Efficient Federated Learning on Resource-Constrained Edge Devices: Integrating Data Distillation and Semi-Supervised Learning

Abstract

This paper presents a Dataset Distilled SemiSupervised Federated Learning (DD-SSFL) approach aiming at optimizing federated learning efficiency and safeguarding privacy through the integration of labeled data distillation techniques and the exploitation of unlabelled data at edge devices. Our approach harnesses dataset distillation on labeled data locally stored within edge devices, facilitating the generation of distilled datasets transmitted to a central server for model training. Upon training, the model is distributed to edge devices for local inference tasks. Post-inference, edge devices utilize the trained model to predict labels for unlabeled data, employing a straightforward pseudolabeling mechanism with a predefined threshold. Subsequently, the newly pseudo-labeled data is distilled and transmitted to the server for further model training iterations. This iterative process enables the model to leverage both labeled and unlabeled data effectively, resulting in refined models. Moreover, this approach significantly reduces the number of communication rounds, minimizes transmitted data size, and alleviates processing burdens on edge devices, thereby demonstrating its efficacy in resourceconstrained edge computing environments. The simulation results indicate that the DD-SSFL approach consistently achieves higher accuracy with fewer communication rounds, smaller transmitted data size, and reduced resource requirements compared to traditional federated learning schemes.

Authors 4

1. Mahdi Barhoush Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH University,Chair of Information Theory and Data Analytics (INDA),Aachen,Germany

2. Ahmad Ayad Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH University,Chair of Information Theory and Data Analytics (INDA),Aachen,Germany

3. Mohammad Kohankhaki Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH University,Chair of Information Theory and Data Analytics (INDA),Aachen,Germany

4. Anke Schmeink Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH University,Chair of Information Theory and Data Analytics (INDA),Aachen,Germany