چكيده لاتين
Today, as production speeds up and data becomes more diverse and dispersed, the importance of distributed processing has increased. Edge computing, as a distributed solution, improves privacy, reduces communication costs, and enhances efficiency by processing tasks closer to data sources. In this context, Federated Learning in edge servers enables the training of shared models without the need to transfer local data. Furthermore, Federated Semi-Supervised Learning methods have been developed to address issues of labeling cost in these environments. Most existing methods, like Federated Averaging, incur high communication costs due to local model parameter transfer. In contrast, knowledge distillation reduces these costs by transferring only local model outputs, though it may achieve lower accuracy and often assumes shared data for training, which is usually not feasible in practice. The GFD-SSL method introduced in this study uses knowledge distillation and semi-supervised GANs without specific data distribution assumptions. This method has two stages: First, each client trains its model on local data to converge the modelʹs distribution with the local data distribution. In the next stage, each client generates synthetic data based on their data ratio and sends it to the server. Clients then compute their model outputs on this synthetic data and send these outputs to the server for aggregation. Finally, clients compare their outputs with the aggregated server results to calculate a regularization term and update their model weights. The goal is to align each clientʹs classifier output with the global aggregated output, thereby transferring global knowledge to local models.
In semi-supervised GANs, simultaneously training the discriminator to identify synthetic data and classify real data can reduce model performance. To address this issue, this study introduces the TGFD-SSL method, which separates the tasks of the discriminator and the classifier using a triple GAN technique. To improve results on Non-IID datasets, the JSA-GFDSSL method introduces a novel weighting approach using Jensen-Shannon divergence, enhancing accuracy during model output aggregation. evaluation of the proposed methods on various datasets showed that the GFD-SSL method achieved, on average, up to 15% higher accuracy, particularly on Non-IID datasets, compared to state-of-the-art methods. Additionally, the TGFD-SSL and JSA-GFDSSL methods improved model accuracy by 2.27% and 3.15%, respectively. Furthermore, the proposed method offers better optimization in terms of communication efficiency.
