GRAPE boosts adversarial robustness 5% with 21% fewer parameters
New training method evolves parameters progressively for leaner, tougher models
A team from the University of Science and Technology of China and China Mobile has introduced GRAPE (Guided Parameter-Space Evolution), a novel adversarial training framework that questions the conventional wisdom of training all parameters from the start. Instead, GRAPE treats robust model learning as a progressive process: it first stabilizes optimization in a small exposed parameter space, then gradually releases new dimensions guided by an adversarial spectral utilization score. This score directs newly freed capacity toward high-pressure modules, ensuring efficient allocation.
On CIFAR-10 under the standard ℓ∞ threat model with a fixed ResNet-18 architecture, GRAPE achieves 56.94% PGD-20 robust accuracy—a 5.24% improvement over standard adversarial training (51.70%)—while reducing parameters by 21.4% at nearly identical FLOPs (1.009x). A variant with sequential growth reaches 56.52%, proving the benefit comes from the exposure path, not just architecture. This suggests that guided parameter evolution can deliver both compactness and robustness.
- GRAPE improves PGD-20 robust accuracy from 51.70% to 56.94% on CIFAR-10 with ResNet-18
- Parameter count reduced by 21.4% at near-identical compute (FLOPs ratio 1.009x)
- Uses adversarial spectral utilization to guide new parameters toward high-pressure modules
Why It Matters
Enables more efficient, compact robust models without sacrificing accuracy—critical for deploying AI in security-sensitive, resource-constrained environments.