Federated deep learning for smart city edge-based applications

The growing quantities of data allow for advanced analysis. A prime example of it are smart city applications with forecasting urban traffic flow as a key application. However, data privacy becomes a real issue. This problem can be addressed by using federated learning trusted authority principle. In this paper, we investigate a novel federated deep learning approach to urban traffic flow forecasting that graph learning, and trusted authority mechanism. The road network is first pre-processed to eliminate the noise from the traffic data. Next, detecting anomalous features is performed to prune irrelevant edges and patterns. The generated graph is then utilized to learn a graph convolutional neural network for calculating the future city’s traffic flow. We extensive evaluate our federated learning-based framework, where a case study on predicting the future traffic flows has been carried out using multiple datasets. We examine it with different baseline techniques as well. The findings show that the suggested framework greatly outperforms the baseline methods, particularly when the graph has a lot of nodes. Importantly, our approach is the first one that integrates trusted authority principle in federated learning and, by doing so, it is able to efficiently secure model data. Moreover, the average precision of the developed model reached 84%, while the baseline solutions did not exceed 77%.