Integrating non-terrestrial and terrestrial networks (NTN-TN) is essential to provide ubiquitous coverage in beyond-5G systems. This demo showcases a cloud-native testbed where an NTN UE and a TN UE communicate through a unified 5G core over an emulated geostationary satellite. The setup combines SDR hardware, realistic channel emulation, and automated lifecycle management of mobile network functions using ETSI NFV MANO and open-source platforms.
Accurate signal matching across spatially distributed receivers is essential for applications including Time Difference of Arrival (TDOA)-based localization and collaborative sensing. A key challenge arises in data transmission when signal structure and modulation are unknown, often necessitating the transmission of raw I/Q data, which is often infeasible due to bandwidth constraints, as traditional decoding-based methods fail. This paper investigates compression techniques that retain matching capability while significantly reducing data volume. We evaluate classical lossy methods (sample resolution reduction and downsampling) alongside a deep learning approach using a Siamese network to learn compact embeddings. Using real-world Mode S radar signals, we assess matching accuracy and compression ratio. Results show that moderate downsampling improves performance by suppressing noise, and that learned embeddings enable compression ratios two orders of magnitude higher than classical approaches with minimal accuracy loss.
Semantic communication, particularly Deep Joint Source-Channel Coding (DJSCC), has emerged as a promising solution for efficient image transmission in future communication systems. However, the broadcast nature of wireless communication and correlation between the compressed and original data poses significant security risks, particularly in the form of Model Inversion Eavesdropping Attacks (MIEAs). Existing defense mechanisms suffer from limitations, such as high computational overhead and information loss due to quantization and adversarial training. As a result, these methods are unsuitable for applications that demand fast, secure, and reliable information transmission. In this paper, we propose a secure, lightweight, and reversible framework named Key-Assisted Protection for Deep Joint Source-Channel Coding (KAP-DJSCC), which utilizes a Diffie-Hellman (DH)-based key exchange to construct a shared secret transformation matrix. This pluggable process is applied to obscure the latent representation of the input without altering its structure or requiring retraining. Additionally, we introduce a novel MIEA variant, Key-assisted Model Inversion Eavesdropping Attack (KMIEA), in which the attacker attempts to guess the key. Simulation results confirm that KAP-DJSCC significantly degrades the eavesdropper’s reconstruction performance in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Perceptual Image Patch Similarity (LPIPS) while preserving high fidelity for the legitimate receiver across varying Signal-to-Noise Ratio (SNR) levels.
This study describes our experience in detecting network scanning of IoT devices through energy consumption analysis. We carried out a set of experiments where five benign activities (web requests, updates, file transfers, audio streaming, and video streaming) and one malicious activity (Nmap network scanning) were considered. Energy traces were collected using a high-accuracy power monitor. Machine learning classification using a Neural Network achieved 97.14% accuracy in identifying the malicious scanning activity. Results demonstrate that energy-based detection can effectively distinguish between normal operations and cyberattacks, offering a promising, passive security approach for IoT devices.
Widely recognized as a candidate technology for 6G, joint communication and sensing (JCAS) is expected to enable a range of emerging applications in wireless networks. However, incorporating radio sensing capabilities introduces new design challenges and demands novel approaches across various research and engineering disciplines. This paper studies the waveform classification task not only for dedicated communication and radar but also for fully integrated JCAS waveforms. To this end, we generate wireless waveforms that belong to eight classes, including four pure communications standard signal waveforms, two classical radar waveforms, and two fully integrated JCAS candidate waveforms. We examine the dataset’s statistical properties and evaluate several machine learning and deep learning classifiers based on expert feature extraction and raw data. Numerical simulations show that expert feature-based classifiers exhibit limited accuracy for the considered task, even for a very high signal-to-noise ratio (SNR). The classifier that learns from raw data yields high accuracy, but requires either a high-performance computing platform or further architectural optimization to meet real-time requirements. Additionally, we investigate the complexity, inference time, and memory footprint of the designed classifiers in the context of their practical implementation in wireless communication systems.
