Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training
Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training
Tiansheng Huang,Sihao Hu,3 作者,Ling Liu
2023 · DBLP: conf/nips/HuangHCIT023
Neural Information Processing Systems · 引用 23 次
TLDR
Lockdown is presented, an isolated subspace training method to mitigate the poison-coupling effect in Federated learning, and achieves superior and consistent defense performance compared to existing representative approaches against backdoor attacks.