AI4C6 – AI/ML Solutions for Communications
Thursday, 4 June 2026, 16:30-18:00, room Sala 4 (1st floor)
Session Chair: Sergio Fortes (Univ. Málaga, ES)
PyOMNeT-AI: Enabling Efficient OMNeT++ Network Optimization Through Low-Overhead, Hybrid Python AI Integration
Moustafa Roshdi (Fraunhofer IIS & Friedrich Alexander University (FAU), Germany); Arindam Chakraborty, Amir Amri and Sahana Raghunandan (Fraunhofer IIS, Germany); Reinhard German (Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany)
Integrating AI/ML into OMNeT++ based network simulations is increasingly important for studying and optimizing adaptive network behaviors under realistic discrete-event dynamics. A prominent example is 5G-Advanced and emerging 6G RAN optimization, where learned policies can support scheduling, mobility management, and resource control. However, integrating Python-native AI/ML workflows with OMNeT++ remains challenging: C++-only integrations limit model agility, while socket-based RL coupling introduces substantial serialization and process coordination overhead. This paper presents PyOMNeT-AI, the first comprehensive framework for Python-native AI integration in OMNeT++, providing three complementary patterns under unified design principles. PyCoSim enables bidirectional co-simulation with Python modules as first-class simulation components for message-driven AI logic. PyEmbed provides scoped interpreter embedding for efficient inference-centric workloads, achieving 0.28 𝜇s baseline overhead and sustaining over 3 million calls per second. ShmGymBridge offers high-throughput shared memory coupling for Gym-compatible RL training, providing 2-6× speedup over socket-based alternatives depending on workload and payload size. We present the design rationale, performance characterization across integration patterns, and selection guidelines validated through Simu5G-based RAN optimization studies, with applicability to broader OMNeT++ domains including IoT, vehicular networks, and smart grids.
Hybrid Deep Learning Framework for CSI-Based Activity Recognition in Bandwidth-Constrained Wi-Fi Sensing
Alison Michel Fernandes (UTFPR – Universidade Tecnológica Federal Do Paraná, Brazil); Hermes I Del Monego (UTFPR, Brazil); Bruno Chang (Federal University of Technology – Paraná, Brazil); Anelise Munaretto (UTFPR, Brazil); Helder Fontes (INESC TEC and FEUP, Portugal); Rui Campos (INESC TEC and Faculty of Engineering, University of Porto, Portugal)
This paper presents a novel hybrid deep learning framework designed to enhance the robustness of CSI-based Human Activity Recognition (HAR) within bandwidth-constrained Wi-Fi sensing environments. The core of our proposed methodology is a preliminary Doppler trace extraction stage, implemented to amplify salient motion-related signal features before classification. Subsequently, these enhanced inputs are processed by a hybrid neural architecture, which integrates Inception networks responsible for hierarchical spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks that capture temporal dependencies. A Support Vector Machine (SVM) is then utilized as the final classification layer to optimize decision boundaries. The framework’s efficacy was systematically validated using a public dataset across 20, 40, and 80 MHz bandwidth configurations. The model yielded accuracies of 89.27% (20 MHz), 94.13% (40 MHz), and 95.30% (80 MHz), respectively. These results confirm a marked superiority over standalone deep learning baselines, especially in the most constrained low-bandwidth scenarios. This study underscores the utility of combining Doppler-based feature engineering with a hybrid learning architecture for reliable HAR in bandwidth-limited wireless sensing applications.
Near Real Time Dynamic Backdoor Attacks on Data-Driven O-RAN xApps
Mikolaj Borys (K.U. Leuven, Belgium); Davy Preuveneers (DistriNet, KU Leuven, Belgium); Rafael Cavalcanti (Keysight Technologies, Belgium)
ML-based xApps are rapidly proliferating in O-RAN to manage the growing complexity of 5G networks. These xApps support diverse use cases such as traffic steering, mobility management, interference mitigation, and anomaly detection. Because ML models directly influence decisions, they themselves become targets, enabling adversaries to manipulate xApp behavior. We provide an overview and analysis of major attack classes on such models, arguing that existing state-of-the-art attack techniques in their unmodified form are ill-suited to the real-time, architectural and data structural constraints of an xApp’s domain. We then introduce a novel black-box attack framework grounded in a threat model informed by prior O-RAN security analyses and known vulnerabilities. The framework carries out a backdoor attack that allows to trigger attacker-desired behavior by adding an iteratively optimized, sample-specific perturbation to the training set. In parallel, a trigger injector neural network is trained to compute and allow an attacker to apply the learned triggers using UEs connected to the network in real time during deployment of the ML model. We test the approach on an anomaly-detection xApp. Despite the domain constraints that make the creation of these triggers non-trivial, the method can introduce a structurally valid trigger injectable by malicious UEs that reduces model accuracy by 40% in real-time, while the poisoned model still maintains its validation performance. It significantly outperforms prior backdoor and other attack types in effectiveness, stealthiness, and timeliness, succeeding where most other methods are inapplicable.
Semantic Communications with RIC-Centric AI-Driven Orchestration for 6G Networks
João Henrique Henrique Inacio de Souza (Aalborg University, Denmark); Mattia Merluzzi (CEA-Leti, France); Nicola Cordeschi (CNIT, Italy); Mateus Pontes Mota (CEA-Leti, France); Lanfranco Zanzi (NEC Laboratories Europe, Germany); Peizheng Li (Toshiba Europe Ltd., United Kingdom (Great Britain) & Bristol Research & Innovation Laboratory, Toshiba, United Kingdom (Great Britain)); Xinyi Lin (Toshiba Europe Ltd., United Kingdom (Great Britain)); Zihan Chen and Tony Q. S. Quek (Singapore University of Technology and Design, Singapore); Olivier Forceville (HPE FRANCE, France); Idir Djouab (HPE France, France); Roberto Fantini (Telecom Italia SpA, Italy); Paolo Di Lorenzo (CNIT, Italy)
Semantic and goal-oriented communications are emerging as a key paradigm for future 6G networks. However, many existing approaches focus on semantic models and algorithms in isolation, often overlooking the architectural, operational, and time-scale constraints of real network deployments. In this paper, we address this gap by presenting a RIC-centric framework for semantic communications grounded in the O-RAN architecture, enabling the coordinated realization of semantic intelligence across heterogeneous control layers and time scales. Building on a unified semantic management framework spanning the Non-Real-Time and Near-Real-Time RIC, we show how semantic objectives, AI-based orchestration, and adaptive L1/L2 control can be jointly integrated while accounting for energy efficiency, computation costs, and deployment realism. The framework is evaluated through three representative case studies covering semantic-aware MAC operation for IoT data collection, collaborative edge intelligence with goal-aware inference, and AI-driven RAN control over a commercial-grade O-RAN-compliant network. The results demonstrate how semantic communications can be effectively instantiated across different layers and time scales, providing actionable insights toward deployable, AI-native 6G systems.
Evaluating LLM-Driven Sensitive Data Redaction for Privacy Preservation in O-RAN Digital Twins
Dinushika Chithrani Pasyala Acharige, Madhusanka Liyanage and Liam Murphy (University College Dublin, Ireland)
Open Radio Access Networks (O-RAN) introduce open and programmable architectures that enable data-driven intelligence across radio access networks. O-RAN Digital Twins (DTs) further support this vision by providing virtualized environments for testing, and validating machine-learning based network functions. However, the collection and sharing of such data between the physical network and its DT raise significant privacy risks, particularly when sensitive user, device, and operational information is exposed. As DTs are typically hosted on external edge or cloud servers, preserving privacy while maintaining data utility for analytics and network management applications remains a critical challenge. “LLM Twin-Shield”, our proposed Large Language Model (LLM)- based privacy-filtering approach, automatically identifies and filters highly sensitive information in O-RAN data. By leveraging LLMs’ contextual understanding capabilities, the proposed method performs semantic-sensitive data detection. The effectiveness of our approach is demonstrated in an intrusion detection scenario, where privacy-sensitive data can be redacted before sharing with the DT and subsequently reconstructed with high analytical accuracy. Our findings highlight the potential of LLM-driven privacy preservation as a scalable solution for AI-driven applications in O-RAN environments.
