AI4C2 – AI/ML Solutions for Communications
Wednesday, 3 June 2026, 17:00-18:30, room Sala de Conferencias 2 (2.2) (1st floor)
Session Chair: Riccardo Trivisonno (Huawei Technologies Duesseldorf GMBH, DE)
Data-Efficient Domain Adaptation for Receiver-Invariant Radio Frequency Fingerprinting Identification
Cem Ayyıldız (GOHM Electronics, Turkey); Fatih Emre Yıldız (GOHM Electronics & Muğla Sıtkı Koçman Üniversitesi, Turkey); Barış Ethem Süzek (Muğla Sıtkı Koçman Üniversitesi, Turkey)
Lightweight authentication that avoids excessive cryptographic overhead is essential for the security of future 6G wireless networks, particularly in massive IoT deployments. Radio Frequency Fingerprinting Identification (RFFI) uses hardware impairments for physical-layer device authentication, but receiver replacement introduces domain shift that can severely degrade identification accuracy. This work addresses three practical deployment questions: whether unsupervised domain adaptation can recover identification performance after receiver replacement, the minimum amount of unlabeled target data required for effective adaptation, and the optimal strategy for sequential receiver replacements. We evaluate adversarial discriminative domain adaptation on a synchronized experimental testbed comprising 30 identical transmitters and 3 receivers. Results show that adaptation restores high identification performance, achieving F1 classification scores of 89 to 92 percent for same-model receivers and 70 to 74 percent for heterogeneous receivers. This performance gap indicates a hardware ceiling imposed by receiver architecture differences. Adaptation achieves practical convergence with only 50 to 100 unlabeled packets per transmitter class, enabling energy-efficient updates by reducing transmission and computation overhead. For sequential receiver replacements, adapting directly from the original source model outperforms adapting incrementally through each previous receiver in most scenarios, providing guidance for multi-stage receiver replacements. The complete dataset has been released publicly to support future research on receiver-invariant RFFI.
HELENA: High-Efficiency Learning-Based Channel Estimation Using Dual Neural Attention
Miguel Camelo Botero (University of Antwerp – imec, Belgium); Esra Aycan Beyazıt (IDLab, Imec- University of Antwerp, Belgium); Nina Slamnik-Krijestorac (University of Antwerp-IMEC, Belgium); Johann Marquez-Barja (University of Antwerpen & IMEC, Belgium)
Accurate channel estimation is critical for high-performance Orthogonal Frequency-Division Multiplexing (OFDM) systems, particularly at low signal-to-noise ratios and under stringent latency constraints. This article presents High-Efficiency Learning-based channel Estimation using dual Neural Attention (HELENA), a compact deep learning model that combines a lightweight convolutional backbone with two efficient attention mechanisms: patch-wise multi-head self-attention for capturing global dependencies and a squeeze-and-excitation block for local feature refinement. Compared to CEViT, a state-of-the-art vision transformer-based estimator, HELENA reduces inference time by 45.0% (0.175ms vs. 0.318ms), achieves comparable accuracy (-16.78dB vs. -17.30dB), and requires 8x fewer parameters (0.11M vs.0.88M), demonstrating its suitability for low-latency, real-time OFDM-based wireless systems.
DiTraPos: Learned Channel Tokens for Distributed Transformer-Based Positioning
Florian Langenstein, Till Ruppert and Mohammad Asif Habibi (German Research Center for Artificial Intelligence, Germany); Dennis Salzmann (German Research Center for Artificial Intelligence DFKI GmbH, Germany); Florian Herrmann (Deutsches Forschungszentrum Für Künstliche Intelligenz, Germany); Christoph Fischer (German Research Center for Artificial Intelligence, Germany); Hans D. Schotten (RPTU Kaiserslautern-Landau, Germany)
Distributed TraPos (DiTraPos) extends our earlier work, Transformer-based Positioning (TraPos) [1], toward distributed inference for disaggregated 6G radio access networks (RANs) and industrial deployments. Instead of centrally processing high-dimensional multi-antenna channel measurements, DiTraPos splits inference into an edge-side, per-antenna encoder that forms tokenized channel representations and a remote-side global Transformer that performs cross-channel fusion. A learned tokenwise compression/decompression interface constitutes the edge-cloud boundary and communicates compact channel tokens rather than raw observations, providing explicit control over fronthaul payload while preserving frequency resolution at the token level. The architecture reuses the core TraPos components and extends the conditional-computation motif throughout the pipeline, including compression stages and the final regression head. We evaluate DiTraPos on real 5G sounding reference signal (SRS) measurements in an industrial-like indoor environment with motion-capture ground truth and eight phase-coherent receive channels. Across angle of arrival (AOA), time difference of arrival (TDOA)-derived distance, and direct position regression, the error distributions of DiTraPos closely match those of TraPos, indicating that distributed inference with tokenized compression preserves predictive accuracy. At the same time, both approaches consistently outperform a 2D Multiple Signal Classification (MUSIC) baseline, while DiTraPos enables reduced fronthaul payloads and a more generalizable per-antenna model through shared network components.
SynthBGP: Synthetic BGP Traffic Generation for Enhanced Cybersecurity Anomaly Detection
Shadi Motaali and Jorge E. López de Vergara (Universidad Autónoma de Madrid, Spain); Luis de Pedro (Universidad Autonoma de Madrid, Spain); Ivan Gonzalez (Universidad Autónoma de Madrid, Spain)
Border Gateway Protocol (BGP) anomalies cause large-scale Internet disruptions, yet the extreme scarcity of labeled training data constrains machine learning-based detection methods. Existing approaches either ignore protocol semantics or require prohibitive expert effort. This paper systematically evaluates synthetic BGP traffic generation for anomaly detection. We propose a consensus-based multi-level labeling pipeline combining five unsupervised detectors to create high-confidence labels from RIPE collector streams, and benchmark thirteen generators from five families (rule-based, deep generative, oversampling, hybrid, statistical) using a protocol-aware, 16-metric fidelity framework. SMOTE_kmeans achieves 97.5% in-distribution fidelity and preserves detector performance (F1>0.99), but cross-collector evaluation reveals significant limitations: F1 scores degrade to 0.79, with false negative rates of 35-41% on unseen anomalies. GAN neural generators degrade further under a distribution shift.
Online Learning for Predictive Satellite Telemetry Analysis: Towards a Digital Twin for Constellation Management
Arnau Dolz Puig (Space Communications Research Group, Fundació i2CAT, Internet i Innovació Digital a Catalunya, Spain & Universitat Politècnica de Catalunya, Spain); Adrian Perez-Portero (i2CAT Foundation, Spain); Juan A. Fraire (Inria/INSA Lyon & CONICET, National University of Córdoba, France); Anna Calveras (Universidad Politecnica de Catalunya (UPC), Spain)
The deployment of Low Earth Orbit (LEO) satellite constellations is transforming global connectivity, yet their management remains challenging due to dynamic topologies and intermittent connectivity inherent in Non Terrestrial Networks (NTNs). Traditional offline approaches fail to capture real-time operational changes, and batch machine learning methods cannot adapt to the continuous, evolving nature of satellite telemetry. To address this, we propose and validate an Online Machine Learning (OML) framework for predictive telemetry analysis, designed as the analytics engine of a broader Digital Twin (DT) architecture that integrates the Distributed Satellite Simulator (DSS-SIM) for orbital and network modeling. We develop a comprehensive statistical methodology that combines Principal Component Analysis (PCA) based clustering, Granger causality, and change-point detection to identify optimal predictor variables from high-dimensional telemetry streams. Validated on real 6GStarLab (6GSL) mission data, the framework identifies Attitude and Determination Control System (ADCS) variables as precursors to battery voltage changes, achieving 82-minute early warning before a confirmed power anomaly event and a Mean Absolute Error (MAE) of 0.04% through prequential evaluation on 401K predictions. These results demonstrate the operational viability of online learning for real-time satellite state prediction, providing a foundation for proactive NTN resource management in sparse LEO constellations.
