ConceptRobustness-Critical Parameters / RCPs

Robustness-Critical Parameters

In adversarial training, the sparse subset of parameters identified through adversarial gradient saliency as contributing most to robustness; updating these parameters effectively improves model robustness.

Timeline (last 90 days)

Oct 30

Gradient analysis found that adversarial optimization is highly concentrated on a small number of parameters (robustness-critical parameters, RCPs), with robustness encoded in a sparse subspace.

Unverified50%

All Facts (1)

Unverified

Gradient analysis found that adversarial optimization is highly concentrated on a small number of parameters (robustness-critical parameters, RCPs), with robustness encoded in a sparse subspace.

50%

Source Articles

Sparse Fine-Tuning SAFT: Updating Only 5% of Parameters for Stronger Adversarial RobustnessOct 30