Search Results (1,101)

Reliable quality monitoring at the front end of sustainable cold chains requires direct observations of product condition in addition to environmental measurements. This study developed an energy-efficient edge vision system for six coffee fruit ripeness and visible-condition classes on an RK3576 neural processing unit (NPU). YOLO11 was reconfigured by combining a Branch-Integrated Feature Aggregation module for fine-grained representation, a YOLOv10-derived Cross-Stage Partial Network with Compact Inverted Block for efficient multi-scale fusion, and a Simplified Regression Head with reduced regression channels. On a public benchmark derived from 996 source images, YOLO11 and YOLO11BCS yielded mean average precision values at an intersection-over-union threshold of 0.5 (mAP50) of 81.5% and 83.2%, respectively, under the matched 320-pixel protocol. At a 640-pixel input, the computational complexity decreased from 6.3 to 5.5 giga floating-point operations. The largest class-level mAP50 difference was 9.7 percentage points for overripe fruit. Under controlled synthetic degradations, the model produced higher point estimates than YOLO11 for illumination variation and rain, whereas severe Gaussian noise and motion blur remained limiting. Board-input measurements showed throughput increases of 24.2% and 22.1% at 640- and 320-pixel inputs, respectively. The 320-pixel system reached 50.3 frames per second and 10.06 frames per second per watt. Within the supplied benchmark, the matched evaluation supports a comparative assessment of the models, while source-image correspondences across partitions define the scope of the reported absolute metrics. Within these boundaries, the results support low-power visual quality monitoring near the point of harvest. Full article

Edge protection of heterogeneous Internet of Things (IoT) devices needs per-device behavior models, cross-site learning, and blocking that spares legitimate services. This systems study presents the Personalized Federated Autoencoder Framework (PFAF), deployed on a Raspberry Pi access point. PFAF passively aggregates traffic into 60 s windows and keeps a scaler, autoencoder, and threshold per MAC address, sharing only model weights and summary statistics for FedAvg, so raw traffic never leaves the access point. It flags deviations from each device’s learned traffic rather than classifying attacks, blocking only after whitelisting and consecutive-window confirmation. Across 652,413 windows from 10 devices (84.3% from one, 8 modeled), global initialization lowered first-epoch loss 2.2- to 40-fold, and per-device models reconstructed traffic 3.5 to 308 times more accurately than the global model. A low-rate TCP SYN flood was detected about 9 s and blocked about 69 s after its first packet, through RST replies absent from training. A dry-run replay traced most false blocking to persistent legitimate behavior changes, and a whitelist learned over a period including such a change cut would-block decisions from 88.0 to about two per day with traffic. PFAF achieves edge detection and blocking, with enforcement as a separate, dry-run-calibrated stage. Full article

Learned resource allocators for sixth-generation edge networks usually treat operational limits as penalties and may return infeasible decisions. This paper studies energy minimization in multi-cell uplink non-orthogonal multiple access mobile edge computing networks. Treating inter-cell interference as a receiver-side quantity makes the two-user cluster sum rate depend only on the aggregate received power of the pair. Joint optimization of user association, subchannel assignment, transmit power and computation allocation is formulated as a mixed-integer nonlinear program with mandatory service and solved by a safeguarded alternating scheme: a capacity-aware heuristic proposes an association, a conditional diffusion model generates powers through a bounded latent mapping that enforces the power limits by construction, and the computation allocation follows in closed form. Given a feasible initialization, accepting only complete candidate tuples that pass the feasibility and energy tests preserves feasibility and yields a non-increasing objective without requiring block-wise optimality. In simulations, total energy is about 10% lower than with a deep reinforcement-learning baseline and 16% lower than with successive convex approximation, mainly through lower transmission energy (0.98 J vs. 1.32 J). At the default configuration, fewer than 2% of generated candidates fail the terminal test, and one outer iteration takes 3.52 s at the largest tested network, compared with 72.5 s for successive convex approximation. Full article

Federated learning, unlike centralized training, enables multiple clients to train a global model without exposing raw data, making it useful for applications such as medical diagnosis in smart healthcare and mobile edge networks. This setting is vulnerable to dynamic collusive attacks, where colluding clients coordinate abnormal parameter updates and adjust their strategies across training rounds. Existing studies address this problem with isolated mechanisms that focus on a single defense stage, but such mechanisms are insufficient for attacks that evolve across rounds. To address this challenge, we propose reputation-aware Stackelberg incentive aggregation (RSIA) to mitigate the influence of dynamic collusive attacks while balancing model utility in federated learning. Specifically, we first use incentive rules before local training to make repeated collusive behavior less beneficial. We then update client reputation after each round, so earlier unreliable behavior affects later incentives. Finally, we use reputation and update reliability to weight client updates, reducing the influence of abnormal updates on global model training. Experiments on three datasets show that RSIA reduces the average effective attack success rate by 84.9% compared with representative state-of-the-art (SOTA) methods. RSIA also maintains competitive test accuracy, providing a favorable trade-off between model utility and attack suppression. Full article

Mirror segmentation is challenging because reflected contours can be difficult to distinguish from physical mirror boundaries. Depth provides complementary geometric cues, but active sensing can produce invalid or distorted measurements on reflective surfaces. Diffusion-based feature aggregation with a heat conduction operator (HCO) provides global context for mirror localization, while boundary-relevant high-frequency structure is attenuated and remains in the complementary residual. We propose the Diffusion-Band Mirror Feature Enhancement Network (DiBME-Net), which retains the diffusion representation for global mirror localization while recovering and spatially modulating the complementary residual under learned mirror-boundary guidance. Its core module, Mirror-Boundary-Guided Diffusion-Band Feature Enhancement (MB-DBFE), consists of two components. First, HCO-Derived Diffusion Feature Decomposition (DFD) partitions the residual into diffusion-derived bands using differences between heat kernels defined by the same learned diffusion strength used for context aggregation, requiring no manually specified frequency cutoffs. Second, Boundary-Guided Residual Recovery (BRR) maintains a full-residual recovery path and spatially modulates the selected band responses using stage-aligned, band-specific boundary guidance. Because raw depth discontinuities also occur at ordinary object edges and can result from sensor artifacts, a Mirror Boundary Estimation (MBE) head predicts guidance from fused RGB-D features and is supervised using boundaries derived from the ground-truth mask. Controlled experiments support these design choices in terms of band derivation, band composition, boundary guidance, and stage placement. On RGBD-Mirror, DiBME-Net achieves an IoU of 84.82, an MAE of 0.025, and a BER of 6.09, all averaged over three random seeds, yielding the best reported IoU, MAE, and BER among the compared methods. It also transfers to multiclass transparent/reflective object segmentation on TROSD while retaining the proposed decomposition and boundary-guided recovery design. Full article

Retinal vessel segmentation is essential for diagnosing eye diseases. However, complex multiscale vascular structures and low contrast make accurate segmentation difficult. Current methods often overlook structural priors and seldom explore spatial–frequency synergies, resulting in poor segmentation of tiny vessels. To address these issues, we propose a Prior-Guided Spatial–Frequency Interactive Network (PGSFI-Net) that operates in both spatial and frequency domains. The Sobel-Attention Prior module extracts vascular edge priors using a single-scale Sobel operator and spatial attention, followed by multi-scale prior injection to strengthen faint-vessel representations. The Four-Branch Selective Attention Aggregation module captures multi-directional vascular features through strip convolutions and multi-dimensional attention. The Multi-Branch Spatial–Frequency Fusion module achieves bidirectional feature complementation via cross-gating interaction. The Local–Global Fusion module recovers subtle vascular structures by combining differential CNN branches with channel self-attention. On DRIVE, CHASE_DB1, and STARE, PGSFI-Net achieves ACC/F1/AUC values of 0.9646/0.8136/0.9838, 0.9727/0.8290/0.9838, and 0.9746/0.8393/0.9884, respectively. These results demonstrate competitive performance against state-of-the-art methods across the evaluated datasets and challenging low-contrast cases. Full article

Smart-city platforms connect IoT devices across transport, energy, water, healthcare and e-government, so every gateway is an entry point. This paper evaluates a four-layer architecture—edge intrusion detection trained across districts by federated learning, digests anchored on a permissioned ledger, and a cloud orchestrator that correlates and audits—whose contribution, LPRA, is a screening rule with ledger-recorded accept and quarantine decisions. On 211,043 identifier-free IoT flows, a depth-8 tree reached F1 = 0.997 at 17.7 KB and an estimated 0.19 J per million flows. Federated learning across five districts reached 0.971 ± 0.001, against 0.952 ± 0.002 alone. Under scaling, label-flip, Gaussian-noise and stealth poisoning attacks, an adaptive attacker that knows the screen, and a sign-flip attack not used in its design, LPRA matched the best robust baselines under attack, avoided their no-attack heterogeneity tax at five districts and produced a ledger-writable decision; it did not detect the stealth attack, its screen holds to a Byzantine fraction of about 0.36, degrades at 0.44–0.45 and inverts above one half, and the ledger’s guarantee ends at a validator-quorum collusion. Anchoring ran at 146 tx/s on a two-organisation Hyperledger Fabric test network. F1 falls from 0.96 within a dataset to 0.12 across, and from 0.9824 to 0.0005 on a later day; the federated model transfers worse than the edge tree. Training must stay local and continuous; federation delivers it. Full article

Textual-edge graphs represent interactions through both connectivity and relation-specific text. Learning community structure from these graphs requires an encoder that preserves the relevance of individual relations when combining semantic and structural information. We propose Dual-Stage Attention Embedding for Community Detection on Textual-Edge Graphs (DAETG). The first attention stage summarizes incident-edge text, and the second transfers its coefficients into neighboring-node aggregation as a learnable prior. Following topology-reconstruction pretraining, Kullback–Leibler self-training and differentiable modularity jointly refine the encoder and a centroid-based community assignment. Training uses fixed node and edge text embeddings and a supplied community count, without category-membership supervision. Experiments on four benchmark graphs with base and large Bidirectional Encoder Representations from Transformers (BERT) embeddings show competitive community-detection performance, with modest mean improvements that vary across datasets and evaluation metrics. Ablation studies on two graphs support the contributions of correctly associated edge text, the dual-stage encoder, and joint community refinement. Together, these results support using relation-specific text alongside topology to learn community representations. Full article

Accurate recognition of individual appliances from their own power consumption profiles underpins non-intrusive load monitoring (NILM), yet deep learning models often incur high computational overhead and lack operational interpretability. This study proposes a new explainable and lightweight model named β-PruNet to recognize household appliances using daily power consumption signals. The task addressed here is appliance classification from submetered profiles rather than disaggregation of an aggregate household signal, so the model is positioned as the recognition stage of a NILM pipeline. Each 24 h signal (86,400 samples) is transformed into a 720-bin ordinal pattern (OP) distribution, and the distribution is summarized by 27 statistical and spectral descriptors over 48 sliding windows. A compact class-weighted convolutional neural network extracts 128 embedding features. Drawing inspiration from collective swarm intelligence and biological synaptic pruning, a particle swarm optimization (PSO)-tuned regularized extreme learning machine (R-ELM) classifier with beta-weight centrality pruning conducts classification via a diverse softmax ensemble. On an imbalanced 18-class benchmark containing 916 daily appliance profiles, β-PruNet achieved 98.9% classification accuracy and a 97.6% macro F1-score while pruning 87.3% of hidden neurons (from 1500 to 191) without accuracy loss. Post hoc explainability analyses using SHAP, LIME, and Grad-CAM confirmed that the pruning criterion strongly correlates with Shapley feature attribution (r = 0.974). These findings indicate that integrating symbolic dynamics with bio-inspired pruning delivers a computationally efficient, highly accurate, and transparent solution suitable for edge-oriented energy management systems. The theoretical contribution is that the magnitude of a hidden unit’s learned output weights provides a usable ranking signal at no additional computational cost beyond the closed-form solution already obtained. The practical contribution is that this signal allows the classifier head to be compressed by 87.3% without a measurable loss in accuracy. Full article

Edge databases must coordinate replica freshness, consistency requirements, and governance restrictions despite intermittent connectivity. This study specifies an Edge-Aware Database Management System (EA-DBMS) architecture centered on the Policy-Aware Distributed Edge Catalog (PADEC). Its architectural novelty within the reviewed literature is a shared metadata interface connecting these constraints to operator placement, local transaction admission, and reintegration. This design provides a basis for keeping permitted operations local during outages and combining authorized results across sites under explicit contracts. A worked example demonstrates how joint filtering rejects prohibited and stale candidates. The analysis formalizes policy-constrained operator placement and distinguishes NP-hard general cases from three subclasses admitting polynomial-time exact solutions. A reproducible analytical evaluation derives conditional latency trade-offs from published measurements, documented proxies, and explicit assumptions. At a 10 ms proxy catalog cost, local reads benefit in one of four paired network scenarios at the low end of the published execution-time range. Under baseline assumptions, reductive aggregation becomes beneficial above approximately 401 kB per window at a reduction factor of 0.189. In the modeled mixture, global coordination reduces the mean-latency advantage. The design targets industrial telemetry, disconnected local processing, and data-residency-constrained analytics, with catalog lookup cost and matched edge–cloud execution times as priorities for empirical evaluation. Full article

The rapid growth of Internet of Things (IoT) networks has increased the need for intrusion detection mechanisms that provide accurate threat detection while preserving privacy, resisting malicious clients, and limiting communication overhead. Existing Federated Learning (FL)-based intrusion detection systems (IDS) remain challenged by heterogeneous non-IID client data, exposure of model updates, adversarial client participation, and frequent model transmission. To address these limitations, this paper proposes FLZT-IDS (Federated Learning with Zero-Trust for Intrusion Detection in IoT). The framework employs gradient-deviation monitoring to mitigate the effects of heterogeneous client updates, while Graph Neural Networks (GNN) and an Attention-based BiLSTM capture relational and temporal attack patterns, respectively. Secure Aggregation with Local Differential Privacy (SALDP) protects transmitted model updates, whereas dynamic trust-based client filtering limits the contribution of potentially malicious participants. An adaptive communication mechanism transmits local updates only when the model change exceeds a predefined threshold, thereby reducing unnecessary communication. Experiments on CICIDS2017 and Edge-IIoTset show that FLZT-IDS achieves 97.8% and 96.2% accuracy, respectively. Compared with FL-IDS, the adaptive communication and trust-filtering mechanisms reduce cumulative communication overhead by more than 30% after 100 communication rounds. Under model-poisoning experiments with 30% malicious clients, FLZT-IDS maintains accuracies of 89.3% on CICIDS2017 and 87.1% on Edge-IIoTset, demonstrating improved robustness under the evaluated adversarial setting. Full article

Accurate detection, quantitative assessment, and localization of delamination damage in composite materials remain major challenges in structural health monitoring (SHM). This paper proposes Adaptive Graph Transformer (AGFormer), a framework integrating a graph convolutional neural network with Transformer encoders, for size quantification, severity grading, and localization of circular delamination damage in carbon-fiber-reinforced polymer (CFRP) laminates. First, ten-dimensional edge features, including root-mean-square deviation, correlation coefficient, energy ratio, time-of-flight shift, and spectral variation, are extracted from differences between multipath Lamb-wave baseline and damaged-state signals. These features are aggregated into 42-dimensional enhanced node features to fuse multipath damage information at the sensor nodes. Second, a multimodal adjacency-matrix construction strategy combining sensor spatial distance with signal time–frequency distance is proposed to provide an adaptive topology for the graph convolutional network. On this basis, the AGFormer model employs three residual graph-convolution modules to extract local topological features, a multilayer perceptron for dimensional mapping, and two multi-head self-attention Transformer encoder layers to capture long-range dependencies among sensors. Attention-gated pooling and parallel output heads then simultaneously perform damage-radius regression and four-class severity classification. In addition, damage-centroid localization is achieved by combining multipath damage indices with the RAPID (Reconstruction Algorithm for Probabilistic Inspection of Damage) elliptical imaging algorithm. Experimental validation was conducted on a 450 mm × 450 mm six-ply CFRP laminate instrumented with a 12-piezoelectric-sensor network, under progressively increasing simulated delamination damage. The model was trained and evaluated using five-fold cross-validation grouped by damage size, while Gaussian-noise injection and random masking were introduced for data augmentation to improve generalization. The results demonstrate that the proposed method can effectively quantify and grade delamination damage of different sizes, achieving a mean absolute error of 4.38 mm, a root-mean-square error of 5.37 mm, and an R² of 0.920 for damage-radius regression, an accuracy of 81.3% for four-class severity classification, and a mean localization error of 6.50 mm, thereby offering a new technical approach to intelligent health monitoring of composite structures. Full article

Semantic segmentation of crops and weeds is an important component of selective weeding systems, but accurate segmentation networks can impose substantial computational and memory demands on resource-constrained agricultural platforms. This paper presents a systematic study of the accuracy–efficiency trade-offs involved in compressing semantic segmentation networks for agricultural edge computing. Six architectures spanning conventional CNN encoder–decoders, atrous and pyramid-pooling decoders, a mobile-oriented CNN, and two transformer scales are evaluated on the Crop/Weed Field Image Dataset (CWFID), with three random seeds per architecture. The most favourable compact model is subsequently evaluated under unstructured global magnitude pruning and dependency-aware structured channel pruning, using one-shot and iterative prune-and-fine-tune schedules at sparsity levels from 10% to 80%, with three seeds per configuration. Among the dense baselines, U-Net with a MobileNetV3-Large encoder matches the accuracy of the larger SegFormer-B2 model (88.85% versus 88.67% mIoU, a difference of 0.18% across three seeds, p = 0.58) with 3.7× fewer parameters than SegFormer-B2. Iterative unstructured pruning reaches 80% weight sparsity while retaining 89.04% mIoU; however, because the dense tensor dimensions are unchanged, this sparsity does not reduce the parameter count, GFLOPs, or model size on conventional hardware. Structured channel pruning, which does reduce all three, remains viable only up to 30% sparsity, reducing GFLOPs by 2.1× at a cost of 4.21 mIoU points and collapsing beyond 40%. Without fine-tuning, structured pruning causes severe degradation even at 10% channel sparsity. The results show that architecture selection has a major influence on the accuracy–efficiency trade-off, while the effects of pruning depend strongly on the pruning structure and fine-tuning strategy. Crop segmentation consistently degrades before the aggregate mIoU indicates severe failure, highlighting the need for class-wise evaluation when compressing crop–weed segmentation models. Full article

Low-altitude remote sensing utilizing unmanned aerial vehicles (UAVs) has evolved into a core pillar of the space–air–ground integrated perception architecture. However, research on detection tasks in UAV aerial imagery grapples with three principal hurdles: feature annihilation of small targets, insufficient robustness under intricate environmental variations, and constrained computational capacity of onboard edge devices. To tackle these issues, a lightweight aerial small-target detection model denoted as Aero-YOLO26 was proposed for aerial image object detection. It introduced a C2f_GhostECA lightweight feature extraction module and a SlimSPPF streamlined multi-scale pooling module, and optimized the global receptive field of the Efficient Channel Attention (ECA) mechanism. Furthermore, it reconstructed the backbone and Path Aggregation Network (PAN)–Feature Pyramid Network (FPN) fusion network, enabling lightweight co-optimization of extraction, pooling, and fusion. Experiments on the VisDrone2019 dataset show that Aero-YOLO26n reduces the parameter count by 22.6% compared with YOLO26n, while improving precision, recall, mAP@50, mAP@50–95, and F1 score by 14.9%, 15.2%, 11.4%, 11.7%, and 14.8%, respectively. On the self-built UAV-MCT dataset, it attains 85.61% precision, 83.78% recall, mAP@50 of 90.82%, and mAP@50-95 of 59.60%, with average precision (AP) for cars and pedestrians reaching 99.4% and 96.6%. The proposed lightweight computer-vision object detection model using UAV-acquired imagery markedly bolsters the detection performance for small objects in UAV-acquired aerial imagery while achieving extreme model size compression, and its excellent performance demonstrates its potential for a wide range of application scenarios. Full article

Security dashboards summarise large event streams into convenient hourly, daily, weekly, and monthly views, but aggregation can change what analysts see and therefore what they prioritise. This case study examines four overlapping Check Point Quantum Spark 1530 reports from a single enterprise edge gateway: 1 h, 24 h, 7 days, and 30 days, each ending on or near 15 September 2023. We reconstruct traffic series, normalise volumes by duration, compare application concentration, reconcile summary and detailed tables, and separate control-confirmed events from unadjudicated indicators. The one-hour window corresponded to a duration-normalised 10.10 GB/day, 2.86 times the 30-day rate of 3.53 GB/day, whereas the 24 h and seven-day rates were 3.86 and 3.51 GB/day. Top five application concentration declined monotonically from 88.8% at one hour to 17.5% at 30 days. The hourly and daily summaries reported zero Anti-Bot, Anti-Virus, and intrusion-prevention events. Yet, the seven-day report contained 15,641 IPS incidents and four Anti-Virus events, and the 30-day report contained 136,057 IPS incidents, 28 Anti-Virus events, and two Anti-Bot events. Reconciliation also exposed non-additive rounded totals, partial-bin coverage, conflicting high-risk-application counts, and a weekly QUIC session count larger than its 30-day counterpart. An alignment-controlled comparison of the exactly nested 7-day/30-day pair showed nearly identical traffic rates in the final week and the preceding 23 days (3.51 versus 3.54 GB/day). In contrast, a containment bound showed that part of the apparent collapse in leading-application dominance reflects vendor reporting logic rather than aggregation. Within the limits of this single-gateway case study, the results indicate that short windows are useful for burst triage but unsafe as stand-alone evidence of security state. We propose, but do not yet validate, a multi-timescale workflow for converting vendor telemetry into cyber-risk intelligence through alignment, normalisation, reconciliation, corroboration, and explicit confidence labels. Full article