STARLoc is a federated vehicle localization framework that combines a Graph Attention Network (GAT), a Long Short Term Memory (LSTM) temporal model, and a transformer based fusion module.GNSS, inertial and odometric sensing, and Vehicle to Vehicle (V2V) messages provide complementary inputs; GAT emphasizes reliable neighbors, and LSTM captures temporal continuity.Secure federated aggregation uses differential privacy and homomorphic encryption.The contribution lies in integrating graph attention and temporal modeling with staleness aware secure aggregation, differential privacy, homomorphic encryption, and gossip of incremental deltas with compression.RMSE equals 2.30 m in urban canyon and 1.20 m on highway, improving over ST FTL (3.10/1.80m) and FedVCP (4.70/2.50m), with convergence in 44 ± 6 rounds.Privacy preservation via differential privacy and homomorphic encryption increases RMSE by 1.74% relative to the non private federated baseline.The framework scales from 100 to 500 vehicles with modest degradation.Within the evaluated SUMO settings and baselines (ST-FTL and FedVCP), STARLoc achieves the best accuracy while preserving privacy and communication efficiency.Comparison with additional federated localization frameworks such as FedLoc, FedPylot, and hierarchical federated learning is left to future work.