Walaa Alayed

Future sixth-generation (6G) Internet of Everything (IoE) networks require zero-touch management systems capable of autonomously allocating communication and computational resources, satisfying service-level constraints, and providing transparent explanations for their decisions. Existing neuro-symbolic explainable artificial intelligence (XAI) twins improve interpretability by coupling neural prediction with symbolic reasoning; however, most rely on statically engineered causal graphs, fixed rule sets, and immutable knowledge bases. Such closed-world assumptions are unsuitable for non-stationary 6G environments in which radio conditions, mobility patterns, interference relationships, and service requirements continuously evolve. To address this limitation, this paper proposes LL-NS-XAI, a Lifelong Learning Neuro-Symbolic XAI Digital Twin for zero-touch 6G IoE management. The framework integrates three tightly coupled planes: an online implicit learner at the physical edge for time-varying service association and uplink/downlink resource allocation; a dynamic explicit learner in the virtual core for directed acyclic graph (DAG)-based Bayesian reasoning, most-probable-explanation inference, and explainable trust scoring; and a meta-cognitive layer that monitors reasoning residuals, detects parameter drift, structural drift, and ontological emergence, and autonomously revises the causal graph, probabilistic rule sets, conditional probability tables, and knowledge base. Unlike static neuro-symbolic twins, LL-NS-XAI treats realized post-decision outcomes, including achieved channel quality indicator (CQI), measured end-to-end delay, and throughput, as interventional evidence for online causal-model adaptation. Experimental evaluation using a non-stationary 5G-derived dataset with injected parameter drift, structural drift, and emergent IoE service classes demonstrates that LL-NS-XAI maintains higher long-term trustworthiness, lower reasoning residuals, and bounded performance degradation compared with static XAI twins, non-stationary bandits, and pure continual-learning baselines. The framework additionally accommodates a previously unseen IoE service class through autonomous causal-graph expansion, recovering near-oracle performance without offline retraining. These results demonstrate that lifelong causal adaptation can improve the robustness, explainability, and autonomy of AI-enabled zero-touch management in evolving 6G IoE networks.

Indoor millimetre-wave (mmWave) small-cell deployments, a key use case for sixth-generation (6G) wireless networks, increasingly rely on data-driven decision controllers for signal classification, beam adaptation, and spectrum allocation in low-mobility, pedestrian environments. These controllers are commonly trained on hybrid datasets that mix real world sensor measurements with large volumes of synthetically generated traces. However, the divergence in noise characteristics, temporal dynamics, and spectral profiles between real and simulated sensor data induces distributional drift that can cause model collapse, manifesting as degraded inference stability, entropy loss, and over the air (OTA) performance degradation. To address this challenge, we propose X-MCNet, an explainability guided training framework designed for intelligent sensing systems operating under synthetic data contamination. The approach continuously monitors Shapley value attributions of critical physical layer features such as SNR, Doppler spread, delay spread, and angle of arrival. It quantifies attribution divergence using a Kullback–Leibler based Attribution Drift Index (ADI), and applies a convex dual reweighting mechanism to dynamically restore training stability by reintegrating high entropy OTA sensor instances. Experiments on a real time Software Defined Radio (SDR) sensor testbed demonstrate that X-MCNet achieves up to 19.7% lower exclusion AUC, 5.3% higher inclusion AUC, and a 29% reduction in bit error rate at 25 dB SNR, all with under 8% additional compute overhead. The framework’s empirically calibrated residual risk bound and edge-ready design confirm its viability for deployment in real time, resource constrained sensor platforms. X-MCNet thus offers a robust, interpretable, and spectrum-aware solution for indoor mmWave sensor-driven wireless systems operating under heavy synthetic-data reliance.

Connected and autonomous vehicles increasingly employ federated learning (FL) for privacy-preserving collaborative intrusion detection. However, existing FL systems with explainable AI (XAI) assume honest participation in explanation aggregation, creating a critical vulnerability. We show that Byzantine vehicles can poison global interpretability maps by submitting falsified SHAP feature importance scores, suppressing security-critical indicators (e.g., packet length, CAN ID frequency) to evade detection. To address this threat, we propose RobustFedX, the first Byzantine-robust federated explainability framework securing both model parameters and explanation integrity via three complementary defenses: (1) geometric median-based robust aggregation with outlier detection for explanation vectors; (2) adversarial explanation training to ensure SHAP stability under perturbations; and (3) cryptographic hash commitments binding explanations to model parameters. Experiments on CICEVSE2024 and CICIoV2024 show that RobustFedX sustains 93.4% detection accuracy and 87.3% explanation fidelity under 30% Byzantine hybrid attacks, compared to 74.0% accuracy and 31.4% fidelity for prior XAI-FL-IDS methods. Ablation analysis confirms progressive robustness gains, achieving 95.9% attack resistance. With moderate overhead (50% training time, 24% communication), RobustFedX enables secure and transparent intrusion detection for safety-critical vehicular systems.

Federated learning is a distributed machine learning paradigm that enables multiple devices to collaboratively train a shared model while keeping their raw data localized. Federated learning has become an attractive solution for intrusion detection in Internet of Things (IoT)-based wireless sensor networks because it enables collaborative model training without transferring raw traffic data. However, real deployments rarely satisfy the common assumption that client data are independent and identically distributed (IID). In practical wireless sensor networks, data heterogeneity naturally arises from spatial variation, uneven attack exposure, traffic imbalance, and differences in sensing conditions, which can substantially affect detection reliability and deployment feasibility. This study presents an application-oriented evaluation of federated intrusion detection under controlled non-IID conditions using three representative datasets: WSN-DS, CIC-IDS-2017, and UNSW-NB15. An LSTM-based intrusion detection model is trained in a federated setting and assessed using three aggregation strategies, namely, FedAvg, FedProx, and SCAFFOLD, under label skew, quantity skew, and feature skew scenarios. The results show that standard FedAvg degrades markedly as heterogeneity increases, with accuracy reductions of up to 23.4 percentage points and substantially higher communication cost under extreme non-IID settings. In contrast, FedProx and SCAFFOLD improve convergence stability and reduce the impact of client drift, with SCAFFOLD showing the strongest overall robustness and up to 45% lower communication cost than FedAvg due to faster convergence. These results demonstrate that non-IID awareness is essential for building deployable privacy-preserving intrusion detection systems for resource-constrained IoT environments. The study provides practical guidance for selecting federated aggregation strategies in wireless sensor network security applications where robustness, bandwidth efficiency, and real-world data heterogeneity must be jointly considered.

In recent years, object detection has become an indispensable component in various critical applications, ranging from autonomous driving to surveillance systems. However, achieving high detection accuracy alone is insufficient; models must also be transparent and interpretable to foster trust and ensure reliable decision-making, especially in safety-critical domains. Traditional supervised approaches require large amounts of labeled data, which is often expensive and time-consuming to obtain. Additionally, these models typically function as “black boxes,” providing limited insight into their decision processes. This article introduces SemiXAI-Net, a semi-supervised learning framework optimized for object detection tasks while incorporating explainable AI (XAI) techniques to enhance transparency. By utilizing a teacher-student architecture, the framework iteratively refines pseudo-labels, effectively improving both model performance and interpretability. The experimental evaluation was performed using the Common Objects in Context (COCO) dataset (a real-world benchmark) alongside synthetic datasets to simulate semi-supervised conditions. A labeled subset of COCO images was combined with unlabeled data (both from COCO’s unannotated images and synthetic sources) for training. The framework was implemented on NVIDIA Tesla V100 GPUs using PyTorch, with hyperparameters optimized for stability (batch size: 32, learning rate: 0.001) and efficiency. To enable real-time interpretability, the student model integrated Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-Agnostic Explanations (LIME) techniques, generating visual explanations during inference. Training spanned 10 epochs with early stopping to mitigate overfitting risks. SemiXAI-Net achieves a superior accuracy of 93%, precision of 94%, recall of 95%, and an F1-score of 94.5%. In comparison, traditional methods reach a maximum accuracy of 87%. The explainability score of SemiXAI-Net stands at 94%, significantly surpassing the traditional method’s score of 71%. Additionally, the framework demonstrates a generalization ability of 94% on unseen data, making it highly adaptable to new environments. These results highlight SemiXAI-Net’s potential for real-world applications, particularly in critical domains like autonomous driving, where both high performance and model transparency are essential.

The increasing trend of phishing sites is among the important threats against the Internet security, associated with monetary loss, data leakage, and identity swindle. In response to this urgent problem, this paper proposes a new phishing website detection framework based on the Modified Differential Evolution (mDE) algorithm in conjunction with state-of-the-art machine learning classifiers. The proposed mDE integrates with dynamic mutation and crossover strategies to improve the global search capability and the convergence speed, which is superior to traditional single optimization methods. We conduct experiments on two benchmark datasets: the UCI Phishing Websites dataset and the large-scale Mendeley Phishing URLs dataset. The proposed optimizer is implemented as an improved variant, mDE+, which augments the adaptive control parameters with opposition-based initialization, guided mutation, and self-adaptive control of the mutation and crossover factors. Using Random Forest and gradient-boosting classifiers, mDE+ selects compact feature subsets that are competitive with, and on most settings superior to, those produced by standard DE and other nature-inspired optimizers (GWO and COA), with statistically significant gains over DE and COA on UCI and negligible additional computational overhead. Experimental results demonstrate that the proposed approach outperforms the state-of-art methods in terms of accuracy, precision, recall and F1-score, which further verify the effectiveness and robustness of the proposed method. This paper shows that mDE is a promising optimization method for training phishing detection systems for real-world cyber security applications.