Open Positions

The field of application of the theses below concerns the design and optimization of AI techniques for 5G/6G and satellite networks.

• Spiking Neural Networks for Intrusion Detection in 5G/6G Networks The rapid growth of connected devices and the advent of 5G/6G networks introduce significant security challenges, exposing base stations to increasingly sophisticated cyber-attacks. While traditional deep learning approaches achieve strong performance, their computational and energy demands are incompatible with the resource constraints of next-generation network nodes. This thesis investigates Spiking Neural Networks (SNNs), a biologically inspired and event-driven paradigm that encodes information as sparse temporal spike trains and offers substantial gains in energy efficiency over conventional artificial neural networks, for efficient intrusion detection and classification. Using the fully labeled 5G-NIDD dataset collected on a real 5G testbed, SNN-based models will be trained and evaluated, and benchmarked against conventional approaches in terms of detection accuracy, convergence speed, and energy consumption.
• Fingerprinting of Federated Learning Models via Network Metadata Federated Learning enables distributed clients, such as user equipment (UEs) in 5G/6G networks, to collaboratively train machine learning models by exchanging model updates rather than raw data. While these updates may be encrypted, the associated network traffic metadata—e.g., packet sizes, uplink/downlink ratios, burst patterns, and interarrival times—can still expose distinctive signatures of the underlying training process. This thesis investigates whether a passive network observer can infer the model family or training configuration solely from traffic traces, without access to payload content. The study will build traffic-level representations of FL rounds (statistical features and burst-level sequences) and train classifiers to identify training workloads under realistic variability (heterogeneous clients, fluctuating bandwidth/RTT, background traffic).
• Graph-Based Modeling of Federated Learning with GNN-Driven Client Selection Federated Learning systems involve a large number of distributed clients (e.g., User Equipments in 5G/6G networks) with heterogeneous computational capabilities, data distributions, and connectivity conditions. An efficient client selection is crucial to ensure fast convergence, fairness, and resource-aware training. This thesis proposes to represent the Federated Learning process as a graph, where each node corresponds to a client and edges encode relationships such as statistical similarity (data distribution), communication latency, channel conditions, or past contributions to the global model. Building upon this graph representation, Graph Neural Networks (GNNs) will be leveraged to learn an adaptive client selection strategy that enhances training efficiency while maintaining high model accuracy.
• Why Should You Join Federated Learning? Federated Learning only works if different organizations agree to collaborate. This thesis studies how to design incentives — such as rewards, reputation systems, or fair contribution rules — that motivate participants to join and stay in the learning process. You will simulate different strategies and determine the ones leading to fair and stable collaboration.
• Using LEO Satellite Links to Sense the Weather: Low Earth Orbit (LEO) satellite constellations—such as Starlink and OneWeb—continuously exchange signals with ground stations, with these transmissions passing through the Earth’s atmosphere. Changes in weather conditions, such as heavy rainfall or high humidity, can cause variations in signal strength and delay. This thesis investigates how to use such variations as weather indicators. By leveraging AI-based models for data fusion and pattern recognition, the study aims to transform satellite communication networks into large-scale, intelligent environmental sensing systems.