Adversarial training with structured confidence perturbation based on Fisher geometry
Start here
On November 10, an arXiv paper proposed adversarial training with structured confidence perturbation based on Fisher geometry: targeting pseudo-attacks that raise loss without changing the confidence structure, it generates perturbations on the confidence manifold that maximize distributional distance using a mixed objective of cross-entropy and KL divergence; theoretical analysis shows that the optimal perturbation direction is consistent with the principal eigenvector of the Fisher information matrix, and experiments show its adversarial robustness and robust transferability outperform existing methods.
Generated by AI from reports · Updated 2 days ago
Related research
This event's paper
- Structured Confidence Perturbation via Fisher Geometry for Adversarial Training
arXiv · 2026-11-10 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningStructured confidence perturbation adversarial training based on Fisher geometry
To address "pseudo-attacks" in adversarial training that push up loss without changing the confidence structure, researchers propose a mixed objective of cross-entropy and KL divergence, and generate perturbations on the confidence manifold that maximize distributional distance. Theoretical analysis shows that the optimal perturbation direction is consistent with the principal eigenvector of the Fisher information matrix. Experiments on multiple datasets and model architectures show that the method's adversarial robustness and robust transferability outperform existing methods, and it is especially good at structured perturbations within high-confidence regions.
Related discussion
Attention trend
New sources per day
- Nov 11: 1 new sources (1 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)