New method for grouping data while keeping information private

A new approach called Evolving Federated Density Peak Clustering uses a shuffler to protect data during one-shot learning.

Researchers developed Evolving Federated Density Peak Clustering, or eFDPC. This method uses a one-shot federated clustering approach. It combines evolving fuzzy data-stream clustering with density peak clustering.

In federated clustering, clients do not share raw samples. However, each round can reveal model information. This may increase the chance to infer properties of local datasets.

The eFDPC method works by having each client evolve micro-clusters from private data mini-batches. These micro-clusters are created through cluster addition and merging. Density peak clustering is then applied to these micro-clusters.

To limit communication, only selected peaks are transmitted instead of full local models. A shuffler-based layer sits between clients and the server. This layer anonymizes contributions before global peak selection.

The researchers tested a peak-only variant called eFDPC-P. They also tested eFDPC-G, which uses server-side label propagation. Experiments on synthetic and real-world benchmark datasets show eFDPC outperforms recent federated clustering baselines.

Key facts

  • eFDPC uses a shuffler-based layer to anonymize contributions.
  • The method works in a one-shot federated learning paradigm.
  • eFDPC scales to higher-dimensional, non-IID data.

All sections

Edition