Reproduce “Generating Adversarial Attacks in the Latent Space” 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
Nitish Shukla, Sudipta Banerjee
A new method for generating adversarial attacks operates on the latent feature representations of images, eliminating the need for traditional epsilon-bound constraints while maintaining high visual realism. This approach achieves over 90% attack success rates on MNIST and over 85% on CIFAR10.
Key insights
- Adversarial attacks can be effectively generated by manipulating the latent feature representations of images, rather than pixel values, circumventing the need for epsilon-bound constraints.
- The process can be geometrically interpreted as moving the latent representation of an input from its original class's convex hull towards a target class's convex hull in the feature space.
- Manipulating features that neural networks directly utilize for decision-making can produce more effective and visually realistic adversarial examples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.