Abstract: Urban traffic flow prediction is becoming more and more important in traffic management. However, these data often belong to different institutions and cannot be interconnected, and these data involve the privacy of the traveling public, so there are risks in centralized storage. Some researchers have used federated learning model to predict traffic, but federated learning itself also has privacy risks. This paper proposes a federated learning method for privacy protection of traffic flow prediction, which uses the federated learning algorithm based on flow prediction algorithm GRU with privacy protection ability to predict traffic flow. The specific approach is to introduce differential privacy into the local GRU algorithm, make the DPGRU algorithm on the client meet the requirement of (ε,δ)differential privacy by adding Gaussian noise to the gradient and make model parameters random. This paper analyzes the privacy of DPGRU algorithm on the client, and makes comparative experiments on the actual traffic flow data set. Experiments show that the method obtains better prediction results on the premise of ensuring privacy.

Key words: federated learning, differential privacy, traffic flow prediction, GRU, privacypreserving, data security

摘要： 城市交通流量预测在交通管理中变得越来越重要.然而，这些数据往往属于不同机构，无法互联互通，且数据涉及出行大众隐私，集中存储也存在风险.已有研究者用联邦学习模式进行流量预测，但联邦学习本身也存在隐私隐患.提出一种面向交通流量预测隐私保护的联邦学习方法，采用具有隐私保护能力的基于流量预测算法GRU的联邦学习算法进行交通流量预测.具体做法是将差分隐私引入本地GRU算法中，通过在梯度中添加高斯噪声，使得客户端的DPGRU算法满足(ε,δ)差分隐私，且使得模型参数具有随机性.对客户端的DPGRU算法进行了隐私性分析，并在实际交通流量数据集上进行了对比实验.实验表明，在保证隐私的前提下，方法得到了较优的预测结果.

关键词: 联邦学习, 差分隐私, 交通流量预测, GRU, 隐私保护, 数据安全