Propose with the Gradient, Accept with the Loss: One Hedged Attack for Pixel-Sparse Robustness
A single attack matched or exceeded the sparse-AutoAttack ensemble’s strength within one point in 17 of 18 tested settings.
Gradient-guided Random Search tests image classifiers by changing only a few pixels, using gradients to propose changes and loss alone to accept them. Across 13 CIFAR-10 configurations and five ImageNet models, the authors report significantly stronger attacks than the ensemble in eight settings. Each step requires one forward and one backward pass. On CIFAR-10, it ran 3.7–5.0 times faster than running the ensemble’s members in full.
ArXiv · AI/CL/LG's note
Gradient-guided Random Search tests image classifiers by changing only a few pixels, using gradients to propose changes and loss alone to accept them. Across 13 CIFAR-10 configurations and five ImageNet models, the authors report significantly stronger attacks than the ensemble in eight settings. Each step requires one forward and one backward pass. On CIFAR-10, it ran 3.7–5.0 times faster than running the ensemble’s members in full.
ArXiv · AI/CL/LG's note
score 4