This paper addresses the challenges in resource allocation within disaggregated Radio Access Networks (RAN), particularly when dealing with Ultra-Reliable Low-Latency Communications (uRLLC), enhanced Mobile Broadband (eMBB), and Massive Machine-Type Communications (mMTC). Traditional traffic steering methods often overlook individual user demands and dynamic network conditions, while multi-connectivity further complicates resource management. To improve traffic steering, we introduce Tri-GNN-Sketch, a novel graph-based deep learning approach employing Tri-subgraph sampling to enhance link prediction in Open RAN (O-RAN) environments. Link prediction refers to accurately forecasting optimal connections between users and network resources using current and historical measurements. Tri-GNN-Sketch is trained on real-world 4G/5G RAN monitoring data. The model demonstrates robust performance across multiple metrics, including precision, recall, F1 score, and ROC-AUC, effectively modeling interfering nodes for accurate traffic steering. We further propose Tri-HyperGNN-Sketch, which extends the approach to hypergraph modeling, capturing higher-order multi-node relationships. Using link-level simulations based on Channel Quality Indicator (CQI)-to-modulation mappings and LTE transport block size specifications, we evaluate throughput and packet delay for Tri-HyperGNN-Sketch. Tri-HyperGNN-Sketch achieves an exceptional link prediction accuracy of 99.99% and improved network-level performance, including higher effective throughput and lower packet delay compared to Tri-GNN-Sketch (95.1%) and other hypergraph-based models such as HyperSAGE (91.6%) and HyperGCN (92.31%) for traffic steering in complex O-RAN deployments.

Graph-based Intrusion Detection Systems (IDS) based on Graph Neural Networks (GNNs) such as GCNs and GATs, have demonstrated significant effectiveness in identifying malicious network traffic. However, these models are vulnerable to adversarial attacks, and their lack of transparency complicates the graph-based model classification. To address these challenges, this work proposes a unified framework that combines a novel Residual Generative Adversarial Attack (ResGAN-FGSM) with a counter-factual explanation module for high-order perturbations, alongside Fast Gradient Sign Method (FGSM)-based refinement. This combination generates semantically coherent and highly disruptive adversarial examples, exposing vulnerabilities in GCN and GAT-based IDS systems that prior methods fail to uncover under full white-box access. Additionally, we develop Counterfactual GraphLIME, a novel explanation framework for graph-structured data. This framework iteratively perturbs a node's 1-hop neighborhood until its label flips, identifying the minimal feature changes required to alter predictions, thereby improving transparency and supporting security diagnostics. Through comprehensive experiments on CICIDS2017, we demonstrate that ResGAN-FGSM reduces GNN-based IDS accuracy by up to 30% under modest perturbation budgets$\epsilon$, outperforming FGSM alone. The counterfactual explanations reveal sensitive features, such as inter-arrival time and burst rate, providing valuable insights for model hardening.

The Federated Learning (FL) approach in the Internet of Health Things (IoHT) is vulnerable to membership inference attacks (MIAs) and adversarial attacks such as model poisoning, both of which threaten privacy and robustness. Model poisoning attacks compromise the global model by aggregating malicious client updates, while MIAs exploit information leakage during FL communication. Current techniques for FL robustness, such as adversarial training and differential privacy (DP) address these threats differently. DP provides the benefit of noise injection into gradients, ensuring privacy for data shared on a central aggregation server while maintaining a privacy-utility trade-off. However, it is not robust against model poisoning attacks on clients in a collaborative FL deployment within the IoHs. To address the challenges with the robustness of the DP, this work proposes a differentially private federated adversarial training (DP-FAT), integrating adversarial training with DP (DP-PGD and DP-FGSM). Using the ClaMP_Integrated-5184 dataset, proposed DP-FAT approach ensures a robust malware detection model while preserving privacy. Experimental results demonstrate that DP-FAT mitigates adversarial attacks and protects sensitive data during federated rounds, offering a scalable solution for secure IoHTs. Furthermore, privacy-utility trade-offs for DP- PGD reveal optimal Rényi Differential Privacy (RDP) guarantees, with configurations such as$(\alpha=3, \epsilon=6)$or$(\alpha=2,\ \epsilon\in[5,7])$effectively balancing robustness and privacy in IoHT malware detection systems.

Brain tumours are abnormal and uncontrollable multiplication of cells in the brain. They can have significant implications for patients, ranging from headaches and nausea to more severe symptoms such as seizures and memory difficulties. An accurate diagnosis is vital to implement a successful treatment plan. Current deep learning approaches for classifying brain tumours rely on the availability of large amounts of labelled data. This can be challenging, as highly trained medical professionals need to manually label each image. To address these limitations, this work proposes a few-shot learning method that uses a varying number of shots, ranging from 1-shot to 10-shots. Three categories of brain tumours are classified: glioma, meningioma, and pituitary tumours. The proposed approach demonstrates 96.30 % and 96.96 % accuracies for the 1-shot and 10-shot models, respectively. The state-of-the-art CNN method achieved a 97.07% accuracy. However, this approach required significantly larger quantities of labelled data. This development could improve the classification of brain tumours, particularly for rare diseases or instances where there is a scarcity of labelled data.

Agronomic forecasting, especially crop yield prediction, faces significant challenges due to data sparsity, irregular sampling, and the limited availability of labeled time-series data. To address these issues, this study proposes a Spatiotemporal Generative Adversarial Network (ST-GAN) framework designed to generate temporally consistent synthetic data for time-series forecasting tasks. We evaluate the performance of an LSTM regression model trained on both real-world and ST-GAN-generated datasets, focusing on predicting crop yield from synthetic features like temperature, rainfall, and pesticide usage. Experimental results show that training the LSTM model with ST-GANgenerated data yields substantial improvements in predictive accuracy, with a 29.3 % reduction in mean squared error (MSE), highlighting the model's ability to capture complex temporal dependencies effectively. Importantly, the synthetic sequences produced by ST-GAN maintain critical spatiotemporal coherence, allowing the model to generalize better to unseen real-world data, resulting in lower error rates on validation samples. The incorporation of semi-supervised learning within the ST-GAN framework further enhances its utility, enabling improved forecasting in low-resource environments where labeled data is scarce. By leveraging the strength of both synthetic data generation and semi-supervised learning, our approach overcomes the limitations of traditional methods, offering a robust solution for agronomic forecasting.

Modeling molecular interactions such as HLA-peptide binding using graph neural networks (GNNs) is challenging due to incomplete and noisy domain knowledge that often leads to suboptimal heuristically constructed graph structures. To address this limitation, this research proposes an end-to-end graph learning framework, named GraphImmuno-IDEAL, which jointly refines graph topology and learns node embeddings. This approach leverages pairwise similarity learning and dynamic structure refinement to correct erroneous edges and discover meaningful interactions. Another new modelling technique, GraphImmuno-IDEALS introduces an anchor-based scalable graph learning mechanism to significantly reduce computational overhead while preserving performance. Experimental evaluation on multiple immunological datasets demonstrates that both latter models consistently outperform baseline classifiers and static GNNs in terms of accuracy and robustness. Furthermore, the learned task-specific graph structures enable interpretability through community detection and key interaction discovery, offering valuable insights for downstream applications in immunology and drug development.

Detecting front-running attacks in Ethereum blockchain transactions is crucial for maintaining security and integrity within decentralized ecosystems. However, existing models struggle to accurately model the complex distributions inherent in tabular data, particularly in the presence of class imbalance and mode collapse. This paper leverages the potentials of Conditional Tabular Generative Adversarial Networks and PacGAN, called a Conditional Packing GAN (cPacGAN), to address these challenges. cPacGAN effectively generates synthetic data that closely mimics the distribution of real transactions, thereby augmenting the dataset and improving the performance of front-running attack detection. PacGAN mitigates mode collapse by incorporating packed samples in the discriminator, improving the diversity of generated samples and improving the stability of the training process. Through experimental evaluations of a real-world Ethereum transactions dataset, cPacGAN demonstrates improved performance across all selected machine learning classifiers, particularly augmenting the effectiveness of Tabular Neural Networks (TabNet).

Calibration in wireless sensor networks (WSNs) poses a significant challenge, particularly in uncontrolled environmental deployments for environmental monitoring, such as air pollution. Traditional calibration methods rely on centralized reference stations, which are costly to maintain, offer limited coverage, and calculate measurements as averages. However, with the rise of the Internet of Things (IoT), sensors present a cost-effective alternative for calibration compared to fixed reference stations. Nevertheless, in uncontrolled environments, sensors require self-recalibration to ensure accurate measurements for the reliable operation of WSNs without human intervention. Existing calibration approaches, such as LSTM, are computationally expensive, have higher memory requirements, and exhibit training instability, making them unsuitable for resource-constrained WSNs. This paper proposes a self-calibration approach for WSNs using the Gated Recurrent Unit (GRU) coupled with the attention mechanism (Attention-GRU). The Attention-GRU selectively focuses on relevant features while capturing long-term dependencies, akin to Recurrent Neural Networks (RNNs), thereby mitigating overfitting. Experimental results demonstrate that the Attention-GRU model outperforms other models with an R-squared value of 0.97 and accelerated learning. These accurate sensor recalibration predictions promote sustainability by supporting IoT-enabled air pollution monitoring efforts.

Ubiquitous computing has potentials to harness the flexibility of distributed computing systems including cloud, edge, and Internet of Things devices. Mobile edge computing (MEC) benefits time-critical applications by providing low latency connections. However, most of the resource-constrained edge devices are not computationally feasible to host deep learning (DL) solutions. Further, these edge devices if deployed under denser deployments result in topological dependencies which if not taken into consideration adversely affect the MEC performance. To bring more intelligence to the edge under topological dependencies, compared to optimization heuristics, this article proposes a novel collaborative distributed DL approach. The proposed approach exploits topological dependencies of the edge using a resource-optimized graph neural network (GNN) version with an accelerated inference. By exploiting edge collaborative learning using stochastic gradient (SGD), the proposed approach called CGNN-edge ensures fast convergence and high accuracy. Collaborative learning of the deployed CGNN-edge incurs extra communication overhead and latency. To cope, this article proposes compressed collaborative learning based on momentum correction called cCGNN-edge with better scalability while preserving accuracy. Performance evaluation under IEEE 802.11ax-high-density wireless local area networks deployment demonstrates that both the schemes outperform cloud-based GNN inference in response time, satisfaction of latency requirements, and communication overhead.

Wireless power transfer techniques to transfer energy have been widely adopted by wireless rechargeable sensor networks (WRSNs). These techniques are aimed at increasing network lifetime by transferring power to end devices. Under these wireless techniques, the incurred charging latency to replenish the sensor nodes is considered as one of the major issues in wireless sensor networks (WSNs). Existing recharging schemes rely on rigid recharging schedules to recharge a WSN deployment using a single global charger. Although these schemes charge devices, they are not on-demand and incur higher charging latency affecting the lifetime of a WSN. This paper proposes a collaborative recharging technique to offload recharging workload to local chargers. Experiment results reveal that the proposed scheme maximizes average network lifetime and has better average charging throughput and charging latency compared to a global charger-based recharging.

Carrier sense multiple access mechanism with collision avoidance (CSMA/CA) in IEEE 802.15.4‐based wireless body area networks (WBANs) may impair the transmission reliability of emergency traffic under high traffic loads, which may result in loss of high valued medical information. Majority of the recent proposals recommend an early retransmission of failed frame while ignoring the history of past failed transmissions. More importantly, these proposals do not consider the number of failed transmissions experienced by each sensor node, thereby affecting the reliability of retransmissions. In this paper, we propose a dynamic retransmission adaptive intelligent MAC (RAI‐MAC) scheme. In our proposed scheme retransmission class of each sensor node is decided by the coordinator according to the number of failed transmissions of each node as observed by the coordinator during the last superframe. Based on the retransmission class received from the coordinator, each node adjusts its next backoff value. The proposed scheme increases the probability of successful frame retransmissions without incurring extra overhead. The simulation results prove that the proposed scheme based on its adaptive retransmission mechanism achieves higher average throughput and average end‐to‐end delay, while not compromising on energy efficiency as compared to the IEEE 802.15.4 and Block Acknowledgment (Block Ack). Moreover, our scheme appears more stable in terms of average throughput, end‐to‐end delay, and energy efficiency under different values of beacon order (BO) and superframe order (SO).

Broadcasting is a fundamental operation underlying different routing, multicasting and address resolution protocols. Broadcasting in a network requires that all the nodes in the network receive the broadcast packet. Mobility in the network induces link failures which cause some nodes to lose the broadcast packets. Objective of all broadcasting protocols is to achieve high reachability while keeping the broadcast redundancy as low as possible. In this paper we propose a cross layer protocol, called cross layer mobility adaptive broadcasting (CMAB) to handle the mobility in mobile ad hoc networks (MANETs). CMAB uses two Disjoint Sets of Broadcast Relay Gateways (BRGs1 and BRG2) to ensure high reliability in case of high mobility. Our approach minimizes broadcast redundancy by activating the second set of BRG2s only in highly mobile scenarios. A further reduction in broadcast redundancy is achieved by forcing the second Disjoint BRG2 to rebroadcast only if it is covering a maximum number of 2-hop neighbors of upstream sender or source to be covered by BRG1. The proposed protocol balances the retransmission redundancy avoiding the broadcast storm problem and increasing reachability in highly mobile and denser network scenarios. Simulation results show that CMAB provides high delivery ratio, low forwarding ratio and low end-to-end delay in highly mobile and denser network scenarios.