LAFED: Towards robust ensemble models via Latent Feature Diversification

Wenzi Zhuang, Lifeng Huang, Chengying Gao, Ning Liu

Abstract: Highlights•We observe that a high similarity between the training sets of ensemble members typically results in weak robustness.•We introduce an unbalanced feature combination strategy aimed at diminishing the similarity of captured features within the ensemble members.•We design hierarchical label smoothing strategy for guiding sub-models to learn latent representations scattered with different degrees.•Our method enhances both white-box and black-box adversarial robustness without compromising clean accuracy.•Our method exhibits strong generalizability, as the inclusion of more members effectively enhances the robustness of ensemble.