Reproduce “Feature Distillation is the Better Choice for Model-Heterogeneous Federated Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yichen Li, Xiuying Wang, Wenchao Xu, Haozhao Wang, Yining Qi, Jiahua Dong, Ruixuan Li
FedFD addresses model-heterogeneous federated learning by distilling client features, rather than only final class scores, into a server model. It groups features by architecture and aligns them with learned orthogonal projections before distillation. On CIFAR-10 with α=1.0 and client models “a-d-g,” the global model reaches 89.64±0.23% accuracy, above MOON-hetero’s 87.92±0.17%; evidence is limited to tested image-classification benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.