Vol. 2 No. 2 (2026)

Published October 7, 2026
Download Full Issue (PDF)

Articles in This Issue

Abstract

The combination of the Internet of Things (IoT) and Software-Defined Networking (SDN) has quickly revolutionized the world of industry, and also brought with it some severe security risks. Despite its recent success, Deep learning (DL) can be a powerful method for Network Intrusion Detection systems (NIDS), but often the computational and memory requirements are too high for edge devices. In this paper, a collaborative optimization framework with lightweight deep-learning IDS models suitable for SDN-IoT environments is proposed. We formalize and test three collaborative compression schemes based on combining pruning, clustering, and quantization, yet maintaining the structural advantages of each stage: Sparsity-Preserving Quantization-Aware Training (PQAT), Cluster-Preserving Quantization-Aware Training (CQAT), and Sparsity-and-Cluster-Preserving Quantization-Aware Training (PCQAT). The three schemes are evaluated on four canonical architectures (LSTM, RNN, DNN and CNN) on the proposed SICA dataset with respect to DDoS, Probe and Brute-Force (BFA).The results show that aggressive single-stage compression such as clustering can shrink model size by up to 94%, but at the cost of predictive confidence; PCQAT, by contrast, acts as a stabilizer that recovers this lost confidence. In particular, CNN-PCQAT and DNN-PCQAT deliver the most favorable trade-off, sustaining high detection accuracy on all three attack classes with inference latency below 0.01 ms. The framework therefore offers a practical, deployment-oriented guideline for implementing real-time IDS on resource-constrained SDN-IoT edge devices.

Abstract

With the rapid popularization of 5G edge communication systems, their decentralized architecture and the ultra-large attack surface brought by massive access devices have created severe network security challenges. Although current artificial intelligence-based intrusion detection systems and conventional federated learning schemes can achieve high intrusion detection accuracy, they universally suffer from two major limitations: excessive computational overhead and vulnerability to poisoning attacks, which make them unable to adapt to resource-constrained edge deployment environments. To address this issue, this paper proposes a lightweight anti-poisoning federated learning framework for low-latency secure 5G edge communications, with two core design elements: First, it integrates lightweight GRU/CNN models and a trust-aware aggregation mechanism to weaken the negative impact of malicious model updates; second, it introduces a deviation-based poisoning detection strategy to identify and filter adversarial updates during the federated learning process. We used the public CICIDS2017 dataset to complete tests in a simulated 5G edge environment equipped with 100 to 1000 edge devices, covering multiple scenarios including DDoS, MITM, and poisoning attacks. The results show that compared with existing similar schemes, the proposed framework reaches a maximum detection accuracy of 98.3\%, reduces communication latency by 30-35\%, cuts computational overhead by 40-60\%, and sees only a negligible drop in accuracy under poisoning attacks. It can support a wide range of real-time 5G edge applications such as intelligent transportation, smart cities, and the industrial Internet of Things.

Original Article
Decentralized Federated Learning for Multi-Class Predictive Maintenance in Industrial IoT Systems
PDF Full Text
Abstract

Predictive maintenance (PdM) is a key enabler of reliability and cost reduction in Industrial Internet of Things (IIoT) systems. However, conventional centralized learning approaches require raw data aggregation, raising serious concerns regarding data privacy, communication overhead, and single points of failure. Federated Learning (FL) addresses part of this challenge by enabling distributed model training, yet centralized FL architectures still rely on a central aggregator. This paper proposes a decentralized federated learning (DFL) framework based on a ring topology for multi-class predictive maintenance in IIoT environments. The proposed approach enables collaborative model training among distributed edge nodes without relying on a central server, thereby enhancing system robustness, scalability, and data privacy. A lightweight deep neural network (DNN) is employed to classify five machine operating states using multi-sensor data, including vibration, temperature, and current measurements. Extensive experiments are conducted using a real-world sensor dataset collected from three-phase AC motors under multiple fault conditions. The performance of the proposed DFL approach is systematically compared against a non-federated baseline and a centralized FL (CFL) scheme. Experimental results demonstrate that the decentralized framework achieves clas-sification performance comparable to centralized FL, with an accuracy of up to 99.83%, while eliminating the dependency on a central coordinator. The findings confirm the feasibility of decentralized federated learning for predictive maintenance applica-tions, particularly in distributed industrial settings where data privacy, fault tolerance, and communication efficiency are critical.

Abstract

The increased computational loads of the Internet of Things (IoT) devices have been increasing exponentially, approaching 29.4 billion connected units in the world by 2030, which is beyond the processing capabilities of small-scale IoT devices. Mobile Edge Computing (MEC) reduces latency by offloading tasks to edge servers, but coordinating offloading across heterogeneous MEC servers, while minimizing energy use and meeting strict service-level agreements (SLAs), remains an open problem. Rule-based load balancers cannot handle dynamic, non-stationary IoT workloads, and existing DRL methods converge slowly and waste energy under variable server loads. In this paper, a proposal of an Energy-Aware Load Balancing framework of MEC-IoT settings is presented, which utilizes Double Deep Q-Network (DDQN) and Prioritized Experience Replay (PER) to learn adaptive offloading policies, and Dynamic Voltage and Frequency Scaling (DVFS) to control fine-grained MEC server energy expenditure. The DDQN agent sees a single state description of the IoT task features (type, size, deadline) and real-time load on the MEC server, queue lengths, and leftover energy budgets and chooses an offloading destination: local processing, one of five edge servers, or cloud to optimize a hybrid reward function that punishes both task completion time and cumulative system energy usage. The framework is tested with the high-fidelity SimPy discrete-event simulation, which is, in turn, calibrated to the real MEC testbed parameters, and compared to seven baselines with different IoT device counts (20 to 200 devices) and workload intensities. The proposed framework cuts total energy use by 45.9% versus random offloading and 19.5% versus the best DRL baseline (A3C). It also reduces mean latency to 41.8ms (70.6% lower than random offloading) and keeps the SLA violation rate at 3.2%, below the 5% target. Its advantage holds across all tested device scales.​

Abstract

Breast cancer is one of the leading causes of mortality among women worldwide and necessitates an accurate and clinically valid mammographic assessment. Although recent deep learning frameworks have achieved promising performance in breast lesion detection, many existing frameworks focus primarily on lesion localization without providing comprehensive decision-support functionalities for clinical practice. In this study, a multi-layer clinical decision support framework is developed for mammographic breast lesion assessment by integrating YOLOv8x-based detection, ROI-based quantitative image descriptors, ROI/contour-derived geometric descriptors, physician-selected Breast Imaging Reporting and Data System (BI-RADS) assessment with associated risk visualization, and physician-editable structured report generation, all within a single workflow. The framework was developed and evaluated using 501 mammographic images. The dataset was divided into 350 training, 100 validations, and 51 test images. In the standard curve-based evaluation, the detector achieved 99.50% mAP@50 and 99.16% mAP@50-95. At the fixed confidence operating point of 0.40, the detector achieved 97.87% precision, 100.00% recall and a 98.92% F1-score. The platform extends beyond automated detection into quantitative ROI analysis, physician-editable reporting, automated PDF generation, and clinician-supervised mammographic decision support. The system does not perform autonomous diagnosis or automatic BI-RADS assignment.