This demo presents a novel implementation of physical layer key generation (PLKG) in UAV communication systems by integrating deep learning for key reconstruction. Secure communication is a significant challenge in UAV networks due to their high mobility, frequent topology changes, and vulnerability to eavesdropping. Traditional cryptographic methods are inefficient in such dynamic environments because of high computational costs and latency. Our approach leverages multimodal learning to enhance resilience against Doppler effects and dynamic channel variations. The demonstration showcases a working prototype that extracts channel state information (CSI), predicts UAV trajectory, and reconstructs cryptographic keys with improved consistency and lower mismatch rates compared to conventional methods. Our system utilizes ESP32 microcontrollers for real-time CSI acquisition and Raspberry Pi 4 for deep learning-based processing. We provide a graphical user interface (GUI) that visualizes real-time CSI fluctuations and reconstructed key bits, demonstrating the framework’s resilience to dynamic channel variations.
TCP's slow start phase is particularly inefficient over most wireless networks, especially high-latency, high-bandwidth paths such as satellite networks, often exiting too early or too late (after packet loss). To address this, the Slow start Exit At CHokepoint (SEARCH) algorithm is designed to improve exit decisions during slow start by analyzing delivery trends across sliding RTT-based windows. This paper presents a first implementation of SEARCH in the FreeBSD kernel using FreeBSD's modular congestion control framework. We evaluate our implementation on a testbed with an actual GEO satellite link with similar to 600 ms RTT and 150 Mb/s capacity. Preliminary results show that SEARCH exits slow start more effectively than HyStart and HyStart++, achieving higher throughput and better utilization.
Digital twin technology plays a pivotal role in smart manufacturing by enabling real-time monitoring, predictive maintenance, and operational optimization. However, traditional digital twins often rely solely on text-based sensor data, resulting in limited system granularity and incomplete process representation. To address this gap, we incorporate image data from mobile devices to enrich digital twin construction with greater detail and accuracy. Given the high computational demands and strict latency requirements of image processing, we propose a Mobile Edge Computing (MEC)-based framework that offloads intensive tasks to nearby edge servers. To minimize system-wide delays, we develop the Min-Max Completion Time (MMCT) algorithm, which strategically prioritizes tasks with the longest processing time to reduce the overall completion time. Simulation results under realistic parameter settings demonstrate that MMCT significantly outperforms baseline approaches—including Pure Local Processing, Pure Offloading, and the Game-Theoretic Offloading Decision Method (GTODM)—especially in scenarios involving high data volume and computational complexity. By improving task responsiveness and reducing latency, our framework ensures timely and comprehensive digital twin updates. We also highlight the potential of integrating generative AI to further enhance task allocation and decision-making in MEC-driven digital twin systems.
The next-generation of mobile networks will increasingly rely on Artificial Intelligence (AI)/Machine Learning (ML) for effective network automation, resource orchestration and management. This translates into performing classification and regression tasks on time series data. Unfortunately, the existing AI/ML models are inherently complex and hard to interpret, which hinders their deployment in production networks. Further, the vast majority of the existing EXplainable Artificial Intelligence (XAI) techniques are either primarily conceived for computer vision and natural language processing and thus fail to provide useful insights.In this paper, we take the research on XAI for time series classification and regression tasks one step further proposing ChronoProf, a new tool that builds on legacy XAI techniques. By creating a linearized version of the original model for different observations, ChronoProf provides insights about the dynamic changes in the model decision-making process across observations and is agnostic to the influence of feature magnitude, which is a key limitation of legacy explainers. Thus, ChronoProf highlights the real influence of model parameters on the output. Our extensive evaluation with real-world mobile traffic traces shows that ChronoProf is able to measure the feature importance, especially in classification tasks where linearized explanations across observations show high consistency.
Low-Earth Orbit (LEO) satellite constellations are becoming a critical component of the global networking infrastructure. Integration of mesh networks of LEO satellites to the Internet and other mission-critical systems has the potential to revolutionize the global networking infrastructure. Routing traffic is a crucial need for the emerging LEO satellite networks to support communication between ground stations. In this paper, we propose a distributed routing mechanism, Two-Phase Geometric Routing (TPGR), that utilizes the geometry of the satellite constellation. TPGR splits the task of delivering a packet to its destination satellite into two phases: (1) Reaching the destination plane and (2) Reaching the destination satellite. This divide-and-conquer approach enables TPGR to handle all possible voids smaller than a hemisphere and all dead-ends, without any support from the Global Positioning System (GPS). On various Walker Delta satellite constellations, we compare TPGR to shortest-path routing and a simpler geometric routing mechanism. Our experiments show that TPGR attains path failure rates similar to shortest-path routing while incurring about 5ms additional latency on our baseline constellations with high inclinations and 2ms additional latency on the Starlink and Kuiper mega-constellations.
Thanks to its ability to provide secure and efficient communication capabilities, network slicing has been adopted as a key technology, not just in commercial networks but also in tactical networks. Based on multiple virtual networks, each dedicated to a specific service, a sliced architecture meets the diverse requirements of highly heterogeneous tactical services. However, tactical networks operate in challenging environments with inherent power constraints, where the need for energy-efficient network slicing management solutions is paramount. In this direction, we tackle a joint slice activation/deactivation and user association problem with the aim of studying trade-offs between energy efficiency and user quality of service. To solve the problem, we introduce our original approach: a split reinforcement learning-based energy-efficient slicing deployment algorithm, namely SNOW. SNOW divides the deep neural network into multiple sections, where the front-end part of the model is trained over multiple user devices and then the back-end part of the model is trained by the central nodes (i.e. base stations in this case), without sensitive data sharing. Extensive simulation results reveal that the proposed scheme is superior to the considered benchmarks in improving energy efficiency while maintaining network performance.
Cryptographic agility is becoming increasingly essential in mobile networks due to the rapid evolution of network architectures, the growing complexity of security threats, and the anticipated arrival of quantum computing. As mobile networks transition from 5G to 6G, ensuring the adaptability of cryptographic systems to emerging standards and threats is paramount. This paper introduces a novel Cryptographic Management Function (CMF), a centralized software-defined cryptography framework designed to achieve seamless cryptographic agility in mobile networks. The CMF is introduced to enable dynamic, seamless updates to cryptographic algorithms and protocols without disrupting ongoing operations, making it ideal for environments with varying device capabilities and evolving security requirements. It operates by decoupling cryptographic functions from the other network components, centralizing the enforcement of cryptographic policies, and supporting hybrid cryptographic schemes, including Post-Quantum Cryptography (PQC).
The Slow start Exit At Right CHokepoint (SEARCH) algorithm is designed to exit the TCP slow start phase after the flow has reached the link capacity but before packets have been lost. To do this, SEARCH keeps a history of the bytes delivered over a recent time window, aggregated into bins. Unfortunately, this delivery history must be kept per-flow, adding additional memory load for each TCP connection. We address this per-flow memory load by observing that SEARCH only needs the relative number of bytes delivered and propose a bit-shifting technique that dynamically compresses bin values as needed. Our approach is tunable to the memory-use reduction required compared to the delivery precision needed. Evaluation of our approach over a satellite network shows SEARCH bin memory use can be reduced by 50% or even 75% without any significant sacrifice in SEARCH algorithm accuracy. Our approach is generalizable to other network algorithms, too, reducing memory use for algorithms that use sliding windows and historical data tracking.
This paper presents empirical results from the use of multi-UAV ad hoc networks that also perform edge computing through the deployment and execution of microservices requested by user equipment on the ground. For this purpose, the most important aspects on how the UAV-based mesh network is configured and how user equipment on the ground accesses it are detailed. Four different tests are provided, all conducted with affordable user-grade hardware, explained and analyzed. Results demonstrate how the use of mesh-connected swarms of UAVs is a viable solution for providing connectivity and edge computing capabilities to large areas of terrain where there is no traditional network infrastructure. Given a suitable distance between UAVs, useful throughput values of at least 10 Mbps are observed, reaching up to 25 Mbps in specific situations. In addition to that, it is also observed that the communication channel between UAVs is unstable and that special attention must be paid to the distance between them.
The Distributed Sensing and Computing Over Sparse Environments (DISCOVER) testbed is a pioneering cyberinfrastructure initiative designed to advance research in distributed computing and networking tailored to rural, remote, and sparsely populated regions. Supported by the National Science Foundation (NSF), DISCOVER integrates a network of configurable Internet-of-Things (IoT) nodes—including stationary sensors, drones, and terrestrial rovers—across three key sites: Northern Arizona University (NAU), Clemson University, and Navajo Technical University (NTU). This collaboration offers a unique platform to explore innovative algorithms and methodologies addressing the technical challenges of under-served areas, with an emphasis on environmental and civil disaster response. The testbed enables a wide range of experiments, such as regional-scale data collection, heterogeneous networked services, distributed artificial intelligence (AI), distributed multi-robot control, and communication-aware software for resource-constrained networks. An online portal enhances accessibility, allowing researchers to request resources, upload experimental code, and retrieve data, with pre-integrated deep learning models for applications like human posture detection, object detection, and wildfire detection. This paper outlines the testbed’s architecture, operational sites, supported experiment types, ongoing research efforts, and its educational and outreach impacts, highlighting its role in fostering scientific innovation.
Joint Communication and Sensing (JCAS) is expected to play a critical role in next-generation wireless networks such as 6G. For complex sensing tasks, such as 3D pose estimation for virtual reality (VR) applications, accurate channel impulse response (CIR) or I/Q samples as well as processing using a neural network is required. Due to the higher bandwidth and antenna array sizes of future wireless networks, it is expected that offloading this data to a remote server for processing would require data rates in the order of 100s of Megabits per second, which is an unreasonable amount of overhead. Therefore it is necessary to preprocess the sensing data locally, and reduce the raw data to useful intermediary features, to mimimize the sensing data transmission overhead, especially when using multiple sensing devices. This paper proposes a method leveraging split inference to distribute neural networks across multiple devices, which achieves high accuracy while addressing the sensing data transfer bottleneck. We evaluate the performance of the proposed method in a VR gaming scenario, where mmWave Wi-Fi signals are used for 3D pose estimation. We show that split inference allows for reducing the communication overhead by three orders of magnitude compared to the centralised approach, while only losing 10% of accuracy. These results pave the way for future work, exploring highly distributed multi-static JCAS as a practical and efficient method of sensing.
The advent of virtual reality and immersive communication technologies is effecting a transformation in user experiences, enabling high-quality, low-latency interactions. In order to meet these requirements, edge caching, in particular for the transmission of virtual reality content, has become an efficient strategy for the mitigation of transmission latency and the decrease of backhaul traffic loads. This paper provides an introduction to a two-stage hybrid caching framework developed to manage the typical obstacles related to the caching of 360 degrees video. The proposed framework comprises two stages: a learning stage, which employs a Deep Q-Network to predict cache replacement actions, and a solving stage, which utilizes Integer Linear Programming to refine and optimize caching decisions. Furthermore, an L2 edge cache architecture is designed with the goal of enhancing cache utilization and further alleviating backhaul traffic. Performance evaluations illustrate that the proposed framework significantly enhances the cache hit ratio and reduces latency and backhaul usage compared to other methods.
The evolution of mobile networks towards more diverse services and open architectures has led to the emergence of mobile network function virtualization. This allows to match the reserved resources for network operation to the actual resources that are needed in the network at a certain place and time. Mobility-related network functions, specifically handover, can also be virtualized in this new paradigm. This virtualization can be facilitated by studying handover behavior at the base station level. In this work, using agglomerative hierarchical clustering, we show the existence of different base station profiles in terms of handovers, including three primary profiles: producer, receiver, and balanced. We also show that the use of these profiles, in addition to the dynamic reconfiguration enabled by virtualization, can reduce reserved resources by more than 50% compared to the current static system.
With the development of 5G, network slicing was proposed to enable service provisioning for diverse sets of Ultra-Reliable Low-Latency Communications (URLLC), enhanced Mobile Broadband (eMBB), and massive Machine-Type Communications (mMTC) users which are characterized by different Quality of Service (QoS) demands. Within network slicing, Radio Access Network (RAN) slicing plays a central role for efficient resource management. In addition, user admission control poses a major challenge. In this context, the problem of joint slice dimensioning and user admission control is investigated in this paper. To this end, an optimization problem based on symbol-level resource allocation with the objective of maximizing an operator's revenue while fulfilling the traffic requirements of all users is formulated. Afterward, the optimization problem is reduced to a knapsack problem and integrated into a Long-Term Revenue Maximization (LTRM) algorithm. Using data from real-world 5G measurements, the efficiency of the LTRM algorithm is verified, and the impact of various resource granularities in the time domain (symbol vs. slot) and frequency domain (varying Resource Block Group (RBG) sizes) is investigated. The revenue gain of the proposed joint algorithm over a sequential slice dimensioning and user admission control scheme is 24%, while symbol-level resource allocation offers at least 13% gain over a slot-based allocation for specific slices.
Sensing the temperature of liquids in containers is a critical process in various applications such as food safety, healthcare and environmental monitoring. Traditional approaches usually require a physical contact with the liquid, which may not be suitable for the liquids in sealed containers, and may also pose a risk for contamination in particular for the products used in healthcare. The contactless temperature sensing approaches that mainly rely on optical or laser based solutions, however, can only detect the temperature of a certain point or only the surface temperature of the liquid. Sensing through thermal cameras can provide scalability, however, they come with high costs. In this study, leveraging the ubiquitous availability of low-cost devices integrated with WiFi interfaces, we explore the feasibility of measuring the liquid temperature using fine-grained WiFi features. The proposed approach utilizes the amplitude variations across WiFi channel subcarriers extracted from Channel State Information (CSI) to detect the liquid temperature. In order to enhance the prediction accuracy, we also integrate principles from Newton’s Law of Cooling for regularization during the training of the neural network and propose a physics-guided machine learning framework (PMLF). Through our experiments with different liquids, temperature ranges and containers, we have demonstrated that the proposed cost-effective and scalable solution provides promising results in predicting the liquid temperatures.
While extensively studied in Global Positioning Systems, meaconing-i.e., the delay, amplification, and replay of a signal-is often regarded as impractical in cellular positioning systems due to the potential risk of communication disruption. We challenge this belief by experimentally validating full-frame meaconing attacks on 5G systems. Using off-the-shelf hardware, we demonstrate how an attacker can replay entire 5G frames, introducing o(mu s) controlled TOA biases while maintaining uninterrupted communication. Our findings reveal the real world viability of these attacks, highlighting the urgent need for robust countermeasures to protect 5G localization systems.
This paper introduces the first implementation of physical layer key generation (PLKG) on a real-world unmanned aerial vehicle (UAV) platform. To tackle the unique challenges of high mobility and dynamic communication environments in Internet of Drones (IoD) networks, we propose a novel multimodal learning framework for enhancing PLKG in UAV-to-ground communications. Static channel state information (CSI) features and dynamic UAV trajectory data are extracted using convolutional neural networks (CNN) and long short-term memory (LSTM) networks, respectively. Leveraging conditional embedding techniques, the predicted UAV position and velocity are integrated into the input space of the key reconstruction network as conditional features. The trained network serves as a Doppler-resilient feature extraction mapping function, thus achieving robust and consistent key generation under varying mobility conditions. Experimental results demonstrate lower key mismatch rates and higher reliability compared to existing CSI-based key generation methods.
Wildfires pose a serious hazard to both life and the environment, necessitating accurate and timely detection methods to minimize their impact. However, traditional approaches have limitations in precision, response time, and adaptability to environmental factors. To address this gap, machine learning (ML) and deep learning (DL) approaches are increasingly being adopted for fire detection. This study proposes a hybrid approach that combines deep feature extraction using Transfer Learning (TL) with EfficientNetB0 and traditional Gradient Boosting Models (GBM), specifically LightGBM, XGBoost, and CatBoost, for wildfire detection using UAV imagery. EfficientNetB0 is used to extract features, which are then dimensionally reduced using Principal Component Analysis (PCA). Optuna framework is used for hyperparameter tuning, to maximize the performance of the model. The results show that the GBM models enhanced with TL outperform their baseline counterparts, with CatBoost + TL achieving the highest precision (92. 37%). These findings confirm the role of TL in improving model precision, efficiency, and computational resource utilization, making it ideal for real-time wildfire detection.
In this paper, we present MAGIC, a novel approach to gesture recognition utilizing mmWave multiple-input multiple-output (MIMO) Channel State Information (CSI). Unlike existing mmWave gesture recognition methods that often rely on radar signals, MAGIC leverages CSI extracted from mmWave MIMO integrated sensing and communication (ISAC) systems. While advanced radar systems, such as those operating in frequency-modulated continuous wave (FMCW) mode, can achieve high frequency and spatial resolution, they typically require dedicated sensing infrastructure, which increases system complexity. In contrast, MAGIC utilizes high-granular CSI from orthogonal frequency-division multiplexing (OFDM) systems, enabling fine spatial, temporal, and frequency-domain information for robust gesture recognition. This eliminates the need for dedicated radar transceivers, simplifying the system and reducing transmission overhead. MAGIC employs a learning-based architecture, integrating a temporal convolutional network (TCN) to classify gestures by capturing long-range temporal dependencies. To address the critical challenge of domain adaptation in gesture recognition, we propose adaptive temporal embedding network (ATEN), a meta-learning framework that combines the temporal modeling capabilities of TCN with task-specific adaptation mechanisms. We evaluate MAGIC through a comprehensive data collection campaign involving two subjects performing 10 micro gestures across three different environments, with synchronized video streams providing the ground truth. The proposed system achieves a baseline accuracy of 99.24% using TCN. The system continues to perform well achieving up to 98.82% accuracy - when adapting to new domains using ATEN, outperforming other state-of-the-art domain adaptation methods by 14% on average.
As 5G networks evolve to meet the growing demands for high data rates, low latency, and superior quality of service, hardware acceleration becomes essential for managing computationally intensive tasks. This paper demonstrates an implementation of Orthogonal Frequency Division Multiplexing (OFDM) symbols (both downlink and uplink transmission) using Field Programmable Gate Arrays (FPGAs), offloading the execution of these critical Low-Physical Layer (Low-PHY) functions away from the general-purpose CPU approach found in the open-source mobile software platform OpenAirInterface (OAI). The full integration of FPGA-generated OFDM symbols with an Open Computing Language (OpenCL) framework and the OAI protocol stack yields significant performance improvements, including a a 90% processing time reduction compared to the traditional CPU-based implementation. This demonstration highlights the effectiveness of FPGA acceleration in enhancing the efficiency of 5G Radio Access Network (RAN) functions implemented by open-source efforts, thus offering a path forward to implementing additional Low-PHY functions on accelerated hardware.
