Federated clustering sends only density peaks: one round, and shuffled contributions the server cannot trace back
Synopsis
Federated clustering previously revealed more model information to the server each round, and even without raw samples, repeated exchanges made client contributions linkable. eFDPC combines evolving fuzzy clustering with density peak clustering in a federated setting: each client evolves micro-clusters from private mini-batches and transmits only density-peak candidates as additive shares through a shuffler layer, so the server reconstructs peaks after mixing and selects global peaks. On synthetic and real benchmarks, it outperforms recent federated clustering baselines in one-shot settings and scales to higher-dimensional, non-IID data.
Interpretation
Federated clustering can now run in a single communication round, with clients sending a small set of density-peak candidates instead of full local models. Earlier federated clustering methods were iterative, repeating local reclustering and server aggregation over multiple rounds, which raised communication cost and made prototype trajectories linkable across rounds. On synthetic and real benchmark datasets, eFDPC outperforms recent federated clustering baselines in one-shot settings and scales to higher-dimensional, non-IID data.
Each client first evolves micro-clusters incrementally from private mini-batches, then applies density peak selection on those micro-clusters and transmits only the selected peaks. Evolving clustering originally learned sample by sample online; here it is modified to update in parallel over mini-batches, matching the federated constraint that per-sample updates cannot be sent. A shuffler layer sits between clients and server: each client splits every peak statistic into additive shares, shufflers permute the messages, and the server reconstructs peaks only after mixing, so it does not know which client produced each reconstructed peak.
Perspective
The method targets unsupervised clustering in a horizontal federated setting, where clients hold local data and the server is honest-but-curious; privacy protection focuses on not exposing raw samples and on keeping client contributions unlinkable. Experiments cover synthetic datasets and real benchmarks, including the high-dimensional MNIST and KDD99, with sensitivity results over client count, batch size, and non-IID level. A PyTorch implementation is publicly available.
When the decision graph is flat and no clear density peaks exist, peak selection and label propagation can be sensitive to local perturbations and neighborhood construction. Overall quality depends on hyperparameters such as merging aggressiveness and neighborhood size, which must be chosen consistently across datasets. Peak summaries still add communication overhead relative to pure centroid-only baselines, and the shuffler anonymization layer adds system complexity and communication overhead. The number of peaks P is given as the true cluster count in experiments, which is unknown in practice; the authors note that no unsupervised criterion recovers the true number of clusters reliably.