search

ybdai7 / Chameleon-durable-backdoor

[ICML 2023] Official code implementation of "Chameleon: Adapting to Peer Images for Planting Durable Backdoors in Federated Learning (https://proceedings.mlr.press/v202/dai23a)"
29 stars 4 forks source link

low accuracy about backdoor #11

Open lu0802mark opened 2 weeks ago

lu0802mark commented 2 weeks ago

Hello, I would like to know how the backdoor accuracy can reach 100% without defenses in the case of a semantic backdoor. When I perform a single attack by a single adversary after the model has converged, the backdoor accuracy is only around 85%.

ybdai7 commented 1 week ago

Hi,

Basically we assume that the local learning rates for benign clients decline as the global round, while the lr for poisoned client could stay the same. Thus, the backdoor accuracy could reach 100% eventually.