GRAPE Improves Adversarial Robustness with Parameter-Space Evolution

This research introduces GRAPE (Guided Parameter-Space Evolution), a novel training framework aimed at developing more compact and robust neural networks against adversarial attacks. Unlike traditional adversarial training (AT) methods that optimize a fixed parameter space from the outset, GRAPE explores whether the sequence in which parameters become optimizable can influence the final robust model, even when the architecture and computational budget are similar. GRAPE operates by stabilizing robust optimization within the currently exposed parameter space, then gradually releasing new optimizable dimensions. It uses an adversarial spectral utilization score to guide this new capacity towards high-pressure modules, effectively evolving the model's parameter space. On the CIFAR-10 dataset under an L-infinity threat model, GRAPE significantly improved PGD-20 robust accuracy from 51.70% to 56.94% with a comparable computational budget, while also reducing the parameter count by over 21%. These results indicate that a guided evolution of the parameter space can lead to more robust and efficient model configurations.