Multi-head ensemble of smoothed classifiers for certified robustness

Kun Fang, Qinghua Tao, Yingwen Wu, Tao Li, Xiaolin Huang, Jie Yang

Abstract: Highlights•An effective and efficient certified defense outperforming those ensembling DNNs.•A multi-head structure with a cosine constraint mitigates heavy computation costs.•The communication between classifiers via circular-teaching for better ensemble.•Easy and hard samples under this certified robustness task are discussed.•Circular-teaching among heads via self-paced learning boosts certified robustness.