Abstract Statistical studies on the number and types of firefighter interventions by region are essential to improve service to the population. It is also a preliminary step if we want to predict these interventions in order to optimize the placement of human and material resources of fire departments, for example. However, this type of data is sensitive and must be treated with the utmost care. In order to avoid any leakage of information, one can think of anonymizing them using Differential Privacy (DP), a safe method by construction. This work focuses on predicting the number of firefighter interventions in certain localities while respecting the strong concept of DP. A local Differential Privacy approach was first used to anonymize location data. Statistical estimators were then applied to reconstruct a synthetic data set that is uncorrelated from the users. Finally, a supervised learning approach using extreme gradient boosting was used to make the predictions. Experiments have shown that the anonymization-prediction method is very accurate: the introduction of noise to sanitize the data does not affect the quality of the predictions, and the predictions faithfully reflect what happened in reality.

中文翻译：

---

使用基于本地差异隐私的数据预测每个地区的消防员干预数量

摘要 按地区对消防员干预的数量和类型进行统计研究对于改善对民众的服务至关重要。例如，如果我们想预测这些干预措施以优化消防部门的人力和物力资源配置，这也是一个初步步骤。然而，这种类型的数据是敏感的，必须非常小心地处理。为了避免任何信息泄露，可以考虑使用差分隐私（DP）来匿名化它们，这是一种安全的构造方法。这项工作的重点是预测某些地区的消防员干预次数，同时尊重 DP 的强大概念。首先使用本地差异隐私方法来匿名化位置数据。然后应用统计估计量来重建与用户不相关的合成数据集。最后，使用使用极端梯度提升的监督学习方法进行预测。实验表明，匿名化预测方法非常准确：引入噪声来净化数据不会影响预测的质量，并且预测忠实地反映了现实中发生的情况。