The study [9] concludes that regardless of the number of devices, both vertical and horizontal federated learning achieve an accuracy rate of over 90% when the learning rate is set at 0.1. The learning rate is set to 0.1 for both scenarios. In the case of two terminal devices, the horizontal FL achieves a maximum performance of 98.7%. However, when the number of terminal devices increases to six, the horizontal FL still achieves the best performance at a learning rate of 0.1. The most suitable option can be determined. The learning rate for the KDD CUP 1999 dataset is approximately 0.1. By contrast, the accuracy rate is below 90% in all scenarios when the learning rate is set to 0.01. Furthermore, it may be inferred that the vertical FL has a lesser performance compared to the horizontal FL. The reason for this is that in vertical FL, each terminal device is presumed to possess distinct data categories. Studies have demonstrated that the accuracy of Federated Learning (FL) can decline by up to 50% when the data is highly non-identically and independently dispersed (Non-IID). However, in practical situations, a horizontal FL approach is more practical and the increased involvement of participants leads to a higher level of accuracy.
In [14] the evaluation results have indicated that the preference for utilizing a Hier- archical approach in FL is owing to its high accuracy, low overhead, and the generation of small block sizes for blockchains. Blockchain plays a crucial role in supporting the FL process inside the envisaged Industry 4.0 platform.The system facilitates decentral- ized storing of different blocks that represent FL models. This enables all participants to easily obtain updates to global and group models. Distributed FL minimizes the quantity of transactions and the additional burden on the network. FL incorporates multiple functionalities that enable the blockchain to effectively confront diverse obsta- cles. By exclusively storing the model changes, the amount of data contained in a block is reduced, thereby improving the scalability and efficiency of the blockchain. This optimization also enhances the storage and computational skills of the participants while safeguarding the privacy of the original data stored on their primary devices.
To proceed in [15] the fault detection algorithm, trained on SD data, demonstrates inferior performance in TD due to disparate data distributions resulting from varied operating circumstances of wind turbines. The suggested adaption technique has the potential to enhance the fault detection performance of the SVM-based model, par- ticularly when there is a limited amount of TD data.The efficacy and resilience of the adaptive technique is confirmed through the case studies, using the data gath- ered from the operational wind turbines. The findings demonstrate that the system is appropriate for the intended purpose.
The authors in [17] which have deployed a machine learning-based IDS to detect and classify attacks occurring within the fog layer, which is the device located nearest to the edge of the network. The AdaBoost classification model has achieved an accu- racy result of up to 99.85% on the given dataset KDD-Cup 1999, making it the most accurate algorithm. In order to improve the accuracy of future works, it is necessary to strengthen the Support Vector Machine and kernel approaches.
max width=
In the context of a variety of research initiatives, Table 1 provides a comparative review of a variety of approaches and the performance indicators that correlate to them. The following information is included in the table: main concept, methodolo- gies utilized, datasets utilized, performance outcomes, and authors who contributed content.
A decentralized and asynchronous FL framework is presented in the first entry of the table. This framework incorporates both vertical and horizontal FL paradigms. The methodology makes use of DP-GAN, which stands for Differentially Private Gen- erative Adversarial Networks, as well as the technology known as Blockchain. Using the KDD Cup dataset, the experiments were carried out, and the results demonstrated an accuracy of 95.3% for vertical FL and 98.7% for horizontal FL.
Continuing on, the second entry introduces a platform that utilizes DT, Blockchain, and Convolutional Neural Networks (CNN) to create a HFL platform that is helped by Blockchain technology. An accuracy of eighty percent was achieved using the dataset that was utilized, which was related to optical-wireless cyber-physical systems which are employed for manufacturing services.
The third column provides a description of an adaptive fault detection method that is based on machine learning and is specifically designed for wind turbine gearbox systems. The approach makes use of a support vector machine (SVM) classifier that has a Gaussian kernel. In order to conduct the evaluation, real-time data is utilized, which ultimately leads to accuracy rates of 86.30%, 90.95%, and 82.89% for three distinct case scenarios.
The fourth column presents IDS that is based on machine learning and is intended to identify attacks on the fog layer. On the KDD-Cup dataset, a number of differ- ent classifiers are utilized. These classifiers include Support Vector Machines (SVM), Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Trees (DT), Random Forests (RF), and AdaBoost. The performance measures demonstrate high levels of accuracy, with SVM achieving 97.52%, LR achieving 96%, KNN reaching 99.85%, DT achieving 99.7%, RF scoring 99.84%, and AdaBoost achieving 99.8%.
The application domains that are addressed by each technique are distinct. In con- trast to adaptive fault detection, which focuses on fault detection in specific machinery such as wind turbine gears, decentralized federated learning is primarily concerned with the handling of data in dispersed contexts through collaborative processing.This method proves its robustness in the management of distributed learning tasks by achieving an accuracy of 95.3% for vertical FL and 98.7% for horizontal FL on the dataset known as the KDD Cup.
IDS, on the other hand, are designed to protect the integrity of networks, particu- larly in scenarios that use fog computing settings.With SVM, LR, KNN, DT, RF, and AdaBoost achieving accuracies that were greater than 95%, the intrusion detection system (IDS) that utilized a number of different machine learning classifiers on the KDD-Cup dataset shown outstanding levels of accuracy. The achievement of accuracy rates more than 99% by KNN, RF, and AdaBoost is particularly noteworthy. These results demonstrate the efficacy of machine learning in spotting attacks on the fog layer.
The incorporation of DT, Blockchain, and CNN into an HFL platform demon- strated an accuracy of 80% when applied to the context of optical-wireless cyber- physical systems for industrial functions. Although it is not as high as the FL framework, this platform nonetheless offers a feasible option for certain applications that are used in the industry.The efficiency of federated learning in collaborative data processing is demonstrated by the fact that it achieves high accuracies in many configurations, including vertical and horizontal ones.
The analyisis sheds light on the variety of approaches that are utilized for data anal- ysis across a wide range of applications, as well as the success of these approaches. FL provides a complex method for the collaborative processing of data, whereas adaptive fault detection and IDS are designed to accommodate to certain application domains by employing approaches that are specifically adapted to those domains. It is essential to have a thorough understanding of the benefits and drawbacks associated with each methodology in order to select the approach that is best suitable for the application on the basis of the needs and limits situations that occur in real time, demonstrating that it is suitable for fault detection tasks. The robustness of IDS in identifying intru- sions in fog computing settings is highlighted by the fact that they constantly display high accuracies across all approaches.