moni Shahar, Yehezkel S. Resheff
This work by Fisher, Shahar, and Resheff presents neural fingerprints, a randomized method for detecting adversarial examples that remains effective even when attackers possess full knowledge of the defense system. Leveraging distinctive patterns in neural activations, the approach achieved over 93% detection rates with low false positives on ImageNet against strong attacks like PGD, demonstrating its robustness and practicality.
Key insights
- Randomizing the detection mechanism at test time can effectively counter white-box adversarial attacks, making it difficult for attackers to craft examples that bypass the defense.
- Neural activations within pre-trained networks contain distinct patterns that can be leveraged to differentiate between clean and adversarial examples.
- A small number of carefully selected neural fingerprints (around 20-40) can provide high detection accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.