[Paper Review] Federated clustering with GAN-based data synthesis

Jie Yan, Jing Liu+2 more|arXiv (Cornell University)|Oct 29, 2022

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper proposes SDA-FC, a federated clustering framework that uses client-local GANs to generate synthetic global data, enabling effective global similarity learning without sharing private data. The method achieves superior clustering performance and robustness against non-IID data and device failures by training K-means or FCM on the synthetic global dataset, outperforming existing baselines in both accuracy and resilience.

ABSTRACT

Federated clustering (FC) is an extension of centralized clustering in federated settings. The key here is how to construct a global similarity measure without sharing private data, since the local similarity may be insufficient to group local data correctly and the similarity of samples across clients cannot be directly measured due to privacy constraints. Obviously, the most straightforward way to analyze FC is to employ the methods extended from centralized ones, such as K-means (KM) and fuzzy c-means (FCM). However, they are vulnerable to non independent-and-identically-distributed (non-IID) data among clients. To handle this, we propose a new federated clustering framework, named synthetic data aided federated clustering (SDA-FC). It trains generative adversarial network locally in each client and uploads the generated synthetic data to the server, where KM or FCM is performed on the synthetic data. The synthetic data can make the model immune to the non-IID problem and enable us to capture the global similarity characteristics more effectively without sharing private data. Comprehensive experiments reveals the advantages of SDA-FC, including superior performance in addressing the non-IID problem and the device failures.

Motivation & Objective

• To address the challenge of learning global similarity in federated clustering without sharing private client data.
• To improve clustering robustness and performance under non-independent-and-identically-distributed (non-IID) data distributions across clients.
• To enhance resilience to device failures in federated clustering systems.
• To provide a privacy-preserving alternative to centralized clustering that maintains high clustering quality.
• To evaluate and advocate for the Kappa metric as a more reliable alternative to NMI for clustering evaluation.

Proposed method

• Each client trains a local GAN on its private data to generate synthetic samples.
• The server aggregates synthetic data from all clients to construct a global synthetic dataset.
• K-means or fuzzy c-means is applied on the global synthetic dataset to compute global cluster centroids.
• These global centroids are sent back to clients for local cluster assignment based on proximity.
• The framework supports one-round communication and asynchronous execution, minimizing overhead.
• The method is evaluated across multiple datasets (MNIST, Fashion-MNIST, CIFAR-10, STL-10) under varying non-IID levels and device failure rates.

Experimental results

Research questions

• RQ1 Can GAN-based synthetic data effectively approximate the global data distribution in federated clustering without sharing private data?
• RQ2 How does the proposed SDA-FC framework perform compared to state-of-the-art federated clustering methods under non-IID data conditions?
• RQ3 To what extent is the clustering performance of SDA-FC robust to device failures in federated systems?
• RQ4 How does the non-IID level affect the sensitivity of clustering performance to device disconnections?
• RQ5 Is the Kappa metric a more reliable evaluation metric than NMI for assessing federated clustering results?

Key findings

• SDA-FC outperforms existing methods, including k-FED and FFCM, in clustering performance across all datasets, especially under high non-IID levels.
• The global synthetic dataset generated by SDA-FC closely approximates the real global dataset, as confirmed by t-SNE visualization and distribution overlap.
• SDA-FC demonstrates superior robustness to device failures, with performance degradation being less severe than baseline methods.
• The sensitivity of clustering performance to device failures increases with higher non-IID levels (higher p values), indicating greater data heterogeneity reduces data substitutability.
• The Kappa metric is found to be more reliable than NMI for evaluating clustering results, as NMI values are deemed potentially misleading in some scenarios.
• The framework achieves high performance with minimal communication overhead, requiring only one round of aggregation and supporting asynchronous execution.