GRAPE: Guided Parameter-Space Evolution for Compact Adversarial Robustness

In three linesGRAPE proposes an adversarial training method that progressively exposes network parameters rather than optimizing a fixed space. On CIFAR-10 under ℓ∞, GRAPE improves ResNet-18 PGD-20 robust accuracy from 51.70% to 56.94% with 21.4% fewer parameters and nearly matched computation budget (1.009x FLOPs).

Summary generated by Claude — human-verified