The proliferation of digital artifacts with various computing capabilities, along with the emergence of edge computing, offers new possibilities for the development of Machine Learning solutions.These new possibilities have led to the popularity of Federated Learning (FL).While there are many existing works focusing on various aspects of the FL process, the issue of the effective problem diagnosis in FL systems remains largely unexplored.In this work, we have set out to artificially simulate the training process of four selected approaches to FL topology and compare their resulting performance.After noticing concerning disturbances throughout their training process, we have successfully identified their source as the problem of exploding gradients.We have then made modifications to the model structure and analyzed the new results.Finally, we have proposed continuous monitoring of the FL training process through the local computation of a selected metric.

Federated learning (FL) allows multiple devices to jointly train a global model without sharing local data.One of its problems is dealing with unbalanced data.Hence, a novel technique, designed to deal with label-skewed non-IID data, using adversarial inputs is proposed.Application of the proposed algorithm results in faster, and more stable, global model performance at the beginning of the training.It also delivers better final accuracy and decreases the discrepancy between the performance of individual classes.Experimental results, obtained for MNIST, EMNIST, and CIFAR-10 datasets, are reported and analyzed.

Federated learning (FL) is a decentralized approach that aims at training a global model with the help of multiple devices, without collecting or revealing individual clients' data.The training of a federated model is conducted in communication rounds.Still, in certain scenarios, numerous communication rounds are impossible to perform.In such cases, a one-shot FL is utilized, where the number of communication rounds is limited to one.In this article, the idea of one-shot FL is enhanced with the usage of adversarial data, exploring and illustrating the possibilities to improve the performance of resulting global models, including scenarios with non-IID data, for image classification datasets: MNIST and CIFAR-10.