Abstract

The proliferation of Internet of Things sensors in smart buildings and Internet of Vehicles networks generates privacy-sensitive time series data at unprecedented scales, posing significant challenges for traditional learning paradigms. Centralized learning approaches are hampered by privacy risks and difficulties in handling heterogeneous data, while federated learning incurs server computational bottlenecks. This thesis presents a comprehensive framework progressing from privacy-preserving edge analytics to fully decentralized coordination for anomaly detection and long-term times series forecasting. The thesis contributions are threefold. First, we introduce FSLSTM, a federated stacked long short-term memory approach for anomaly detection across heterogeneous building sensors (HVAC, lighting, access control). Without centralizing raw telemetry, FSLSTM achieves competitive performance on various real-world datasets while converging twice faster compared to centralized baselines and reducing communication costs by 45%. Second, building on this foundation, we develop FedTime, a federated large language model framework that adapts Llama-2 for long-term time series forecasting. Through cluster-aware federation, channel independence, patching and task alignment techniques, FedTime outperforms strong forecasting baselines across several benchmarks including a production-scale electric vehicle charging data. The FedTime framework demonstrates thrice faster convergence than centralized training, while reducing communication overhead by 50%. Third, to eliminate residual trust assumptions, we propose BLEND, a blockchain-enhanced decentralized system that replaces server-based aggregation with smart contract orchestration. Through a novel Proof-of-Forecast consensus mechanism that validates updates based on predictive accuracy, BLEND yields strong forecasting performance while reducing network overhead by 63% and demonstrating resilience to 20\% Byzantine participants. Collectively, these contributions establish a complete pipeline from edge sensing to decentralized learning, demonstrating that privacy preservation, computational efficiency and decentralized trust can be achieved simultaneously without sacrificing predictive performance.