This paper presents a project-based learning module that unifies sensing, communications, and edge computing around a cooperative ultrasonic "radar" (SONAR) network. The module was delivered during the Internet Festival as a pre-university outreach lab for Italian high-school students (ages 16–19). Teams of students assembled scanning ultrasonic nodes, operating at 40 kHz, streamed measurements via MQTT using a private network, and executed a Moore–Penrose pseudoinverse multilateration pipeline on an edge laptop inside the same private network to track objects on a tabletop arena. By integrating hardware prototyping, networking, and mathematical modeling, the module provided a tangible demonstration of how modern distributed systems "think together." It also highlighted cooperation as both a technological paradigm, where multiple nodes collaborate to solve an overdetermined sensing problem, and an educational one, where students collectively build and reason about an intelligent network. The resulting hands-on activity effectively bridged theory and practice, offering students an accessible entry point into the principles underlying the Internet of Everything.

Continuous and objective monitoring of livestock behavior plays a key role in precision farming, animal welfare assessment, and reproductive management. This study proposes a non-invasive framework for sheep behavior and reproductive activity monitoring that integrates wearable actigraphy, machine learning, and a cloud-based data processing architecture. Tri-axial accelerometer data were collected at 30 Hz using collar-mounted ActiGraph sensors under real farming conditions. Raw acceleration signals were processed without temporal aggregation, preserving full temporal resolution that includes axis-specific acceleration, vector magnitude, and delta magnitude features. Several supervised learning models were evaluated for behavior classification, including BLSTM, LSTM, CNN–BLSTM, Random Forest, and Support Vector Machine, targeting behaviors such as standing, walking, grazing, lying, flehmen, and mating. The results indicate that both deep learning and classical machine learning approaches achieve high classification performance, with Random Forest obtaining an overall accuracy of 0.82, while deep sequential models effectively capture temporal patterns and behavioral transitions. Furthermore, a scalable cloud architecture is introduced to automate data ingestion, preprocessing, inference, storage in InfluxDB, and visualization through an interactive web application. The proposed framework supports continuous monitoring and offers practical tools for precision livestock management.

The integration of Quantum Key Distribution (QKD) into prospective 7G communication systems represents a promising approach to provide information-theoretic security for future networked applications. In such environments, QKD infrastructures are expected to support heterogeneous services with diverse security and performance requirements under strict resource constraints. However, QKD performance is inherently sensitive to dynamic channel conditions, which directly affect key metrics such as the Quantum Bit Error Rate (QBER) and the Secret Key Rate (SKR). Since these parameters determine the achievable Quality of Service (QoS), their accurate and efficient prediction is essential for effective system optimization. In this work, we propose an ultra-efficient QoS prediction framework for QKD systems based on a Fuzzy-Augmented Neural Network (FANN) architecture. The proposed approach decomposes the prediction task into a regression stage, followed by an application-specific fuzzy inference layer that maps predictions into QoS classes. This decoupled design enables a lightweight, modular, and extensible solution suitable for resource-constrained QKD nodes. The proposed framework supports dynamic quantum channel scheduling, allowing applications to access key generation resources according to their security and throughput requirements, with direct benefits in terms of energy efficiency at both control and physical layers. Experimental results on a publicly available QKD dataset demonstrate that the proposed approach achieves comparable or improved F1-score performance with respect to a conventional Feed-Forward Neural Network (FFNN) classifier, while reducing inference time by over 98% and memory usage by more than 71%, even compared to FFNN models that leverage classical compression techniques, i.e., pruning and quantization.

Software Defined Wide Area Network (SD-WAN) is gaining attention in view of the integration of Non-Terrestrial Network (NTN) to guarantee a resilient connection, also in remote areas. In this paper, we present a real test-bed implementation that poses the basis for an innovative SD-WAN architecture designed for the 6 G era that integrates terrestrial links and Low-Earth-Orbit (LEO) satellite connectivity. In particular, we describe a control-plane centric design that leverages ONOS as the SDN controller, Open vSwitch as the programmable data plane, and OpenFlow for southbound control, implemented and validated on a small-scale, real testbed with a Starlink connection. Moreover, we introduce a policy-driven module for link selection and failover. The ultimate goal is to provide a reproducible tutorial for deploying the SD-WAN testbed, thereby offering the scientific community a realistic experimental environment to evaluate advanced failover mechanisms and to investigate the optimal integration of LEO satellite networks.

This paper presents DroneOPERA, an ns-3 simulation module to model the energy consumption of onboard computing in energy-constrained unmanned aerial vehicles (UAVs) operating in Internet of Drones (IoD) networks. As drone networks are increasingly integrated with IoT and cloud services, energy-efficient processing becomes a key challenge, especially in applications requiring real-time analytics under strict power budgets. DroneOPERA extends ns-3 with a configurable model that couples UAV mobility, battery dynamics, and CPU workloads associated with on-board tasks. By processing data locally instead of offloading raw information to external servers, DroneOPERA enables the study of privacy-preserving mission designs, where communication is limited to compact, processed results. The module exposes APIs to define mission profiles, processing workloads, and hardware configurations, and computes the resulting energy expenditure at fine time granularity. We demonstrate the capabilities of DroneOPERA through a representative use case, in agriculture scenario, highlighting the trade-offs between task complexity, flight time, and privacy-preserving on-board processing.

Artificial lighting represents a major source of energy consumption in greenhouse production, while optimal photosynthetic photon flux density (PPFD) requirements vary across space and time. This paper introduces a multi-digital twin adaptive lighting control framework in which multiple digital twins operate in parallel to evaluate heterogeneous control strategies under dynamic environmental conditions. Rather than relying on a single fixed controller, the system dynamically selects both the controller structure and its parameters by continuously simulating alternative control hypotheses at the edge. The proposed architecture is supported by an IoT-enabled communication infrastructure, integrating low-power sensor networks and IP-based actuator control to enable real-time interaction. Experimental validation in a real greenhouse demonstrates accurate multi-zone PPFD regulation and significant energy savings compared to a conventional on/off strategy.

The increasing adoption of Low Earth Orbit (LEO) satellite networks, such as SpaceX's Starlink, introduces new opportunities and challenges in future 6G networks. While LEO satellites offer low-latency global connectivity, their link quality exhibits periodic variations due to satellite handovers and variable atmospheric conditions. This paper investigates the feasibility of using LEO satellite Internet for enterprise networks within a Software-Defined Wide Area Network (SD-WAN) framework. We built and evaluated a dual-tunnel SD-WAN architecture that integrates Starlink with a terrestrial network. Moreover, we compare three switching strategies for tunnel selection: a deterministic, an Exponential Moving Average (EMA)-based deterministic, and a Machine Learning (ML)-based approach. Our trial includes real-world traffic measurements collected in an experimental setup with a real Starlink dish located at Stanford University (US). We report key metrics including average delay, packet loss, and number of switching events. This work aims to offer valuable insights into building robust, adaptive hybrid LEO/terrestrial networks for the 6G era.

Low Earth Orbit (LEO) satellite constellations such as Starlink are emerging as a transformative solution for providing global broadband connectivity, including environments like aeronautical and transoceanic flights for future 6G networks. This paper presents an evaluation of Starlink’s performance during a commercial flight from Hawaii to Japan. By analyzing ICMP ping measurements collected throughout the journey, we assess key metrics such as latency, jitter, packet loss, and temporal stability. We further investigate statistical characteristics and frequency-domain patterns to analyze satellite handovers and oscillatory behavior in delay. Moreover, we performed a throughput evaluation using a standard speed test protocol. Our results show that while Starlink provides generally low-latency connectivity even in flight, its performance exhibits significant fluctuations and occasional disruptions, likely due to the challenges of maintaining stable satellite links in a fast-moving, highaltitude context. Moreover, we released an open-source repository for collecting datasets about in-flight Starlink measurements. This study contributes empirical insight into the feasibility and limitations of LEO satellite Internet for aeronautical mobility applications.

Wildlife monitoring in remote field environments presents significant challenges owing to energy constraints, limited network connectivity, and the need for real-time processing. This study proposes a scalable framework that integrates artificial intelligence (AI) and edge computing for autonomous wildlife detection, class-based object counting in frames, distance measurement, and geolocation estimation of key European wildlife species. The system leverages deep learning-based object detection models, specifically YOLOv8m and YOLOv10m trained by the European Animal Detection Dataset, deployed on power-efficient embedded platforms such as the NVIDIA Jetson Orin Nano and Raspberry Pi 5. The experimental evaluation assesses detection accuracy, computational efficiency, and energy consumption across different hardware configurations, ensuring an optimal trade-off between performance and resource utilization. Our findings demonstrate that the proposed approach enables real-time inference while maintaining high mean average precision (mAP) and minimizing power consumption, ensuring feasibility in resource-limited field conditions. A comprehensive cost-performance analysis highlights the Jetson Orin Nano as the optimal choice for balancing accuracy and efficiency, achieving 40.1 Frame Per Second (FPS) at just 8W in 25W power mode using a TensorRT-optimized YOLO model. Meanwhile, the Raspberry Pi 5 serves as a cost-effective alternative for lightweight models. Beyond technical performance, this work underscores the transformative role of edge AI in supporting ecologists and agronomists with actionable insights for biodiversity conservation, human wildlife conflict mitigation, and sustainable ecosystem management.

Networked Music Performance (NMP) is one of the most challenging real-time applications in which musicians can play together using the internet without being physically together. Nowadays, the connection technology can have performance that guarantees an adequate Quality of Experience (QoE) for NMP. On the other hand, there are millions of musicians who live in remote and rural areas without access to high-speed network connections. The promising LEO satellite internet technology could fill this gap and democratize NMP in remote places. This paper aims to provide an analysis of Starlink’s capability to meet the stringent requirements of NMP. We present an analysis of network metrics that are relevant to NMP applications and assess the NMP software JackTrip in order to test a real NMP application with LEO satellite internet. The evaluation of the RTT, One-way delay, Packet loss, and jitter in different scenarios poses some challenges for the NMP scenario. We present the results of two live NMP jamming sessions with musicians, proposing an adapted QoE model that incorporates both objective and subjective metrics. With a tailored buffer strategy and the ability of a musician, we were able to obtain an adequate QoE, overcoming the challenges introduced by LEO satellite internet.

In rural areas such as national parks, forests, and mountains, recognizing and classifying wildlife is crucial for monitoring animals that may pose risks to crops and human safety. Deep learning provides the most accurate approach due to its dynamic adaptability. However, it is also highly energy intensive. Given the constraints of rural environments, including the use of extreme edge devices and limited power availability, this study employs distributed computing to maximize battery life while enabling the execution of DL tasks. The core concept is Split Computing, specifically applied to the YOLOv8m and YOLOv10 models, which have been provided by researchers as a highly effective solution for wildlife classification. This approach dynamically reallocates different model components across available boards, specifically the NVIDIA Jetson Orin Nano, to optimize energy consumption. This research investigates various model distribution configurations across the head, backbone, and neck components, assessing whether video stream compression between the camera and the computing boards affects energy consumption and network load. Our findings suggest that the choice of model splitting configuration significantly impacts energy efficiency, bandwidth consumption, and computational load distribution.

The evolution of Low Earth Orbit (LEO) satellite systems is poised to play a key role in shaping the concept of Global Net Equity in the 6G landscape. This paper explores the integration of LEO satellite connections within terrestrial and non-terrestrial network (TN/NTN) systems, presenting them as essential components for future global connectivity. Leveraging measurements from the commercial solution Starlink, we analyze the current performance of LEO satellites, highlighting their limitations in stability for real-time applications such as Networked Music Performance (NMP). Using JackTrip, we conducted jamming sessions to evaluate the feasibility of NMP over Starlink, highlighting key challenges related to latency and synchronization. Moreover, our data collection (publicly available) took place during a solar storm, providing a unique opportunity to observe its impact on the Starlink constellation. While significant obstacles remain, notably in achieving the stability required for latency-sensitive applications, the study underscores the promise of LEO satellites in addressing connectivity needs where terrestrial networks fall short.

Federated Learning (FL) has emerged as a privacy-preserving alternative to centralized machine learning, enabling collaborative model training without sharing raw data. This is particularly valuable in smart agriculture, where farmers generate sensitive, distributed data that must remain localized due to privacy, legal, and competitive constraints. However, FL introduces new security vulnerabilities stemming from its decentralized nature. This paper investigates these vulnerabilities through a case study involving real-world microclimate time series data from four agricultural greenhouses. We evaluate the effects of Data Poisoning, Model Update Poisoning, and Backdoor Attacks strategies and implement multiple defense mechanisms, including Robust Aggregation, Differential Privacy, and Sanitization. Experimental results demonstrate that while individual defenses can mitigate specific attacks, a comprehensive, multi-layered defense strategy offers superior resilience. We propose a Final Defense framework that combines k-NN-based data sanitization, Trust-Krum aggregation, adaptive clipping, and differential privacy. This configuration consistently achieves the highest F1-Scores across all attack scenarios, enhancing both robustness and convergence efficiency. Our findings confirm the necessity of integrated defenses for secure, scalable deployment of FL in adversarial environments such as smart agriculture.

This research presents an energy-efficient monitoring system for smart greenhouses aimed at reducing power consumption and enhancing sustainability. The study focuses on optimizing the sampling time of monitored environmental parameters to minimize energy usage while maintaining data accuracy. Utilizing an optimal sampling approach significantly reduces energy consumption compared to fixed sampling intervals commonly used in commercial weather stations. Additionally, the implementation of an energy harvesting system, powered by solar panels, ensures the autonomy of sensors within the greenhouse, reducing the need for human intervention in battery replacement and recharging.

The preservation of biodiversity and the monitoring of remote ecosystems require autonomous and low-impact technological solutions. This work presents an integrated system for wildlife monitoring that combines solar-powered Edge AI smart cameras, local connectivity via Wi-Fi HaLow (IEEE 802.11ah), and long-range backhaul through Low Earth Orbit (LEO) satellite communication using Starlink. The proposed architecture enables real-time animal detection with a custom-trained YOLOv8m model and efficient data transmission to a Green Data Center, which is placed in University of Pisa, Italy, for storage and visualization. The focus of this paper is on the integration of LEO satellite internet for connecting the entire architecture, which is located in a remote area. Performance tests conducted in a Starlink-to-Starlink configuration evaluated latency, jitter, and throughput. Results show that it remains insufficient for bandwidth-intensive or ultra-low-latency applications, but these findings confirm the feasibility of LEO satellite-based connectivity for scalable and sustainable wildlife monitoring in remote areas.

Smart greenhouses offer crucial solutions for reducing our atmospheric impact and resource waste. However, two fundamental challenges persist in their implementation, massive energy consumption and a high level of human intervention, particularly for sensor battery replacement or recharging. Unfortunately, sensors are indispensable in greenhouses and agriculture, such as for monitoring environmental parameters for air quality assessment. Therefore, while sensors cannot be eliminated, it is essential to optimize their energy consumption. This work introduces an energy-efficient monitoring system for smart greenhouses, aiming to reduce the energy consumption of individual sensors and enhance system sustainability. This study focuses on optimizing the sampling intervals of commonly monitored environmental parameters to minimize sensor energy usage while maintaining data acquisition accuracy adequate for the intended purpose. Additionally, to further reduce battery energy draw, an energy harvesting system using solar panels was implemented. In conclusion, adopting an optimal sampling strategy for each parameter significantly reduces energy consumption compared to fixed, inefficient sampling intervals commonly used in commercial weather stations. Furthermore, by employing an energy harvesting system for each sensor, leveraging the light emitted by greenhouse lamps and external sources ensures the autonomy of sensors within the greenhouse, thereby minimizing the need for human intervention for battery replacement and recharging.

Climate change poses significant challenges, particularly in agriculture, where extreme weather events demand more efficient and resilient systems, such as smart greenhouses. These controlled environments require predictive solutions to optimize conditions such as temperature and humidity, which are critical for crop growth. While neural networks are widely employed for climate prediction, their complexity presents a barrier to implementation on edge devices, which are characterized by limited computational resources. In this work, we propose Fuzzy-Augmented Neural Network (FANN), a novel approach based on fuzzy sets, applied in cascade to regressive neural network models, to reduce complexity and energy consumption in greenhouse microclimate classification and prediction. The methodology was tested on four edge devices, including microcontrollers and microprocessors. We compare our FANN approach with standard models (FFNN, BNN, SimpleRNN, GRU, LSTM), highlighting significant reductions in inference time, energy consumption, and memory usage. FANN also offers practical advantages, such as the ability to adapt classification by modifying fuzzification parameters without retraining the model, and the potential to parallelize computations for simultaneously classifying the microclimate of multiple crops. These features make the system flexible and optimal for practical applications in dynamic agricultural contexts.

Modern Software-Defined Wide Area Network (SDWAN) deployments are required to manage traffic over heterogeneous underlay networks while meeting stringent Quality of Service (QoS) requirements. In scenarios where multiple branches share overlay resources, independent tunnel selection decisions often lead to congestion and degraded performance. Existing approaches lack coordination mechanisms to handle the dynamic interactions between agents competing for shared resources. This paper presents a Multi-Agent Reinforcement Learning (MARL) framework for distributed overlay selection in SD-WANs. Each branch is modeled as an autonomous agent that learns routing policies through interaction with the network environment. To account for the mutual impact of decisions across branches, we adopt the Centralized Training with Decentralized Execution (CTDE) paradigm, enabling agents to learn globally consistent behaviors while preserving scalability at inference. To encourage cooperative policies, we introduce a $\lambda$-weighted reward shaping mechanism that balances local QoS goals with global resource fairness. We evaluate our approach using both PPO and DQN algorithms in a simulated SD-WAN environment. The findings highlight the necessity of MARL in addressing resource contention and ensuring equitable shared overlay utilization.

The increasing complexity of modern distributed systems, spanning cloud computing, edge devices, and ultra-low latency applications, needs accurate and scalable network emulation. This comparative study analyzes the fidelity of one of the most popular network emulators (Mininet) against a physical testbed to quantify performance deviations with a focus on scalability. Using iPerf for benchmarking, we measure bandwidth allocation dynamics under incremental client loads (1-20 nodes). Results show that Mininet accurately emulates throughput distribution across different nodes with errors below 9%. We further discuss Mininet's limitations in replicating high-bandwidth links and propose a Python-based topology configuration tool to streamline emulation workflows. This work provides empirical insights into the fidelity and constraints of software-defined network emulation and the possibility of integration in a digital twin system.

Lung disease diagnosis from chest X-ray images is a critical task in clinical care, especially in resource-constrained settings where access to radiology expertise and computational infrastructure is limited. Recent advances in deep learning have shown promise, yet most studies focus solely on either classification or detection in isolation, rarely exploring their combined potential in an embedded, real-world setting. To address this, we present a dual deep learning approach that combines five-class disease classification and multi-label thoracic abnormality detection, optimized for embedded edge deployment. Specifically, we evaluate six state-of-the-art CNN architectures—ResNet101, DenseNet201, MobileNetV3-Large, EfficientNetV2-B0, InceptionResNetV2, and Xception—on both base (2020 images) and augmented (9875 images) datasets. Validation accuracies ranged from 55.3% to 70.7% on the base dataset and improved to 58.4% to 72.0% with augmentation, with MobileNetV3-Large achieving the highest accuracy on both. In parallel, we trained a YOLOv8n model for multi-label detection of 14 thoracic diseases. While not deployed in this work, its lightweight architecture makes it suitable for future use on embedded platforms. All classification models were evaluated for end-to-end inference on a Raspberry Pi 4 using a high-resolution chest X-ray image (2566 × 2566, PNG). MobileNetV3-Large demonstrated the fastest latency at 429.6 ms, and all models completed inference in under 2.4 s. These results demonstrate the feasibility of combining classification for rapid triage and detection for spatial interpretability in real-time, embedded clinical environments—paving the way for practical, low-cost AI-based decision support systems for surgery rooms and mobile clinical environments.