Problem
Research questions and friction points this paper is trying to address.
Enhances certified robustness via multi-head ensemble
Reduces computational costs in training and certification
Improves head diversity and mutual learning efficiency
Multi-head ensemble of smoothed classifiers for certified robustness.
🤖 AI Summary
While multi-model ensembles in randomized smoothing (RS) improve ℓ₂-robustness certification radii, they incur substantial training and certification overhead and neglect inter-classifier collaboration.
Method: We propose Multi-Head Ensemble Smoothing (ME-Smooth), a novel RS framework that deploys multiple parallel noise-injection paths and classification heads within a single deep neural network. It jointly optimizes feature perturbation robustness and decision diversity, introducing the multi-head architecture to RS for the first time. This design decouples noise-robust feature modeling from classification decisions, eliminating additional inference cost.
Results: On CIFAR-10 and an ImageNet subset, ME-Smooth achieves up to 8.2% higher certified accuracy compared to standard RS baselines. At equal certified ℓ₂ radii, it significantly outperforms both single-head RS and state-of-the-art certified defense methods.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Single DNN with multiple augmented heads
Circular communication flow among heads
Cosine constraint for variance reduction
🔎 Similar Papers
No similar papers found.
