Semantics- and Energy-Aware Decentralized Federated Learning: Algorithms Implementation and Performance Analysis

Federated learning allows distributed model training without uploading raw data, but its centralized architecture may cause communication bottlenecks, a single point of failure, and trust dependence in modern edge networks. Decentralized federated learning removes the central server and lets each node train locally and exchange models with its neighbors. However, many scheduling mechanisms designed for a global coordinator cannot be directly used in this setting. This thesis studies how two representative centralized scheduling ideas, i.e., (i) a client selection policy based on the version age of information (VAoI), and (ii) FedBacys, a battery-aware cyclic client scheduling procedure, can be adapted to a decentralized framework. For the freshness-aware direction, centralized VAoI is reformulated as a local neighbor selection rule. Each node maintains its own version age records and combines it with model-distance information to decide which neighbor models should be aggregated. For the energy-aware direction, FedBacys is adapted from a hub-coordinated setting to a decentralized bridged topology. The group structure is preserved through full intra-group connectivity, while bridge nodes support information flow between groups. Three local conditions, including sufficient battery, no pending update, and the cyclic deadline condition, determine whether a node starts training in each slot. The proposed methods are evaluated on the MNIST dataset training a lightweight multi-layer perceptron (MLP) model under independent and identically distributed (IID) and percent-based label-skew non-IID data settings. Experimental results show that decentralized VAoI achieves higher final test accuracy and lower average version age than random sampling under heterogeneous data. Decentralized FedBacys provides a favorable energy–accuracy trade-off in the battery-aware experiments, reducing total energy consumption by up to about 58% in the energyabundant setting, while showing its largest accuracy advantage under non-IID data.