Federated Learning (FL) enables collaborative model training across distributed clients while preserving data locality and reducing the need to share raw data. However, deploying FL over non-terrestrial networks (NTNs), including satellite communication systems, introduces challenges such as propagation delay, intermittent connectivity, heterogeneous link reliability, and asynchronous update arrival. These conditions can increase update staleness and reduce the reliability and utility of client contributions, while privacy-preserving mechanisms may further affect model performance. This paper proposes TAADP-FedAsync, a Trust-Aware Adaptive Differential Privacy framework for asynchronous federated learning in NTN environments. The framework jointly coordinates (i) trust-aware client evaluation based on update quality and communication conditions, (ii) trust-conditioned adaptive differential privacy under Rényi Differential Privacy (RDP), and (iii) trust- and staleness-aware asynchronous aggregation for delayed updates. By coupling these mechanisms, the framework enables client reliability information to influence privacy configuration and asynchronous aggregation, thereby providing a unified approach to the interaction between trust, privacy, and communication staleness. The framework is evaluated in a simulated NTN environment incorporating satellite propagation delay, Doppler effects, packet loss, handover dynamics, heterogeneous client reliability, and asynchronous updates. The evaluation includes baseline comparisons, component-wise ablation, parameter sensitivity analysis, and experiments with varying client populations and network conditions. In the reported experiments, TAADP-FedAsync achieves 37.14
Diabetic Retinopathy (DR) is a major cause of vision impairment worldwide and automated analysis of retinal fundus images can support timely disease screening and severity assessment. This paper presents an end-to-end deep learning framework for five-class DR classification using the Indian Diabetic Retinopathy Image Database (IDRiD). The framework integrates image preprocessing, controlled data augmentation, an Approximate Class Balancer (ACB), transfer learning and Newtonian Particle Swarm Optimization (NPSO)-based hyperparameter tuning. Three pretrained architectures, namely VGG16, VGG19 and MobileNet are systematically evaluated under progressive experimental settings. NPSO incorporates Newtonian force dynamics, particle mass and elapsed-iteration-dependent attraction to improve search flexibility during learning-rate and dropout optimization. The best performance is obtained using NPSO-optimized MobileNet achieving an accuracy of 99.60 99.13 ± 0.18% , demonstrating consistent performance across stochastic initializations. Comparative hyperparameter optimization experiments show that NPSO provides the highest overall mean accuracy with comparatively low configuration-wise variability. The results demonstrate a lightweight NPSO-optimized transfer learning for computationally efficient multiclass DR screening.
To address the state lag and control-command misalignment caused by end-to-end communication and computation delays in edge-assisted automatic train operation (ATO), this paper centers on the co-optimization of delay-aware model predictive control (MPC) and edge computing resources and uses ATO as a representative validation case. First, a train dynamics model incorporating delays and an end-to-end delay model are established, and the trajectory correction problem is formulated as a constrained optimal control problem. Then, a delay-aware MPC (TA-MPC) algorithm is proposed. Furthermore, a communication-computation-control co-optimization framework is constructed, in which the co-scheduling of MPC parameters and edge computing resources is modeled as a mixed-integer nonlinear program and solved by ADMM. Experimental results show that, compared with benchmark methods, the velocity tracking RMSE is reduced by 21.2
The Large Language Models have very powerful features that completely changed the artificial intelligence scenario, but at the same time, their enormous computational demands are the source of serious environmental problems especially in terms of energy use and carbon release. The present day AI practices concentrate on speeding up the process without taking the environmental effect of the power sources into account. As a solution to this problem, this research proposes CarbonBalance framework. This framework introduces the integration of two components i.e. forecasting pipeline and a scheduling pipeline. The forecasting pipeline makes use of a machine learning pipeline where historical energy and weather data is processed. It adopts LightGBM for capturing complicated feature relations and Prophet for modeling temporal patterns in carbon intensity and provides accurate prediction of grid carbon levels for the scheduling pipeline, which uses an optimization scheduler and LLM profiles to make its decisions to direct non-urgent AI tasks to cleaner energy times and locations without affecting the critical tasks. Through this approach, the evaluated scheduling scenarios achieved a 41.6
Secure software updates are essential for maintaining Internet of Things (IoT) systems throughout their operational lifetime. However, software lifecycle information is typically fragmented across software repositories, deployment services, device management platforms, and runtime monitoring systems, limiting interoperability and end-to-end lifecycle visibility. This paper proposes a Digital Twin (DT)-based information model that uses the DT of each device as a shared information contract throughout the operational stages of the IoT secure software update lifecycle. The proposed model extends the DT with software lifecycle information, enabling deployment intentions, installation status, verification results, and runtime observations to be exchanged through a shared representation across heterogeneous software management components. The approach is realised as an extension of Eclipse Ditto and evaluated in a laboratory set-up for telemedicine gateways that combines a real gateway with fleets of emulated devices. A real security update could be followed revision by revision in the gateway’s DT; queries over the DTs identified vulnerable, updated and failed devices correctly, despite injected installation and verification failures and temporarily offline devices; and new lifecycle information and an additional lifecycle component were introduced without changing the existing ones. The coordination carried out through the DT remained small in comparison with the installation work on the device itself, while the throughput of the DT platform bounded how quickly an update could be assigned across a large fleet.
