Propose with Gradients, Accept with Loss: A Hedged Attack for Pixel-Sparse Robustness
Propose with the Gradient, Accept with the Loss: One Hedged Attack for Pixel-Sparse Robustness
Hon Ming Yan and Chun Pong Lau propose the pixel-sparse adversarial attack GRS (Gradient-guided Random Search), which matches or exceeds the best single member of sparse AutoAttack (sAA) on 13 CIFAR-10 configurations and 5 ImageNet models. In this setting, the attacker can change a small number of pixels in an image to arbitrary values; the paper unifies existing sPGD and Sparse-RS as local search with a proposal distribution plus an acceptance rule, deriving five rules from Taylor bounds and measurements at the search state, for example setting new pixel values by the gradient sign and accepting based only on loss. GRS needs only one run, with one forward and one backward pass per step, using 3 rather than up to 7 forward units, and is within one point of or stronger than sAA on 17 of 18 settings, significantly stronger on 8 of them, and 3.7 to 5.0 times faster than the full-run sAA members on CIFAR-10.