Generating Adversarial Attacks in the Latent Space
Nitish Shukla, Sudipta Banerjee; In Proceedings of IEEE/CVF CVPR Workshops 2023
Research Goal
Most adversarial examples add imperceptible pixel-level noise. This work asks whether semantically meaningful adversarial examples can be produced by perturbing a generative model’s latent space instead.
How it works
- Latent perturbations. Adversarial perturbations are injected into the latent space of a generative model, so the resulting changes are structured rather than pixel noise.
- Probing decision boundaries. The approach shows how a network’s decision boundaries can be probed through structured manipulations of latent representations.