While Federated Learning (FL) provides a privacy-preserving approach to analyze sensitive data without centralizing training data, the field lacks an detailed comparison of emerging open-source FL frameworks. Furthermore, there is currently no standardized, weighted evaluation scheme for a fair comparison of FL frameworks that would support the selection of a suitable FL framework. This study addresses these research gaps by conducting a comparative analysis of 15 individual open-source FL frameworks filtered by two selection criteria, using the literature review methodology proposed by Webster and Watson. These framework candidates are compared using a novel scoring schema with 15 qualitative and quantitative evaluation criteria, focusing on features, interoperability, and user friendliness. The evaluation results show that the FL framework Flower outperforms its peers with an overall score of 84.75%, while Fedlearner lags behind with a total score of 24.75%. The proposed comparison suite offers valuable initial guidance for practitioners and researchers in selecting an FL framework for the design and development of FL-driven systems. In addition, the FL framework comparison suite is designed to be adaptable and extendable accommodating the inclusion of new FL frameworks and evolving requirements.

Given the inherent fluctuation of photovoltaic (PV) generation, accurately forecasting solar power output and grid feed-in is crucial for optimizing grid operations. Data-driven methods facilitate efficient supply and demand management in smart grids, but predicting solar power remains challenging due to weather dependence and data privacy restrictions. Traditional deep learning (DL) approaches require access to centralized training data, leading to security and privacy risks. To navigate these challenges, this study utilizes federated learning (FL) to forecast feed-in power for the low-voltage grid. We propose a bottom-up, privacy-preserving prediction method using differential privacy (DP) to enhance data privacy for energy analytics on the customer side. This study aims at proving the viability of an enhanced FL approach by employing three years of meter data from three residential PV systems installed in a southern city of Germany, incorporating irradiance weather data for accurate PV power generation predictions. For the experiments, the DL models long short-term memory (LSTM) and gated recurrent unit (GRU) are federated and integrated with DP. Consequently, federated LSTM and GRU models are compared with centralized and local baseline models using rolling 5-fold cross-validation to evaluate their respective performances. By leveraging advanced FL algorithms such as FedYogi and FedAdam, we propose a method that not only predicts sequential energy data with high accuracy, achieving an R2 of 97.68%, but also adheres to stringent privacy standards, offering a scalable solution for the challenges of smart grids analytics, thus clearly showing that the proposed approach is promising and worth being pursued further.

The utilization of advanced intraoral scanners to acquire 3D dental models has gained significant popularity in the fields of dentistry and orthodontics. Accurate segmentation and labeling of teeth on digitized 3D dental surface models are crucial for computer-aided treatment planning. At the same time, manual labeling of these models is a time-consuming task. Recent advances in geometric deep learning have demonstrated remarkable efficiency in surface segmentation when applied to raw 3D models. However, segmentation of the dental surface remains challenging due to the atypical and diverse appearance of the patients' teeth. Numerous deep learning methods have been proposed to automate dental surface segmentation. Nevertheless, they still show limitations, particularly in cases where teeth are missing or severely misaligned. To overcome these challenges, we introduce a network operator called dilated edge convolution, which enhances the network's ability to learn additional, more distant features by expanding its receptive field. This leads to improved segmentation results, particularly in complex and challenging cases. To validate the effectiveness of our proposed method, we performed extensive evaluations on the recently published benchmark data set for dental model segmentation Teeth3DS. We compared our approach with several other state-of-the-art methods using a quantitative and qualitative analysis. Through these evaluations, we demonstrate the superiority of our proposed method, showcasing its ability to outperform existing approaches in dental surface segmentation.

Personal health data is subject to privacy regulations, making it challenging to apply centralized data-driven methods in healthcare, where personalized training data is frequently used. Federated Learning (FL) promises to provide a decentralized solution to this problem. In FL, siloed data is used for the model training to ensure data privacy. In this paper, we investigate the viability of the federated approach using the detection of COVID-19 pneumonia as a use case. 1411 individual chest radiographs, sourced from the public data repository COVIDx8 are used. The dataset contains radiographs of 753 normal lung findings and 658 COVID-19 related pneumonias. We partition the data unevenly across five separate data silos in order to reflect a typical FL scenario. For the binary image classification analysis of these radiographs, we propose ResNetFed, a pre-trained ResNet50 model modified for federation so that it supports Differential Privacy. In addition, we provide a customized FL strategy for the model training with COVID-19 radiographs. The experimental results show that ResNetFed clearly outperforms locally trained ResNet50 models. Due to the uneven distribution of the data in the silos, we observe that the locally trained ResNet50 models perform significantly worse than ResNetFed models (mean accuracies of 63% and 82.82%, respectively). In particular, ResNetFed shows excellent model performance in underpopulated data silos, achieving up to +34.9 percentage points higher accuracy compared to local ResNet50 models. Thus, with ResNetFed, we provide a federated solution that can assist the initial COVID-19 screening in medical centers in a privacy-preserving manner.

Since the beginning of the COVID-19 pandemic, many different machine learning models have been developed to detect and verify COVID-19 pneumonia based on chest X-ray images. Although promising, binary models have only limited implications for medical treatment, whereas the prediction of disease severity suggests more suitable and specific treatment options. In this study, we publish severity scores for the 2358 COVID-19 positive images in the COVIDx8B dataset, creating one of the largest collections of publicly available COVID-19 severity data. Furthermore, we train and evaluate deep learning models on the newly created dataset to provide a first benchmark for the severity classification task. One of the main challenges of this dataset is the skewed class distribution, resulting in undesirable model performance for the most severe cases. We therefore propose and examine different augmentation strategies, specifically targeting majority and minority classes. Our augmentation strategies show significant improvements in precision and recall values for the rare and most severe cases. While the models might not yet fulfill medical requirements, they serve as an appropriate starting point for further research with the proposed dataset to optimize clinical resource allocation and treatment.

In medical imaging, deep learning models can be a critical tool to shorten time-to-diagnosis and support specialized medical staff in clinical decision making. The successful training of deep learning models usually requires large amounts of quality data, which are often not available in many medical imaging tasks. In this work we train a deep learning model on university hospital chest X-ray data, containing 1082 images. The data was reviewed, differentiated into 4 causes for pneumonia, and annotated by an expert radiologist. To successfully train a model on this small amount of complex image data, we propose a special knowledge distillation process, which we call Human Knowledge Distillation. This process enables deep learning models to utilize annotated regions in the images during the training process. This form of guidance by a human expert improves model convergence and performance. We evaluate the proposed process on our study data for multiple types of models, all of which show improved results. The best model of this study, called PneuKnowNet, shows an improvement of + 2.3% points in overall accuracy compared to a baseline model and also leads to more meaningful decision regions. Utilizing this implicit data quality-quantity trade-off can be a promising approach for many scarce data domains beyond medical imaging.

Data scientists in the Natural Language Processing (NLP) field confront the challenge of reconciling the necessity for data-centric analyses with the imperative to safeguard sensitive information, all while managing the substantial costs linked to the collection process of training data. In a Federated Learning (FL) system, these challenges can be alleviated by the training of a global model, eliminating the need to centralize sensitive data of clients. However, distributed NLP data is usually Non-Independent and Identically Distributed (Non-IID), which leads to poorer generalizability of the global model when trained with Federated Averaging (FedAvg). Recently proposed extensions to FedAvg promise to improve the global model performance on Non-IID data. Yet, such advanced FL algorithms trained on multilingual Non-IID texts have not been studied in industry and academia in detail. This paper compares, for the first time, the FL algorithms: FedAvg, FedAvgM, FedYogi, FedAdam and FedAdagrad for a binary text classification task using 12078 tailored real-world news reports in English, Portuguese, Spanish and Hindi. For this objective, pre-trained DistilBERT and BERT models fine-tuned with these texts are used. The paper results show that FedYogi is the most stable and robust FL algorithm when DistilBERT is used, achieving an average macro F1 score of 0.7789 for IID and 0.7755 for Non-IID protest news. The study also exhibits that BERT models trained with weighted FedAvg and FedAvgM can achieve a similar prediction power as centralized language models, demonstrating the potential of leveraging FL in the NLP domain without the need to collect data centrally.