A privacy-preserving clustering method (FDSC) that runs deep subspace clustering across many clients under a federated-learning schema. Each client trains an autoencoder with a self-expressive layer that learns an affinity matrix for spectral clustering; only the encoder is shared and weighted-averaged (FedAvg-style) into a global model, while the self-expressive layer and decoder remain private, so raw data never leaves the client. A locality-preservation regularizer aligns each client's affinity matrix with its local adjacency (graph) matrix to keep neighborhood structure, boosting feature extraction and clustering quality; demonstrated to beat state-of-the-art baselines on four image datasets.
A machine-learning methods paper (cs.LG) by Zhang, Feng, Wang and Shang introducing FDSC (Federated Deep Subspace Clustering), a privacy-preserving approach that runs deep subspace clustering under a…