JoWUA

Abstract

The fast proliferation of intelligent e-learning systems has caused the need for personalized learning services, along with security challenges related to data privacy, efficient data transfer, and communication problems associated with wireless mobile devices. The conventional methods of centralized machine learning are vulnerable to security attacks and are unable to handle the dynamics of participant involvement and heterogeneity in educational data. While Federated Learning can provide an effective decentralized solution to the problem by allowing the local training process without transmitting data, classical federated learning algorithms are also limited by problems of communication complexity, unstable convergence, and low personalization capability under non-IID learning. Thus, the paper provides an Adaptive Federated Learning Algorithm (AFLA) for privacy-aware personalized learning in e-learning systems with dynamically changing sets of users. This methodology involves the dynamic selection of the participants by considering their ability to communicate effectively, their computational power, and reliability, thus increasing the scalability and convergence of the algorithm. From experimental evaluations, it is evident that the proposed AFLA approach performs better than other approaches such as Centralized Deep Learning (CDL), Traditional Federated Learning (TFL), Differential Privacy Learning (DPL), and Personalized Federated Learning (PFL). For instance, this model achieved an accuracy of 96.2%, a precision of 95.8%, a recall of 95.1%, an efficiency of communication of 91.6%, and a privacy of 96.8%. Moreover, from the analysis performed to study the effects of the individual modules used, it was clear that both the client selection and personalization play significant roles in improving the overall performance of the proposed method.