Adversarial Gradient Saliency
A first-order, gradient-based importance measure that uses the absolute value of the loss gradient with respect to a parameter to quantify the parameter's sensitivity to the adversarial loss, employed to identify robustness-critical parameters.
Timeline (last 90 days)
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.
SAFT (Sparse Adversarial Fine-Tuning) identifies RCPs through adversarial gradient saliency, updates only these parameters, and freezes the rest.
Adversarial gradient saliency is a first-order local approximation; the saliency ranking drifts during training, and updating only the initial subset over the long term may miss parameters that become important later, so subset selection strategies and periodic re-estimation are needed.
All Facts (3)
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%UnverifiedSAFT (Sparse Adversarial Fine-Tuning) identifies RCPs through adversarial gradient saliency, updates only these parameters, and freezes the rest.
50%UnverifiedAdversarial gradient saliency is a first-order local approximation; the saliency ranking drifts during training, and updating only the initial subset over the long term may miss parameters that become important later, so subset selection strategies and periodic re-estimation are needed.
50%