AI4C7 – AI/ML Solutions for Communications
Friday, 5 June 2026, 9:00-10:30, room Sala 2 (1st floor)
Session Chair: Tommy Svensson (Chalmers Univ. Technology, SE)
Reliable Spectrum Sensing for Future Wireless Networks: A CFAR-Unified Study of Classical and ML Approaches
Riya Deshpande, Faheem A. Khan and Qasim Zeeshan Ahmed (University of Huddersfield, United Kingdom (Great Britain))
Reliable spectrum sensing is essential for Cognitive Radio (CR) systems, enabling Secondary Users (SUs) to achieve opportunistic spectrum access while protecting Primary Users (PUs) from interference. Conventional static spectrum sensing techniques, such as Energy Detection (ED) and Matched Filter Detection (MFD), remain widely used, but their performance degrades under low Signal-to-Noise Ratio (SNR) conditions or when strict requirements for prior Channel State Information (CSI) exist. Although Machine Learning (ML) and Deep Learning (DL) approaches have recently been explored as alternatives, existing studies often suffer from inconsistent datasets, mismatched assumptions, or the absence of strict false-alarm control, making fair comparison difficult. This paper proposes a unified and Constant False Alarm Rate (CFAR) controlled evaluation framework to systematically compare conventional sensing techniques (ED, MFD) with supervised ML techniques, such as Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Convolutional Neural Network (CNN). All detectors operate on identical In-phase and Quadrature (IQ) sample inputs derived from Quadrature Phase Shift Keying (QPSK) modulated pilot signals, and thresholds are calibrated exclusively from noise-only samples to ensure a CFAR across all methods. The results of the simulations demonstrate that MFD achieves the best overall performance when the pilot information is known, validating its theoretical optimality. Furthermore, the CNN detector significantly outperforms conventional ML models at low and moderate SNRs and approaches MFD performance above -5 dB, indicating that CNN implicitly learns correlation-like structures similar to MFD.
Towards 6G Intelligent MEC: Context-Aware DRL-Based Task Assignment with Adaptive Rewards
Mona Mohammed Alghamdi (King Abdulaziz University, Saudi Arabia & Queen Mary University of London, United Kingdom (Great Britain)); Atm Shafiul Alam (Queen Mary University of London, United Kingdom (Great Britain)); Asma Cherif (King Abdulaziz University, Saudi Arabia)
The rapid evolution of mobile ecosystems toward 6G and Next-Generation Networks (NGN) is driving a shift toward AI-native architectures and pervasive edge intelligence. This transition introduces increasing demand for computation-intensive and highly heterogeneous workloads, such as Holographic Telepresence and Industrial Digital Twins, placing significant burdens on resource-constrained devices. Mobile Edge Computing (MEC) addresses this challenge by extending computational intelligence to the network edge. However, optimal task assignment in 6G-enabled MEC systems remains challenging due to highly stochastic network conditions, heterogeneous workloads, high-dimensional state spaces, and the non-stationary service requirements. While Deep Reinforcement Learning (DRL) offers a promising framework for adaptive task assignment, conventional approaches often struggle with reward sparsity and limited context awareness under varying task requirements. This paper proposes a context-aware adaptive reward shaping strategy for delay-aware task assignment in heterogeneous 6G-enabled MEC environments. By formulating the problem as a Markov Decision Process and employing a Double Deep Q-Network (DDQN), the proposed approach constructs context-aware performance baselines that enable reward normalization and stable learning under dynamic conditions. Experimental results demonstrate improved learning stability and faster convergence, with service latency reductions of up to 34% compared to DRL baselines and 33% and 58% over local-edge and cloud execution, respectively, for heterogeneous workloads.
ProtoAoA: Prototypical Networks for Few-Shot Angle-of-Arrival Estimation
Elsayed Mohammed and Omar Mashaal (University of Calgary, Canada); Alec Digby, Pasquale Leone and Lorne Swersky (Qoherent Inc, Canada); Ashkan Eshaghbeigi (Qoherent inc, Canada); Hatem Abou-Zeid (University of Calgary, Canada)
Angle-of-arrival (AoA) estimation is a crucial function in wireless communications used for localization, beamforming, interference management, and other applications. Deep learning (DL) solutions have been proposed for AoA to mitigate limitations of traditional AoA estimation techniques such as sensitivity to noise and the inability to generalize across different array characteristics. A challenge, however, of DL-based approaches is their reliance on large data collection campaigns and model training. This paper proposes the application of PrototypicalNetworks (PN) to address this challenge and utilizes a real-world dataset collected on a software defined radio (SDR) testbed to validate the effectiveness of the proposed solution. Prototypical Networks excel in extracting representative embeddings from unstructured input data, establishing class prototypes during training that can be few-shot trained on unseen classes. We demonstrate the efficacy of PNs for AoA classification using complex IQ samples, focusing on its ability to correctly classify new, unseen angles that the model was not trained on previously. Our results show that training our proposed ProtoAoA on only 23% of the AoA dataset classes can attain a mean absolute error (MAE) of 3◦ with only 4-shots of training on the unseen angles − and an MAE of 2◦ with 32-shots of training data. These results demonstrate that the developed prototypical network architecture requires remarkably few data samples to achieve reliable AoA estimation − and highlights its potential for other wireless applications where data availability is limited.
A Transformer Inspired AI-Based MIMO Receiver
Andras Racz, Tamás Borsos and Andras Veres (Ericsson Research, Hungary); Benedek Csala (Budapest University of Technology and Economics Budapest, Hungary)
We present AttDet, a Transformer-inspired MIMO (Multiple Input Multiple Output) detection method that treats each transmit layer as a token and learns inter-stream interference via a radio-aware self-attention mechanism. Queries and keys are derived directly from the estimated channel matrix, so attention scores quantify channel correlation. Values are initialized by matched-filter outputs and iteratively refined. The AttDet design combines model-based interpretability with data-driven flexibility. We demonstrate, through link-level simulations under realistic 5G channel models and high-order mixed QAM modulation and coding schemes, that AttDet approaches near-optimal BER/BLER (Bit Error Rate/Block Error Rate) performance under practical SNR (Signal-to-Noise Ratio) and MIMO operating conditions.
Fair Federated Learning Under Agnostic Contention Diversity in Disaggregated QoS-Aware Networks
Tania Panayiotou (University of Cyprus, Cyprus); Saroj Kumar Panda (LTIMindtree, India); Sadananda Behera (NIT Rourkela, India); Georgios Ellinas (University of Cyprus & KIOS Research and Innovation Center of Excellence, Cyprus)
Federated learning (FL), that keeps data isolated at local clients, has emerged as an effective and promising solution for mitigating privacy concerns in disaggregated networks where multiple operators independently manage different segments. Ensuring fairness in traffic prediction and resource allocation under heterogeneous traffic datasets remains, however, challenging. Fair FL, through the optimization of the q-fair FL (q-FFL) function, seeks to balance accuracy across clients by tuning a single inequality aversion parameter, q. However, modeling contention with a single parameter implicitly assumes contention uniformity across all datasets; in practice, heterogeneity in size, quality, distribution, and participation frequency may lead to varying levels of contention amongst different dataset groups. Therefore, this work proposes Q-fair FL (Q-FFL), a novel optimization function parametrized by a set of q values, enabling clients to collaboratively tune their individual q parameters. Since this requires exploration of a computationally prohibitive parameter space, a metaheuristic-guided Q-FFL optimization framework is introduced, which effectively reduces the search space (by 99%) while also providing a flexible and cost-efficient framework for collaborative networks to best meet their targeted quality-of-service (QoS) criteria. It is demonstrated that Q-FFL enables further improvement in fairness (by 25%) compared to q-FFL, with the use of datasets based on real network traffic traces.
