Differentially Private Decentralized Traffic Flow Prediction Approach based on Federated Learning

Existing centralized deep learning methods get surprising success in traffic flow prediction based on large-scale datasets. However, to protect the privacy of datasets, organizations are not allowed to share local datasets, which causes data exists as silos. The emergence of federated learning has broken this awkward situation. But the new challenge that puzzles us is how to provide meaningful privacy guarantees in federated learning. In this paper, we apply the federated learning to the intelligent transportation domain and propose a traffic flow prediction method based on long short-term memory (LSTM) networks and differential privacy. We introduce blockchain technology to verify the model update parameters in each round and achieve fully decentralized training. Experiments show that we can guarantee the accuracy of the model under a proper privacy budget, and the communication costs of our method are controllable.