Vehicle trajectory prediction in Internet-of-Vehicles requires collaborative learning over sensitive trajectories under intermittent connectivity and partially trusted participants. ChainDrive-FL-VRA coordinates semi-asynchronous federated learning on a permissioned consortium ledger using Practical Byzantine Fault Tolerance (PBFT), while keeping raw trajectories and raw model-update tensors off-chain. Each client submits an on-chain header containing a commitment and hash of the local update, together with zero-knowledge proofs that certify l₂clipping and anchor-consistency. Validators admit only proof-checked updates, compute staleness- and reputation-aware robust weights, and publish a proof of correct aggregation that binds the aggregation commitment and the committed global model hash to the admitted committed updates under fixed-point weights. A contextual-bandit trigger selects aggregation timing under client churn. Experiments on NGSIM US-101 and I-80 show improved ADE/FDE/RMSE and improved robustness under staleness and anomalous updates, while on-chain artifacts remain at kilobyte scale per update and per aggregation event.
Reddy et al. (Mon,) studied this question.