Deep Embedded Clustering of Urban Communities Using Federated Learning
Afra J. Mashhadi, Joshua Sterner, Jeffrey Murray · 2021
Deep clustering utilizes representation learning to learn features in an unsupervised setting. Although successful, the current models rely on the assumption of the centralized dataset, which due to the privacy concerns is becoming less realistic. To address this challenge, we propose a federated deep convolutional embedded clustering framework. Our framework relies on a federated server to orchestrate the training between workers where each participant individually trains the model with the objective of decreasing clustering loss using Kullback–Leibler divergence. To avoid feature space being distorted by the clustering loss, each worker maintains their own local decoder which for privacy reasons is not shared with the federated server. Empirical results with both IID and non-IID client data on benchmark datasets demonstrates the feasibility of our federated training when compared to the centralized counterpart. We also evaluate our model on a real world application of community detection using GPS traces and measure the computational complexity and energy consumption on a smartphone.