Desensitization method of meteorological data based on differential privacy protection

Artificial intelligence (AI) profoundly affects the research of meteorology. As an AI method, deep learning greatly improves the accuracy of weather forecasting. Deep learning training requires a lot of data, while the collection of meteorological data faces some problems, such as the long collection cycle, the high cost, and the desensitization requirement. We propose a deep learning model called MDPGAN (Meteorology Differential Privacy Generative Adversarial Network), which introduces a differential privacy framework to reduce the risk of identifying real data by querying synthetic data. The MDPGAN model can generate synthetic weather data with similar statistical characteristics to real weather data. The data generated by the MDPGAN model meets the requirements of data augmentation and data desensitization at the same time. In this paper, the meteorological data set of Kennedy Airport published by NOAA (National Oceanic and Atmospheric Administration) was used for the experiments of the MDPGAN model. The reliability and validity of generated meteorological data were analyzed and tested. The comparison between the generated data and the real data shows that they have similar statistical characteristics, and the synthetic data has achieved good results in the time series prediction of temperature changes. The MDPGAN model provides a convenient tool for the meteorology researches based on deep learning, which can automatically generate a large amount of safe and reliable data, especially suitable for the meteorology researches with small sample data.