A sparse adversarial attack that optimizes a mask over an I-FGSM perturbation, using a Gaussian-smoothed Heaviside step to approximate the L0 penalty, achieves sparser and faster attacks and attributes misclassification to obscuring and leading noise.
More data can expand the general- ization gap between adversarially robust and standard models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Towards Interpretable Adversarial Examples via Sparse Adversarial Attack
A sparse adversarial attack that optimizes a mask over an I-FGSM perturbation, using a Gaussian-smoothed Heaviside step to approximate the L0 penalty, achieves sparser and faster attacks and attributes misclassification to obscuring and leading noise.