نوع مقاله : مقاله پژوهشی
نویسندگان
1 دانشجوی دکتری، دانشکده مهندسی کامپیوتر، دانشگاه یزد، یزد، ایران
2 دانشیار، دانشکده مهندسی کامپیوتر، دانشگاه یزد، یزد، ایران
3 استادیار، دانشکده مهندسی کامپیوتر، دانشگاه یزد، یزد، ایران
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Federated learning, as a new approach in distributed machine learning, enables training of artificial intelligence models without the need to transfer raw data. However, the classical federated learning architecture depends on a central server, which is not available in many real-world applications such as mobile networks and the Internet of Things. On the other hand, data heterogeneity, node mobility and communication limitations create fundamental challenges in implementing such systems. In this study, a new method for decentralized federated learning is proposed. In this approach each node updates the model solely based on its local communications and without dependence on a central server. A dynamic graph based on node encounters is constructed to model the communication structure and effective nodes in the aggregation process are selected using the computation of the graph dominance set. To accurately weight the selected models, a multi-criteria decision-making method is used that considers criteria such as accuracy, data volume, and node score. In addition, the similarity between models is used the graph combination process. The proposed method is evaluated on three benchmark datasets MNIST, FASHION-MNIST, and CIFAR10 and is compared with two baseline methods, peer-to-peer FedAvg and WAFL. The accuracy results of the models in two environments with 50 and 100 nodes yielding values of 0.964, 0.952, 0.771, 0.682, 0.424 and 0.399 respectively, indicate that the proposed method achieves superior performance. These findings demonstrate the capability of the proposed approach for application in dynamic and mobile environments with a large number of clients and without centralized infrastructure, where it can overcome the scalability issues and client drift problem in decentralized federated learning.
کلیدواژهها [English]