MSc thesis proposals

The following MSc thesis proposals are currently open. Students interested in one of them can get in touch through the Contacts page.

Describing AI Functions to the Network: A Machine-Readable Operational Model for Inference, Learning and Agentic Functions at the Edge

AI functions are moving out of the cloud onto edge servers, far-edge devices and network equipment, where the infrastructure has to place, configure, update and interconnect them as it does with any other network function. To do so, it must know what each function requires and offers. Today every AI component describes itself in its own way: model cards document provenance and intended use, agent cards list skills and endpoints, 3GPP defines an “ML entity” for 5G management, and the W3C Thing Description models the interaction affordances of IoT devices. None of them captures, in one implementation-independent form, the inputs and outputs of a function, its resource needs, its lifecycle states and the conditions under which it can be operated and connected.

The thesis builds a first version of such an operational model bottom-up, from real functions rather than from a designed vocabulary. The student will: (i) compare existing descriptions and formulate the competency questions the model must answer (e.g., which input formats a function accepts, what it needs from the network to cooperate with another function); (ii) extract operational properties from the documentation, interfaces and runtime behaviour of functions deployed in our edge testbed, such as an image detector on a far-edge camera, a telemetry anomaly detector, a federated-learning client and a coordination agent; (iii) formalise the properties in a machine-readable data model and implement a reference service that exposes them at runtime; (iv) test the model by comparing declared and observed behaviour, for instance whether a function accepts the declared input types or whether its exposed state and model version change after a model update.

The result is a data model with a reference implementation, and evidence of which properties are common to all functions and which need function-specific extensions.

Keywords: AI functions; Edge computing; Data modelling; Network management; Interoperability

Duration: 9 months

Prerequisites: Computer networks; programming in Python; basics of machine learning. Interest in data modelling (e.g., YANG, JSON Schema) and software engineering.

Suggested reading

  1. M. Mitchell et al., “Model Cards for Model Reporting,” in Proc. ACM Conference on Fairness, Accountability, and Transparency, 2019.
  2. 3GPP TS 28.105, “Management and orchestration; Artificial Intelligence/Machine Learning (AI/ML) management.”
  3. ITU-T Rec. Y.3172, “Architectural framework for machine learning in future networks including IMT-2020,” 2019.
  4. W3C, “Web of Things (WoT) Thing Description 1.1,” W3C Recommendation, Dec. 2023.
  5. A2A Project (Linux Foundation), “Agent2Agent (A2A) Protocol Specification” (Agent Card), 2025.
  6. ETSI GS NFV-IFA 011, “Network Functions Virtualisation (NFV); Management and Orchestration; VNF Descriptor and Packaging Specification.”
  7. M. Grüninger and M. S. Fox, “Methodology for the Design and Evaluation of Ontologies,” in Proc. IJCAI Workshop on Basic Ontological Issues in Knowledge Sharing, 1995.
  8. G. C. Publio et al., “ML-Schema: Exposing the Semantics of Machine Learning with Schemas and Ontologies,” arXiv:1807.05351, 2018.
  9. M. Bjorklund (Ed.), “The YANG 1.1 Data Modeling Language,” RFC 7950, 2016.

Verifiable Lifecycle Management of AI Functions across Heterogeneous Orchestrators: Management Contracts and Conformance Testing on Kubernetes and NETCONF/YANG

AI functions deployed across the cloud-edge continuum are managed by very different systems: container orchestrators such as Kubernetes with KServe, NFV orchestrators such as ETSI OSM, and SDN controllers based on NETCONF/YANG. Each platform defines what deploy, activate, update or recover mean for its own resources, so the same operation can leave a function in different states depending on which platform executed it. Automated testing has shown that even mature controllers violate their own desired-state semantics, which makes cross-platform management of AI functions hard to trust.

The thesis studies whether lifecycle operations on AI functions can be given a common meaning against which every platform can be checked. The student will: (i) define management contracts for a small set of operations, such as activation, model update with validation, monitoring and recovery, each specifying preconditions, required information, expected state change and the evidence of success or failure (an inference function is active only when it is observably ready to serve requests, not when the request has been accepted); (ii) implement bindings that execute the contracts on at least two platforms, for example Kubernetes/KServe and a NETCONF/YANG controller; (iii) build a conformance-testing harness that runs the same workflows through different bindings, from different initial states and with injected faults (failed nodes, invalid models, lost connectivity), and compares the resulting state transitions with the contract; (iv) quantify interoperability in terms of reused versus platform-specific code, conformance failures and operation latency, using real AI functions on the laboratory edge testbed.

The thesis combines network and service management, cloud-native orchestration and software testing.

Keywords: Network and service management; AI lifecycle management; Kubernetes; NETCONF/YANG; Conformance testing

Duration: 9 months

Prerequisites: Computer networks; Linux and containers (Docker); programming in Python or Go. Knowledge of Kubernetes and NETCONF/YANG is a plus.

Suggested reading

  1. B. Burns, B. Grant, D. Oppenheimer, E. Brewer, J. Wilkes, “Borg, Omega, and Kubernetes,” Communications of the ACM, vol. 59, no. 5, 2016.
  2. X. Sun et al., “Automatic Reliability Testing for Cluster Management Controllers,” in Proc. USENIX OSDI, 2022.
  3. J. T. Gu et al., “Acto: Automatic End-to-End Testing for Operation Correctness of Cloud System Management,” in Proc. ACM SOSP, 2023.
  4. 3GPP TS 28.105, “Management and orchestration; Artificial Intelligence/Machine Learning (AI/ML) management.”
  5. R. Enns, M. Bjorklund, J. Schönwälder, A. Bierman (Eds.), “Network Configuration Protocol (NETCONF),” RFC 6241, 2011.
  6. B. Claise, J. Clarke, J. Lindblad, Network Programmability with YANG, Addison-Wesley, 2019.
  7. D. Kreuzberger, N. Kühl, S. Hirschl, “Machine Learning Operations (MLOps): Overview, Definition, and Architecture,” IEEE Access, vol. 11, 2023.
  8. D. Sculley et al., “Hidden Technical Debt in Machine Learning Systems,” in Proc. NeurIPS, 2015.
  9. A. Clemm, L. Ciavaglia, L. Z. Granville, J. Tantsura, “Intent-Based Networking – Concepts and Definitions,” RFC 9315, 2022.

Negotiated Communication between AI Functions: Compiling Interaction Profiles into eBPF Packet Processing over SD-WAN Tunnels

Distributed AI applications at the edge combine functions developed independently: detectors on far-edge cameras, fusion and learning functions on edge servers, coordination agents. Today two functions can exchange information only if they share a protocol fixed at design time (HTTP/JSON, gRPC, MQTT or agent protocols such as A2A), and the network carrying their traffic, often an encrypted SD-WAN tunnel, knows nothing about what they exchange. Adding a new function, or adapting to a degraded link, means writing new adapters by hand.

The thesis explores a different approach: functions first agree on how to exchange information, and the agreement is then compiled into packet processing executed by the data plane. The student will: (i) define a compact, machine-readable interaction profile (fields, types, encoding, version, sequence number, timestamp, validity, required network service) and a deterministic offer/answer procedure that selects a common profile, binding to an existing protocol when both functions support one; (ii) implement a compiler that turns the agreed profile into a packet layout and into sender- and receiver-side eBPF programs (TC/XDP) that insert, validate and extract the metadata at the tunnel endpoints, outside the encrypted payload, extending the in-band telemetry mechanism for SD-WAN developed in our group; (iii) support profile changes at runtime through a version-tagged two-phase update, e.g. when the path moves from fibre to satellite and a lower-rate profile is agreed; (iv) evaluate the prototype on the laboratory testbed with a wildlife-detection pipeline, measuring agreement and switch time, CPU and tail-latency overhead, throughput and loss against gRPC, HTTP/JSON and hand-written adapters.

Programs are generated only from a closed grammar, never from code supplied by a peer, and must pass the eBPF verifier.

Keywords: eBPF; Programmable data plane; AI functions; Interoperability; SD-WAN

Duration: 9 months

Prerequisites: Computer networks (TCP/IP, tunnelling); good programming skills in C and Python; Linux. Previous experience with eBPF/XDP or P4 is a plus.

Suggested reading

  1. S. Troia, J. P. Asdikian, G. Sguotti, E. Gregorini, M. Li, G. Maier, “In-band Network Telemetry for Software-Defined Wide Area Networks,” Computer Networks, vol. 270, art. 111567, 2025.
  2. T. Høiland-Jørgensen et al., “The eXpress Data Path: Fast Programmable Packet Processing in the Operating System Kernel,” in Proc. ACM CoNEXT, 2018.
  3. M. A. M. Vieira et al., “Fast Packet Processing with eBPF and XDP: Concepts, Code, Challenges, and Applications,” ACM Computing Surveys, vol. 53, no. 1, 2020.
  4. J. Rosenberg, H. Schulzrinne, “An Offer/Answer Model with the Session Description Protocol (SDP),” RFC 3264, 2002.
  5. P. Bosshart et al., “P4: Programming Protocol-Independent Packet Processors,” ACM SIGCOMM Computer Communication Review, vol. 44, no. 3, 2014.
  6. C. Kim et al., “In-band Network Telemetry via Programmable Dataplanes,” ACM SIGCOMM 2015 (demo).
  7. Y. Zhou, Z. Wang, S. Dharanipragada, M. Yu, “Electrode: Accelerating Distributed Protocols with eBPF,” in Proc. USENIX NSDI, 2023.
  8. A2A Project (Linux Foundation), “Agent2Agent (A2A) Protocol Specification,” 2025.
  9. S. Troia, L. Borgianni, G. Sguotti, S. Giordano, G. Maier, “A Comprehensive Survey on Software-Defined Wide Area Network,” IEEE Communications Surveys & Tutorials, 2025.

Control-Plane Automation for Open-Source SD-WAN: Zero-Touch Provisioning and Lifecycle Management of Edge CPEs

Software-Defined Wide Area Networks (SD-WANs) interconnect distributed sites through encrypted overlay tunnels built over heterogeneous access links (fibre, 4G/5G, satellite). Much of their value comes from the control plane: a central controller must install, configure, monitor, update and retire Customer Premises Equipment (CPE) devices without manual intervention, often at sites with no technical staff. Commercial products provide these functions inside closed platforms, while open-source SD-WAN stacks mostly stop at the data plane.

The thesis starts from the open-source SD-WAN CPE developed in our laboratory, which provides WireGuard-based tunnels, BFD liveness detection and fast packet processing, and designs the control plane that manages its whole lifecycle. The student will: (i) design a zero-touch provisioning workflow in which a new CPE authenticates to the controller, receives its identity, keys and initial configuration, and joins the overlay without manual steps; (ii) define a declarative, model-driven configuration interface for tunnels, routing and traffic policies (e.g., YANG with NETCONF or gNMI, or an equivalent intent-based API); (iii) implement Day-2 operations: streaming telemetry, periodic key rotation, staged software updates with rollback, and safe behaviour of the CPEs when the controller is unreachable; (iv) evaluate the system on the laboratory SD-WAN testbed and on an emulated network with tens to hundreds of virtual CPEs, measuring onboarding time, convergence time after a policy change, control-plane overhead and controller scalability.

The expected outcome is an open-source controller integrated with the existing CPE, together with an experimental assessment of its performance.

Keywords: SD-WAN; Control plane; Zero-touch provisioning; Network automation; Intent-based networking

Duration: 9 months

Prerequisites: Computer networks (TCP/IP, routing, VPNs); Linux system administration; programming in Python or Go. Basic knowledge of network security (PKI, TLS) is a plus.

Suggested reading

  1. S. Troia, L. Borgianni, G. Sguotti, S. Giordano, G. Maier, “A Comprehensive Survey on Software-Defined Wide Area Network,” IEEE Communications Surveys & Tutorials, 2025.
  2. G. Sguotti, S. Troia, M. Grieco, G. Maier, “An Open Source SD-WAN CPE with Fast Packet Processing and Secure Connectivity,” in Proc. IEEE HPSR, 2025.
  3. G. Sguotti et al., “On the Service-Oriented Availability Analysis of Software Defined Wide Area Network,” in Proc. IEEE NetSoft, 2025.
  4. MEF 70.1, “SD-WAN Service Attributes and Service Framework,” MEF Forum, 2021.
  5. K. Watsen, I. Farrer, M. Abrahamsson, “Secure Zero Touch Provisioning (SZTP),” RFC 8572, 2019.
  6. B. Claise, J. Clarke, J. Lindblad, Network Programmability with YANG, Addison-Wesley, 2019.
  7. A. Clemm, L. Ciavaglia, L. Z. Granville, J. Tantsura, “Intent-Based Networking – Concepts and Definitions,” RFC 9315, 2022.
  8. D. Kreutz et al., “Software-Defined Networking: A Comprehensive Survey,” Proceedings of the IEEE, vol. 103, no. 1, 2015.
  9. J. A. Donenfeld, “WireGuard: Next Generation Kernel Network Tunnel,” in Proc. NDSS, 2017.
  10. ETSI GS ZSM 002, “Zero-touch network and Service Management (ZSM); Reference Architecture,” 2019.

Service-Aware Data Plane for SD-WAN: SLA-Driven Path Selection, Packet Replication and Fast Packet Processing on Edge CPEs

In an SD-WAN each Customer Premises Equipment (CPE) reaches the other sites through several encrypted tunnels over different access networks, whose delay, jitter, loss and available bandwidth change over time. The services carried on top have different needs: a video stream, a control message, a bulk dataset upload and an AI model transfer cannot be treated in the same way. Many deployments, however, select one path per destination using a single metric such as the lowest round-trip time, leaving most of the diversity of the available links unused.

The thesis designs and implements a service-aware data plane for the open-source SD-WAN CPE developed in our laboratory. The student will: (i) define service classes with explicit SLA targets and a lightweight classification stage (port- or DSCP-based, optionally ML-based for encrypted flows); (ii) measure delay, jitter and loss per tunnel, combining active probes with the in-band telemetry mechanism for SD-WAN developed in our group, and design a steering policy that maps each class to the tunnel, or set of tunnels, meeting its SLA, with hysteresis to avoid oscillations; (iii) add mechanisms for critical traffic, such as packet replication with duplicate elimination over disjoint tunnels and per-packet load balancing with reordering control; (iv) implement the fast path with a programmable packet-processing technology (eBPF/XDP, VPP/DPDK or P4) and quantify its cost in CPU usage and throughput.

The solution will be evaluated on the laboratory testbed with real access links (fixed, cellular and satellite) and with controlled impairments (Linux tc/netem), measuring SLA violation ratio, goodput, tail latency, failover time and CPU load at increasing traffic rates.

Keywords: SD-WAN; Programmable data plane; Application-aware routing; Packet replication; Quality of Service

Duration: 9 months

Prerequisites: Computer networks (TCP/IP, routing, QoS); good programming skills in C and/or Python; Linux networking. Knowledge of eBPF, DPDK or P4 is a plus.

Suggested reading

  1. S. Troia, L. Borgianni, G. Sguotti, S. Giordano, G. Maier, “A Comprehensive Survey on Software-Defined Wide Area Network,” IEEE Communications Surveys & Tutorials, 2025.
  2. G. Sguotti, S. Troia, M. Grieco, G. Maier, “An Open Source SD-WAN CPE with Fast Packet Processing and Secure Connectivity,” in Proc. IEEE HPSR, 2025.
  3. S. Troia, J. P. Asdikian, G. Sguotti, E. Gregorini, M. Li, G. Maier, “In-band Network Telemetry for Software-Defined Wide Area Networks,” Computer Networks, vol. 270, art. 111567, 2025.
  4. S. Troia, F. Sapienza, L. Varé, G. Maier, “On Deep Reinforcement Learning for Traffic Engineering in SD-WAN,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 7, 2021.
  5. T. Høiland-Jørgensen et al., “The eXpress Data Path: Fast Programmable Packet Processing in the Operating System Kernel,” in Proc. ACM CoNEXT, 2018.
  6. D. Barach et al., “High-Speed Software Data Plane via Vectorized Packet Processing,” IEEE Communications Magazine, vol. 56, no. 12, 2018.
  7. N. Finn, P. Thubert, B. Varga, J. Farkas, “Deterministic Networking Architecture,” RFC 8655, 2019 (packet replication and elimination).
  8. D. Katz, D. Ward, “Bidirectional Forwarding Detection (BFD),” RFC 5880, 2010.
  9. Q. De Coninck, O. Bonaventure, “Multipath QUIC: Design and Evaluation,” in Proc. ACM CoNEXT, 2017.
  10. S. Rezaei, X. Liu, “Deep Learning for Encrypted Traffic Classification: An Overview,” IEEE Communications Magazine, vol. 57, no. 5, 2019.

Edge-to-Edge Communication over LEO Satellite Links: Prediction-Driven Control of Delay, Jitter and Packet Loss in SD-WAN

Low Earth Orbit (LEO) constellations are becoming a common underlay for SD-WANs that connect remote edge sites, where terrestrial connectivity is weak or absent. A LEO path is far from a stable pipe: measurements on the OneWeb link available in our laboratory show latency steps of tens of milliseconds caused by gateway changes, and studies of other constellations report jitter, packet reordering and bursts of loss around satellite handovers. A previous thesis built a geometric digital twin that predicts gateway changes and steers traffic to a terrestrial backup before they occur. Moving traffic away is not always possible or convenient, though, and steering alone does not address jitter and loss.

This thesis studies how to keep edge-to-edge flows within delay, jitter and loss targets while they continue to use the satellite path. The student will: (i) extend the measurement campaign from round-trip latency to one-way delay, jitter, loss, reordering and throughput, with synchronised probes at both ends of the SD-WAN tunnel; (ii) characterise how these impairments affect transport protocols (TCP CUBIC, BBR, QUIC) and typical edge applications, such as AI model and dataset transfers, video streams and control messages; (iii) design and implement protection mechanisms in the SD-WAN edge, such as selective packet duplication over satellite and terrestrial links, forward error correction, and de-jittering and reordering buffers at the receiving CPE, activated only when the digital twin or a learned predictor anticipates an impairment, so that the bandwidth overhead stays low; (iv) evaluate them in the laboratory testbed with trace-driven emulation and on the real satellite link.

Results will be expressed as the trade-off between application-level quality and overhead, compared with reactive steering and always-on protection.

Keywords: LEO satellite networks; SD-WAN; Jitter and packet loss; Network digital twin; Edge-to-edge communication

Duration: 9 months

Prerequisites: Computer networks and transport protocols (TCP, UDP, QUIC); Python for data analysis; Linux networking (tc/netem). Basic knowledge of satellite communications is a plus.

Suggested reading

  1. G. Sguotti, S. Troia, G. Maier, “Mitigating Satellite Link Variability in SD-WAN Using Constellation-Aware Edge Routing,” in Proc. ICTON, 2025.
  2. F. M. Cioffi, “A Digital Twin for Predictive Traffic Steering in Satellite-Terrestrial SD-WANs,” MSc thesis, Politecnico di Milano, 2026.
  3. J. Zhao, O. Perrin, A. Ahangarpour, J. Pan, “Measuring the OneWeb Satellite Network,” in Proc. IFIP TMA, 2025.
  4. F. Michel, M. Trevisan, D. Giordano, O. Bonaventure, “A First Look at Starlink Performance,” in Proc. ACM IMC, 2022.
  5. S. Ma et al., “Network Characteristics of LEO Satellite Constellations: A Starlink-Based Measurement from End Users,” in Proc. IEEE INFOCOM, 2023.
  6. R. S. Frederiksen, T. G. Mulvad, I. Leyva-Mayorga, T. K. Madsen, F. Chiariotti, “End-to-End Delivery in LEO Mega-Constellations and the Reordering Problem,” in Proc. IEEE PIMRC, 2024.
  7. S. Kassing, D. Bhattacherjee, A. B. Águas, J. E. Saethre, A. Singla, “Exploring the ‘Internet from Space’ with Hypatia,” in Proc. ACM IMC, 2020.
  8. O. Kodheli et al., “Satellite Communications in the New Space Era: A Survey and Future Challenges,” IEEE Communications Surveys & Tutorials, vol. 23, no. 1, 2021.
  9. N. Cardwell, Y. Cheng, C. S. Gunn, S. H. Yeganeh, V. Jacobson, “BBR: Congestion-Based Congestion Control,” ACM Queue, vol. 14, no. 5, 2016.
  10. M. Watson, A. Begen, V. Roca, “Forward Error Correction (FEC) Framework,” RFC 6363, 2011.

Drift-Aware Continual Adaptation and Efficient Inference of Wildlife Detectors on Far-Edge Camera Traps

Camera traps built on low-cost boards such as the Raspberry Pi can run compact object detectors on site, so that only relevant images or detections travel over the scarce connectivity of rural areas. Their accuracy drops, however, when field conditions differ from the training data: night-time infrared images, new viewpoints, weather, seasonal vegetation. In previous work we built a far-edge/edge/cloud framework in which a YOLO detector running on the camera is adapted at the edge with semi-supervised learning on unlabeled field images, followed by teacher-student weight interpolation, and redistributed over an SD-WAN. The adaptation was performed once, on a fixed batch of data, and the model ran in its original PyTorch format on the device.

This thesis turns the one-shot adaptation into a continual, resource-aware loop for long deployments. The student will: (i) implement lightweight drift detection on the camera (e.g., on confidence statistics or feature distributions) to decide when the deployed model is no longer reliable; (ii) design repeated adaptation rounds at the edge that improve on new conditions without forgetting previous ones, comparing weight interpolation, replay and test-time adaptation; (iii) optimise on-device inference with post-training quantization, pruning and efficient runtimes (ONNX Runtime, NCNN, TFLite) on Raspberry Pi and Jetson-class boards, measuring accuracy, latency and energy; (iv) evaluate the complete loop on day and night field images collected by our camera traps and on additional camera domains, including the cost of each update in transferred data and edge compute time.

The work combines computer vision, embedded AI and edge networking, with access to real camera-trap hardware and field data.

Keywords: Edge AI; Wildlife monitoring; Object detection; Continual domain adaptation; Model compression

Duration: 9 months

Prerequisites: Machine learning and deep learning fundamentals; Python and PyTorch; basics of computer vision. Experience with embedded Linux boards is a plus.

Suggested reading

  1. G. Maier et al., “WatchEDGE: Smart Networking for Distributed AI-Based Environmental Control,” Computer Networks, vol. 243, art. 110248, 2024.
  2. J. P. Asdikian et al., “YOLO-AUG: Enhancing Wildlife Monitoring on Far-Edge Devices with Dynamic Color-Space Augmentation and Efficient Learning,” IEEE Access, vol. 14, 2026.
  3. J. P. Asdikian et al., “WatchEDGE: A Distributed Edge–Cloud Framework for AI-Driven Wildlife Monitoring in Rural Environments,” manuscript under review.
  4. S. Beery, G. Van Horn, P. Perona, “Recognition in Terra Incognita,” in Proc. ECCV, 2018.
  5. M. S. Norouzzadeh et al., “Automatically Identifying, Counting, and Describing Wild Animals in Camera-Trap Images with Deep Learning,” PNAS, vol. 115, no. 25, 2018.
  6. S. Rabanser, S. Günnemann, Z. C. Lipton, “Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift,” in Proc. NeurIPS, 2019.
  7. Q. Wang, O. Fink, L. Van Gool, D. Dai, “Continual Test-Time Domain Adaptation,” in Proc. IEEE/CVF CVPR, 2022.
  8. M. Wortsman et al., “Robust Fine-Tuning of Zero-Shot Models,” in Proc. IEEE/CVF CVPR, 2022.
  9. B. Jacob et al., “Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference,” in Proc. IEEE/CVF CVPR, 2018.
  10. Open wildlife image dataset of the project, GitLab: restart-watchedge/wild-animals-open-dataset.

Learning-Based Pricing and Fast Re-Allocation in Hierarchical Auctions for Multi-Operator RAN Slicing

Network slicing lets several virtual operators share the same 5G/6G infrastructure. In multi-operator scenarios, each Mobile Network Operator (MNO) controls a region of the network and sells slices to Mobile Virtual Network Operators (MVNOs), which resell them to end users. Our previous work modelled this market as a two-level combinatorial auction with pair-bid matching at both levels, VCG-like critical-value payments that support truthful bidding, power-aware placement of CU/DU functions, and dynamic arrival and departure of requests. Two simplifications limit its realism: MNO asking prices are fixed inputs, and the allocation is recomputed from scratch at every round, which suits monthly timeslots but not short-lived slices.

The thesis addresses both limitations. The student will: (i) model the repeated interaction among MNOs as a pricing game and design learning-based pricing agents (e.g., multi-armed bandits or multi-agent reinforcement learning) that adapt asking prices across rounds from observed outcomes; (ii) analyse the resulting market in terms of social welfare, operator revenue, energy cost, price stability and possible tacit collusion, and check which incentive properties of the buyer-side mechanism are preserved; (iii) design an incremental re-allocation scheme that reuses previous CU/DU placements and computes critical-value payments in parallel, to shorten the timeslot towards hours or minutes; (iv) evaluate the approach on the existing simulator over 10- and 36-node metro topologies and against exact ILP solutions on small instances.

The thesis combines optimisation, mechanism design and machine learning applied to network management.

Keywords: Network slicing; Combinatorial auctions; Mechanism design; Multi-agent reinforcement learning; RAN

Duration: 9 months

Prerequisites: Optimisation (linear and integer programming); programming in Python; basics of game theory or reinforcement learning. Knowledge of the 5G architecture is a plus.

Suggested reading

  1. M. Li, S. Troia, J. P. Asdikian, G. Sguotti, Y. Zhang, G. Maier, “Hierarchical Auction-Based 5G Network Slicing with Region-Based Multi-MNO Deployment,” manuscript under review.
  2. M. Li, S. Troia, Y. Zhang, G. Maier, “Pair-Bid Auction Model for Optimized Network Slicing in 5G RAN,” in Proc. IEEE ISCC Workshops (NGMSE), 2025, arXiv:2509.10533.
  3. M. Li, L. M. M. Zorello, S. Troia, Y. Zhang, G. Maier, “Optimization of 5G RAN Network Slicing Based on Auction Models,” Computer Networks, art. 112141, 2026.
  4. L. M. M. Zorello, M. Sodano, S. Troia, G. Maier, “Power-Efficient Baseband-Function Placement in Latency-Constrained 5G Metro Access,” IEEE Transactions on Green Communications and Networking, vol. 6, no. 3, 2022.
  5. K. Zhu, E. Hossain, “Virtualization of 5G Cellular Networks as a Hierarchical Combinatorial Auction,” IEEE Transactions on Mobile Computing, vol. 15, no. 10, 2016.
  6. D. Lehmann, L. I. O’Callaghan, Y. Shoham, “Truth Revelation in Approximately Efficient Combinatorial Auctions,” Journal of the ACM, vol. 49, no. 5, 2002.
  7. N. Nisan, T. Roughgarden, É. Tardos, V. V. Vazirani (Eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (chapters on mechanism design and combinatorial auctions).
  8. E. Calvano, G. Calzolari, V. Denicolò, S. Pastorello, “Artificial Intelligence, Algorithmic Pricing, and Collusion,” American Economic Review, vol. 110, no. 10, 2020.
  9. P. Caballero, A. Banchs, G. de Veciana, X. Costa-Pérez, “Network Slicing Games: Enabling Customization in Multi-Tenant Mobile Networks,” IEEE/ACM Transactions on Networking, vol. 27, no. 2, 2019.
  10. R. S. Sutton, A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed., MIT Press, 2018.

Learning-Aware Orchestration of Hierarchical Federated Learning over SD-WANs: From Round Completion to Time-to-Accuracy

Hierarchical Federated Learning (HFL) lets fixed and mobile edge devices, such as camera traps and drones, train a shared model by merging their updates at intermediate edge aggregators before they reach a global one. When the edge sites are connected by an SD-WAN with limited and time-varying bandwidth, the choice of aggregators, the shape of the aggregation tree and the physical paths of the model transfers must be decided together, since each choice changes the load on the links used by the others. Existing orchestration schemes, including those developed in our group, optimise network-level metrics such as delay, resource cost or the fraction of rounds completed within a deadline, but they do not measure how fast the model actually learns.

The thesis closes this gap. The student will: (i) couple the network orchestration simulator available in our group with a real federated-learning framework (e.g., Flower with PyTorch), so that every decision produces actual model updates on realistic non-IID data, such as wildlife images from field camera traps; (ii) extend the decision space with learning-related options, such as partial participation, aggregation deadlines, semi-asynchronous aggregation and update compression; (iii) design orchestration policies, including deep reinforcement learning, whose objective is time-to-accuracy and fairness among clients rather than round completion alone; (iv) validate the approach in an emulated SD-WAN running the real CPE software and on a small testbed with physical edge devices.

Metrics include time to target accuracy, final accuracy, round failure rate, link and node utilisation and energy. The work combines networking and distributed machine learning.

Keywords: Hierarchical federated learning; SD-WAN; Edge intelligence; Deep reinforcement learning; Network-aware orchestration

Duration: 9 months

Prerequisites: Machine learning and deep learning basics; Python and PyTorch; computer networks. Experience with reinforcement learning or network emulation (Mininet, containerlab) is a plus.

Suggested reading

  1. H. B. McMahan et al., “Communication-Efficient Learning of Deep Networks from Decentralized Data,” in Proc. AISTATS, 2017.
  2. P. Kairouz et al., “Advances and Open Problems in Federated Learning,” Foundations and Trends in Machine Learning, vol. 14, no. 1–2, 2021.
  3. L. Liu, J. Zhang, S. H. Song, K. B. Letaief, “Client-Edge-Cloud Hierarchical Federated Learning,” in Proc. IEEE ICC, 2020.
  4. S. Luo et al., “HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning,” IEEE Transactions on Wireless Communications, vol. 19, no. 10, 2020.
  5. M. Li et al., “On the Optimization of Model Aggregation for Federated Learning at the Network Edge,” IEEE Transactions on Network and Service Management, 2025.
  6. M. Li et al., “Demo: Adaptive Resource Allocation Simulator for Federated Learning in MEC-Driven SD-WANs,” in Proc. IEEE NetSoft, 2025.
  7. S. Wang et al., “Adaptive Federated Learning in Resource Constrained Edge Computing Systems,” IEEE Journal on Selected Areas in Communications, vol. 37, no. 6, 2019.
  8. T. Li et al., “Federated Optimization in Heterogeneous Networks,” in Proc. MLSys, 2020.
  9. D. J. Beutel et al., “Flower: A Friendly Federated Learning Research Framework,” arXiv:2007.14390, 2020.
  10. J. Schulman et al., “Proximal Policy Optimization Algorithms,” arXiv:1707.06347, 2017.