Reproduce “Mix2FLD: Downlink Federated Learning After Uplink Federated Distillation With Two-Way Mixup” 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
Seungeun Oh, Jihong Park, Eunjeong Jeong, Hyesung Kim, Mehdi Bennis, Seong-Lyun Kim
A new framework, Mix2FLD, integrates federated learning and federated distillation to optimize model training over asymmetric wireless channels. It employs a two-way Mixup mechanism to enable privacy-preserving server-side knowledge distillation, resulting in enhanced accuracy and faster convergence, particularly with non-IID data.
Key insights
- The "two-way Mixup" (Mix2up) mechanism effectively resolves the seed sample dilemma in FLD, enabling privacy-preserving generation of hard-labeled synthetic data for server-side knowledge distillation.
- Integrating federated distillation for uplink and federated learning for downlink effectively balances communication efficiency and accuracy in asymmetric wireless environments.
- The inverse-Mixup at the server not only enhances privacy but also serves as a data augmentation method, generating a larger and more robust set of synthetic seed samples for global model training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.