Reproduce “The Limitations of Deep Learning in Adversarial Settings” 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
Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, Ananthram Swami
This paper presents a forward derivative-based attack that efficiently generates adversarial examples for deep neural networks, achieving a 97.10% misclassification rate on MNIST by modifying only 4.02% of input features. The work demonstrates that these adversarial inputs, while fooling machines, remain highly recognizable to human observers.
Key insights
- Deep neural networks are inherently vulnerable to small, carefully crafted adversarial perturbations that can cause misclassification.
- Adversarial saliency maps reveal that only a small subset of input features often needs to be modified to achieve a desired misclassification.
- A fundamental disconnect exists between human and machine perception; adversarial examples that fool machines often remain easily recognizable to humans.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.