Heart disease events can happen suddenly and may become dangerous during driving. In this paper, we propose an edge-cognitive framework for heart disease monitoring in the Cognitive Internet of Vehicles (CIoV) using ECG signals. The main goal is to detect cardiac risk patterns while keeping sensitive data inside the vehicle. The model uses a single ECG lead, as recorded in a vehicle by steering-wheel electrodes or a chest patch. Our method has three stages. First, we pretrain a 1D ECG encoder with self-supervised masked signal modeling on unlabeled ECG windows. Second, we use federated learning to train a global classifier in multiple vehicles without sharing raw ECG data. Third, we apply lightweight driver personalization to adapt the model to individual ECG patterns. To reduce unnecessary warnings, we add uncertainty-aware alerting based on MC-dropout and a dual-threshold policy. We evaluate the framework on the PTB-XL dataset using a five-class diagnosis task (NORM, MI, STTC, CD, HYP) and a binary task. With lead II, the federated model reaches a Macro-F1 of 0.478 and an AUC of 0.785 in the five-class task, almost the same as centralized training (0.469 and 0.789), and an accuracy of 0.803 in the binary task. Self-supervised pretraining makes federated training faster and gives a small gain when labels are scarce. The dual-threshold policy lowers the false alert rate on normal windows from 0.234 to 0.059, but more risky windows are missed. Overall, the proposed approach supports practical, privacy-preserving heart monitoring for CIoV.
Visual Internet of Things deployments operate under coupled bandwidth, energy and channel-reliability constraints, yet conventional video coding optimises pixel-level reconstruction rather than the preservation of task-relevant meaning. This paper presents QISFP, a quantum-inspired semantic fidelity protocol that treats semantic degradation as a stochastic state transformation rather than as additive signal distortion. It couples delta encoding of temporally associated object-level semantic states with a stochastic channel built from classical analogues of the depolarizing and amplitude-damping maps, and reconstructs meaning by energy- and confidence-weighted fusion over a temporal window. The framework is classical throughout: the quantum formalism supplies trace-preserving maps whose diagonal restriction is a stochastic matrix, which keeps per-object cost linear rather than quadratic in the semantic dimension. The induced class channel is shown to be M -ary symmetric, yielding two independent decay constants that characterise a semantic link and fix a multi-hop relaying budget, while a source-channel separation argument gives the attainability limit F ≤ 1 - p_d(1-1/M) , met with equality by uncoded single-observation transmission; every attainable gain therefore lies in the temporal dimension. Simulation on CDNet 2014 and UCF-101 indicates 27.5 10^4 operations per frame, so embedded feasibility is governed by the choice of semantic extractor rather than by QISFP itself.
Vehicular edge computing (VEC) supports latency-sensitive and computation-intensive workloads for intelligent transportation. Existing VEC optimization frameworks mostly minimize latency or energy in isolation and give limited attention to reliability under stochastic vehicular conditions. This paper proposes a reliability-aware energy minimization framework for a multi-tier vehicular edge–cloud continuum comprising vehicles, roadside units (RSUs), edge servers and a cloud-hosted federated aggregator. A joint stochastic optimization problem minimizes expected energy plus a conditional value-at-risk (CVaR) penalty on a failure cost, subject to per-task reliability targets, under stochastic mobility, time-varying wireless channels and node failures. The solution combines Lagrangian decomposition for the continuous resource allocation with a risk-sensitive federated reinforcement learning (FedRL) algorithm for the binary offloading decisions. In SUMO/OMNeT++/Veins simulations at 80 vehicles/km ^2 , the framework reduces per-task energy by 3.7–40.0 _0.95 of per-task energy is 21.3–46.1 ^2 ) the success rate is 94.7
AI models in critical sectors such as healthcare and finance must provide both data privacy and adversarial robustness. Differential Privacy (DP) protects training data by injecting noise, but this noise smooths decision boundaries and leaves models open to adversarial evasion, a tension known as the Privacy-Robustness Trade-off. Although this trade-off is well documented, its internal mechanism remains underexplored: prior work does not reveal how the noise reshapes a model’s reasoning or which features become vulnerable. To close this gap, we propose a Privacy-Aware Adversarial Defense grounded in Explainable AI. Specifically, we introduce the Attention Concentration Score (ACS), which measures how DP training shifts Transformer attention away from task-critical features toward non-functional ones. This attention drift correlates with adversarial vulnerability, providing a mechanistic explanation of the trade-off. Building on this insight, we develop a Manifold-Aligned Semantic Attack that targets the most drifted features, and a TrustScore defense that fuses embedding-level anomaly detection with attention-level consistency checks. We validate across two datasets (Adult Census, MIMIC-IV), two architectures (DeBERTa-V3-Large, LLaMA−3.1-8B), and seven experiments benchmarking five attacks against six defenses. Within the recommended range ( ϵ∈ [5, 10] ), models retain 84.2
Despite the recent increase in aquaculture production, information and communication technologies can significantly improve efficiency and help achieve better results. This systematic literature review studies previous research efforts on the application of intelligent environments to improve freshwater shrimp farming. The reviewed research works’ main contributions and opportunities for improvement are critically analysed to highlight the main benefits and existing gaps in using intelligent environment solutions for freshwater shrimp farming. The review highlights the importance of developing customised systems that address the specific characteristics of different freshwater shrimp farming processes, which are strongly influenced by the environmental context in which they occur, the species used for Pond grow-out, the users’ IT literacy, and other indicators that condition the technology to be used. This personalisation also affects the use of analytics techniques, which must adapt and perform with limited available data representing specific requirements of the process they support. The article proposes an architecture to guide the development of intelligent environments for freshwater shrimp farming, addressing the identified gaps and focusing on customising the behaviour of the system controlling the farming process.
