GRAPE: Guided Parameter-Space Evolution for Compact Adversarial Robustness

The paper introduces GRAPE (Guided Parameter-Space Evolution), a novel training framework aimed at enhancing adversarial robustness in neural networks by optimizing the order of parameter exposure during training. GRAPE achieves a PGD-20 robust accuracy of 56.94% on CIFAR-10 using a fixed-structure ResNet-18, improving upon the baseline accuracy of 51.70% while reducing the parameter count by approximately 21.4% and maintaining a similar computational budget (FLOPs ratio of 1.009x). This approach underscores the importance of parameter-space evolution in achieving compact and robust model configurations, offering valuable insights for practitioners focused on adversarial training strategies.