Reproduce “ProtoFL: Unsupervised Federated Learning via Prototypical Distillation” 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

Hansol Kim, Youngjun Kwak, Minyoung Jung, Jinho Shin, Youngsung Kim, Changick Kim

ProtoFL introduces an unsupervised federated learning framework that combines prototypical distillation with local normalizing flows to address one-class classification in extreme non-I.I.D. data settings. The method significantly enhances performance and communication efficiency, achieving state-of-the-art AUROC on various benchmarks while maintaining privacy.

Key insights

  • Decoupling representation learning from one-class classification allows for a more efficient and scalable federated learning process in privacy-sensitive settings.
  • Prototypical distillation, involving a one-time distribution of class-specific prototypes, effectively handles extreme non-I.I.D. data and limited local samples without continuous data or parameter sharing.
  • Utilizing normalizing flows as local one-class classifiers provides robust density estimation for normal classes, even with scarce local data, enabling powerful anomaly detection.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.