Expand and Shrink: Federated Learning with Unlabeled Data Using Clustering.

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• The integration of the Internet of Things (IoT) with federated learning (FL) promises advanced deep learning while preserving data privacy.
• Current FL models typically require labeled client data for supervised classification, which is often impractical in real-world IoT scenarios.
• Existing methods for handling unlabeled data in FL, like class-prior probabilities or pseudo-labeling, rely on unrealistic or unavailable assumptions.

Purpose of the Study:

• To investigate the feasibility of performing federated learning with unlabeled data in the IoT.
• To propose a novel clustering-based method for client-side data labeling prior to FL.
• To develop a universally applicable solution for classification tasks within the FL framework.

Main Methods:

• Implemented a clustering-based approach for sample labeling directly on the client device before federated training.
• Conducted experiments varying the ratio of labeled data, the number of clusters, and client participation rates.
• Evaluated the performance of the proposed method across different experimental conditions.

Main Results:

• Achieved high accuracy rates of 87% and 90% using minimal amounts of true labels (0.01 and 0.03, respectively).
• Demonstrated the effectiveness of the clustering-based labeling strategy in an FL context.
• Validated the method's suitability for various classification tasks.

Conclusions:

• The proposed clustering-based labeling method enables effective federated learning with unlabeled data, addressing a key limitation in current FL architectures.
• This approach enhances data privacy in IoT environments by allowing labeling at the source.
• The method offers a practical and adaptable solution for privacy-preserving deep learning in diverse classification tasks.