Federated Learning for Enhancing Physical‑Layer Security in 6G Networks

Thesis Proposal Details

Supervisor: Stefano Tomasin

Creation Date: 25/12/2025 20:33

Description

1. Background & Motivation

The next generation of mobile communications (6G) will rely heavily on intelligent, data‑driven techniques to protect the physical layer against eavesdropping, jamming, and spoofing. Conventional security mechanisms (encryption, authentication) operate at higher layers and may not react quickly enough to the ultra‑low‑latency, massive‑connectivity demands of 6G.

Federated Learning (FL) offers a promising avenue: distributed devices train local models on their own observations (channel state information, interference patterns, etc.) and only exchange model updates with a central aggregator. This preserves user privacy and reduces the need to ship raw radio data to a cloud server.

However, FL assumes that the data held by each participant are independent and identically distributed (i.i.d.). In realistic wireless environments, devices experience vastly different channel conditions, mobility patterns, and interference levels, resulting in highly non‑i.i.d. data. When the data distribution diverges, a single global model can under‑perform, jeopardizing the security gains that FL seeks to provide.

The goal of this thesis is to investigate, adapt, and extend state‑of‑the‑art FL strategies to the non‑i.i.d. setting of physical‑layer security, and to validate the proposed solutions on a representative benchmark (CIFAR‑100) before moving to realistic 6G‑style datasets.

2. Objectives

  1. Survey and analyse existing FL techniques that mitigate non‑i.i.d. effects (e.g., personalization, clustered FL, meta‑learning, adaptive weighting).
  2. Design novel FL extensions tailored to physical‑layer security tasks (e.g., anomaly detection, secure modulation classification).
  3. Implement the algorithms using PyTorch, managing the codebase with VS Code and GitHub for reproducibility.
  4. Evaluate performance on the CIFAR‑100 image classification benchmark (as a proxy for heterogeneous data) and on a simulated 6G physical‑layer dataset (channel measurements, adversarial attacks).
  5. Produce a comprehensive technical report documenting methodology, experimental results, and recommendations for integrating FL‑based security in future 6G standards.

3. Methodology

Phase Activities Expected Outputs
Phase 1 – Literature Review & Baseline Setup (Month 1) • Compile recent publications on FL for non‑i.i.d. data (personalized FL, FedAvg variants, hierarchical FL). • Set up a baseline FL pipeline in PyTorch (FedAvg) and replicate CIFAR‑100 results reported in the literature. • Annotated bibliography. • Working baseline code repository (GitHub).
Phase 2 – Algorithm Development (Months 2‑3) • Implement selected mitigation strategies (e.g., FedProx, FedAvg with cluster‑wise aggregation, model‑agnostic meta‑learning). • Propose two original extensions:  1. Security‑aware weighting – clients contributing more informative channel features receive higher aggregation weight.  2. Hybrid personalization – a shared global backbone plus lightweight client‑specific heads for fine‑grained anomaly detection. • Extended codebase with modular implementations. • Documentation of algorithmic modifications.
Phase 3 – Dataset Preparation & Simulation (Month 4) • Use CIFAR‑100 to emulate heterogeneous label distributions (partition classes unevenly across clients). • Generate a synthetic 6G physical‑layer dataset: simulate Rayleigh/Rician fading, mobility, and adversarial attacks (jamming, spoofing). • Encode the data as tensors compatible with the FL pipeline. • Two benchmark datasets (CIFAR‑100 partitions, 6G channel/attack dataset).
Phase 4 – Experiments & Analysis (Month 5) • Run extensive experiments measuring:  - Global model accuracy / detection rate.  - Convergence speed (communication rounds).  - Communication overhead (model size, bits transmitted). • Compare baseline FedAvg against each mitigation strategy and the two novel extensions. • Perform ablation studies on the impact of client heterogeneity level. • Tables and plots summarising performance metrics. • Insightful discussion of trade‑offs (privacy vs. security vs. efficiency).
Phase 5 – Reporting & Dissemination (Month 6) • Write the master‑thesis manuscript (~60 pages) covering motivation, related work, methodology, results, and future directions. • Prepare a technical report for the supervising department and a short conference‑style paper (optional).   • Completed thesis document.. • Presentation slides for the defence.

4. Expected Contributions

Contribution Description
Enhanced FL Algorithms Two security‑aware FL extensions that improve robustness to non‑i.i.d. data in physical‑layer contexts.
Benchmark Suite A reproducible CIFAR‑100 partitioning scheme and a synthetic 6G physical‑layer dataset for future research on FL‑based security.
Performance Insights Quantitative analysis of how heterogeneity, communication budget, and security‑aware weighting affect convergence and detection accuracy.
Open‑Source Toolkit Fully documented PyTorch implementation, ready for integration into larger 6G testbeds.
Thesis Document A self‑contained master‑level dissertation that can serve as a reference for researchers exploring privacy‑preserving security in next‑generation networks.

5. Timeline (6 months)

Month Milestone
1 Literature review, baseline FedAvg implementation, repository setup.
2‑3 Development of mitigation strategies and the two novel extensions; code integration.
4 Creation of heterogeneous CIFAR‑100 splits and synthetic 6G dataset.
5 Full experimental campaign, result aggregation, ablation studies.
6 Thesis writing, final report, code release, defence preparation.

 

Dataset and methods

Dataset type: Already acquired data

Dataset description: images and cellular channels

List of Methods: Python

Tags
6G federated learning physical layer security
Back to proposals list