Federated Robustness Propagation: Sharing Adversarial Robustness in Heterogeneous Federated Learning

Federated learning (FL) emerges as a popular distributed schema that learns model from set of participating users without sharing raw data. One major challenge FL comes with heterogeneous users, who may have distributionally different (or non-iid) data and varying computation resources. As federated would use the for prediction, they often demand trained to be robust against malicious attackers at test time. Whereas adversarial training (AT) provides sound solution centralized learning, extending its usage has imposed significant challenges, many very limited tight computational budgets, afford data-hungry costly AT. In this paper, we study novel strategy: propagating robustness rich-resource can AT, those poor resources cannot it, during learning. We show existing techniques effectively integrated strategy propagate among non-iid propose an efficient propagation approach by proper batch-normalization. demonstrate rationality effectiveness our method through extensive experiments. Especially, proposed is shown grant models remarkable even when only small portion AT Source code accessed https://github.com/illidanlab/FedRBN.