Sparse Fine-Tuning for Parameter-Efficient Adversarial Training
Updates only robustness-critical parameters, about 5 percent, during adversarial training.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv