Mobility and cloud management with federated and distributed learning
Pacheco, Lucas de Sousa · BORIS Theses (Bern Open Research Information System) (Bern University Library, Hochschulstrasse 6, 3012 Bern, Switzerland) · 2025
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
BORIS Theses (Bern Open Research Information System) (Bern University Library, Hochschulstrasse 6, 3012 Bern, Switzerland) · Nakuru County Council · dissertation · 2025
Abstract
Modern vehicular networks face challenges supporting applications like real-time traffic prediction, collision avoidance, and adaptive signal control, which require dynamic topology management, privacy preservation, and low latency. High mobility disrupts vehicle-infrastructure connectivity, hindering data exchange, while heterogeneous sensor data from diverse onboard systems violates homogeneous data assumptions in traditional collaborative frameworks. Additional complexities include mmWave beam alignment demands and adversarial threats.
Conventional approaches like FedAvg struggle due to rigid client selection, centralized aggregation bottlenecks, and one-size-fits-all compression, ignoring vehicular resource disparities. This thesis introduces four integrated frameworks to address these challenges through context-aware adaptations of FL principles. DOTFL (Chapter 3) addresses non-IID data distributions and poisoning attacks via neural similarity metrics and Wasserstein distance-based clustering, achieving 94% malicious update rejection while improving accuracy by 22% over FedAvg in urban mobility scenarios.
DrivePFL (Chapter 4) optimizes bandwidth utilization through Kalman Filter-predicted contact windows and layer-wise model transmission, reducing communication overhead by 10% without compromising 83.4% inference accuracy under vehicular mobility patterns. FLIPS (Chapter 5) integrates SHapley Additive exPlanations (SHAP) for adaptive model pruning, compressing transmissions by 48% while preserving safety-critical features through layer importance scoring. eDAFL (Chapter 6) accelerates mmWave beam selection via dynamic layer clustering, achieving 84% faster sector search than exhaustive search protocols through federated multi-sensor fusion.
The frameworks are validated through simulations combining realistic mobility traces, SUMO traffic models, and NS-3 network emulation. Key innovations include: first: Optimal transport-based distribution alignment for non-IID data without raw data access (DOTFL, Chapter 3); 2nd: Mobility-aware client selection using predicted link durations (DrivePFL, Chapter 4); 3rd: Selective and adaptive layer pruning for throughput optimization (FLIPS, Chapter 5); 4th: Hierarchical aggregation with Centered Kernel Alignment (CKA) for model consistency (eDAFL, Chapter 6).
Experimental results demonstrate scalability to 100-vehicle networks with 200ms inference latency in DrivePFL, 91% accuracy under non-IID data distributions (CIFAR-100) via DOTFL, and 52% parameter reduction through SHAP-guided compression in FLIPS. These contributions advance distributed learning theory by formalizing trade-offs between communication efficiency, adversarial robustness, and model interpretability in vehicular systems.
Open the resolved locator · 10.48549/6848
Metadata licence: OpenAlex metadata (CC0)