Federated Learning (FL), Continual Learning (CL), and Digital Twins (DTs) have emerged as key paradigms for the development of intelligent, adaptive, and privacy-aware systems in various domains. FL enables collaborative model training across decentralized data sources without sharing raw data, thus ensuring privacy. CL allows models to continuously learn from evolving data streams and adapt to dynamic environments, reducing the need for retraining from scratch. DTs provide accurate virtual representations of physical systems, supporting real-time monitoring, simulation, and predictive maintenance. Combining these paradigms is a recent strategy for building physical systems that are decentralized, adaptive, and continuously improve using real-time data in various contexts. For example, in industries the integration of FCL with DTs can enable factories to learn from new sensor data across distributed sites while preserving sensitive data, adapting to equipment changes, and optimizing maintenance cycles. In mobile edge computing, this combination can enhance service reliability and user experience by updating models based on fresh data and dynamic user behavior. However, their combination also amplifies the inherent challenges, such as model drift, system complexity, and resource constraints, that need to be managed. Despite its promising potential, no existing surveys offer a focused and structured analysis of their intersection. This survey presents the first structured and comprehensive analysis of these three paradigms, highlighting not only existing approaches but also discussing their potential synergies and conflicts, outlining open research questions that must be addressed to unlock their full potential in real-world applications. • Identifies key challenges and outlines future research directions in FCL+DT. • Taxonomy and analysis of FCL methods applied to real-world DT systems. • First survey on the convergence of FCL and Digital Twin technologies.

Alzheimer’s disease (AD) is considered to be a significant health challenge that affects the cognitive ability of elderly people. The effects can only be slowed down if the disease is detected at an early stage. Researchers have extensively explored the use of machine learning algorithms to ensure early detection and prediction. However, effective models are complex, hence limiting their interpretability and privacy. Federated learning (FL) approaches have also been proposed to add privacy aspect to the machine learning models, however, FL methods are vulnerable to model related attacks. To address this we propose Agentic ElderFedLearn, a novel framework that proceeds in the following steps: 1) model healthcare institutions as autonomous artificial intelligence (AI) agents training local models on multimodal data [electronic health record (EHR) and synthetic magnetic resonance imaging (MRI)]; 2) apply personalized differential privacy (DP) to gradients, adapting budgets based on dataset size and sensitivity; 3) use multiagent reinforcement learning (MARL) to optimize agent interactions, such as privacy adjustments and communication; and 4) perform effective aggregation via weighted trimmed mean to defend against attacks. This innovation ensures privacy, handles heterogeneity, and achieves 94% accuracy with 0.93 F1-score, outperforming centralized approaches while using synthetic data.

Federated Learning (FL) facilitates decentralized model training while prioritizing data privacy. However, its effective implementation faces significant challenges, primarily related to high communication overhead and energy consumption. Quantization, a key optimization technique, plays a crucial role in enhancing energy efficiency by reducing the bit precision of model updates, thereby lowering computational and transmission costs. This paper explores the impact of quantization on energy-efficient FL, focusing on techniques such as lattice quantization and stochastic gradient quantization methods. By compressing gradient updates and model parameters, quantization significantly reduces bandwidth requirements, enables efficient model aggregation, and prolongs battery life in resource-constrained edge devices. Furthermore, we discuss the trade-offs between quantization levels, model accuracy, and energy savings, emphasizing strategies to mitigate performance degradation while maintaining robust learning. Experimental results demonstrate that quantized FL can achieve up to 50% reduction in energy consumption while maintaining competitive model accuracy. This study highlights the importance of quantization in scalable, sustainable, and energy-aware FL, paving the way for its widespread adoption in real-world applications such as smart healthcare, IoT, and edge AI.

Human Activity Recognition (HAR) represents a significant domain within pervasive computing, facilitating a diverse array of applications ranging from healthcare to smart environments. Traditional HAR models suffer from several challenges, including data privacy and the distributed participation of heterogeneous resource-constrained devices. To mitigate these challenges, the research community popularly uses federated learning (FL). However, selecting clients in FL is a critical issue, mainly when there is a combination of resource-constrained heterogeneous devices. This paper proposes a resource-and performance-aware client selection algorithm for HAR, namely EAPD-CS, that amalgamates the benefits of FL with energy efficiency. The framework allows for the training of machine learning models across multiple devices without the necessity of sharing raw data, thereby preserving user privacy. Additionally, it employs energy-aware strategies to diminish the carbon footprint and reduce the computational costs typically linked to traditional cloud-based HAR systems. Experimental results indicate that the proposed framework achieves more than 90% accuracy, comparable to centralized models, while significantly lowering energy consumption and improving the robustness of the model. This work contributes to the evolving field of green AI, delivering an effective, privacy-preserving, and environmentally sustainable approach for HAR applications.

Analyzing big data using federated learning (FL) requires distributing the data to different clients to process locally and sending the model parameters to the global server for aggregation. In a real scenario, big data distribution is nonindependent and identically distributed (non-IID). Aggregating client updates is an important task in federated learning. Federated averaging (FedAvg) is the simplest and most popularly used method in FL. However, it is unable to handle heterogeneous (non-IID) data. To address this, in this paper we propose a Client Specific Dynamic Aggregation (CSDA) strategy focusing on dynamic aggregation based on client-specific metrics, aiming to improve the robustness and performance of the global model. Each client’s contribution is weighted by the quality and performance of their local updates, enhancing the overall federated learning process. Extensive experimentation has demonstrated that the proposed CSDA strategy, in conjunction with advanced comparison techniques, has the potential to greatly enhance the accuracy of each client across three real-world datasets.

Extreme Learning Machine (ELM) is becoming a popular learning algorithm due to its diverse applications, including Human Activity Recognition (HAR). In ELM, the hidden node parameters are generated at random, and the output weights are computed analytically. However, even with a large number of hidden nodes, feature learning using ELM may not be efficient for natural signals due to its shallow architecture. Due to noisy signals of the smartphone sensors and high dimensional data, substantial feature engineering is required to obtain discriminant features and address the “curse-of-dimensionality”. In traditional ML approaches, dimensionality reduction and classification are two separate and independent tasks, increasing the system’s computational complexity. This research proposes a new ELM-based ensemble learning framework for human activity recognition to overcome this problem. The proposed architecture consists of two key parts: (1) Self-taught dimensionality reduction followed by classification. (2) they are bridged by “Subsampled Randomized Hadamard Transformation” (SRHT). Two different HAR datasets are used to establish the feasibility of the proposed framework. The experimental results clearly demonstrate the superiority of our method over the current state-of-the-art methods.

Numerous sensors on smart devices have made it possible to automatically recognize human movement, which might be helpful for intelligent applications like elder care, smart homes, and health monitoring. Nevertheless, implementing an activity recognition model in practical situations faces two main obstacles. First, machine learning models use a large number of labeled data to recognize human activities, which is not always feasible in real scenarios. Second, existing human activity recognition (HAR) systems cannot dynamically adapt to a new action. Furthermore, current methods fail to separate short-term activities from heterogeneous smart devices with varying positions and orientations that have similar sensory reading patterns. To address these issues, we propose Flexi-HAMR, an intelligent adaptive human activity monitoring and recognition system that dynamically recognizes activities using online, real-time activity signals. Many empirical findings show that the suggested flexible activity recognition model performs competitively on multiindividual activity identification tasks and has a comparatively more vital generalization ability.

The expeditious propagation of Internet of Things (IoT) technologies implanted in different smart devices such as smartphones and smartwatches have a ubiquitous consequence on the modern population. These devices are employed to collect data and to aid in tracking and analyzing the users’ daily activities using various human activity monitoring and recognition (HAR) techniques. However, most current HAMR approaches rely on exploratory case-based shallow feature learning architectures, which endeavor to recognize activities correctly in real-world situations. To address this issue, we offer a unique strategy for HAMR that leverages the attention mechanism with multi-head convolutional neural networks (CNNs) and Long-Short-Term-Memory (LSTM). The accuracy of activity detection is improved in the presented method by integrating attention into multi-head CNNs followed by LSTM for better feature extraction and selection. Verification investigations are carried out using data from the University of California (UCI) repository, which is publicly available. The results show that our proposed framework is more accurate than current frameworks using both the 10-fold and leave-one-subject-out cross-validation. Finally, the proposed method can recognize human activity in real-time, regardless of the type of smart device.

Transition-aware activity recognition is an inherent component of online health monitoring and ambient assisted living. An explosion of technology breakthroughs in wireless sensor networks, wearable computing, and mobile computing has facilitated this. However, real time, dynamic activity recognition is still challenging in practice. As reported in the existing literature, machine learning techniques are successfully used on the presegmented data to deliver transition-aware activity recognition systems. However, these strategies are frequently ineffective when used in a near-real-time context. This article presents an online change point detection (OCPD) strategy to segment the continuous multivariate time-series smartphone sensor data and its application in a transition-aware activity recognition framework. The proposed OCPD strategy is based on the hypothesis-and-verification principle. After the online data stream segmentation using the proposed OCPD strategy, feature engineering is performed to retain the essential features. Then, synthetic minority oversampling technique (SMOTE) is applied to balance the dataset. Finally, practical experiments are carried out to verify the suggested frameworks’ efficiency and reliability. The results reveal that the proposed OCPD strategy with ensemble classifier achieves a greater recognition rate (F-Measure: 99.80%) compared to methods stated in the literature.