计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (11): 60-65.DOI: 10.3778/j.issn.1002-8331.2108-0057

• 理论与研发 • 上一篇    下一篇

面向多元时序数据的个性化联邦异常检测方法

王昊天,郑栋毅,刘芳,肖侬   

  1. 1.国防科技大学,长沙 410073
    2.湖南大学,长沙 410006
  • 出版日期:2022-06-01 发布日期:2022-06-01

Personalized Federated Anomaly Detection Method for Multivariate Time Series Data

WANG Haotian, ZHENG Dongyi, LIU Fang, XIAO Nong   

  1. 1.National University of Defense Technology, Changsha 410073, China
    2.Hunan University, Changsha 410006, China
  • Online:2022-06-01 Published:2022-06-01

摘要: 随着实时传感器在诸如机场、发电厂、智能工厂和医疗保健系统等各种领域的广泛运用,对多变量时间序列数据的异常检测变得更加重要。然而,目前面临两个关键的挑战。数据机构的敏感数据通常以孤岛的形式存在,这使得在保护隐私安全的前提下难以融合数据,无法训练出高性能的异常检测模型。不同数据机构的数据存在统计异构性,在个性化数据场景下,使用统一的异常检测模型的性能不佳。提出了一种面向多元时序数据的个性化联邦异常检测框架FedPAD(federated personalized anomaly detection)。FedPAD基于联邦学习架构,在保护隐私的前提下进行数据聚合,通过微调构建相对个性化的模型。在NASA航天器数据集上的实验表明,FedPAD能够实现准确和个性化的异常检测,相比于基准方法F1分数平均提高了6.9%。

关键词: 多元时序数据, 异常检测, 联邦学习, 个性化

Abstract: With the wide application of real-time sensors in various fields such as airports, power plants, intelligent factories and health care systems, anomaly detection of multivariate time series data becomes more important. However, there are currently two key challenges. First of all, the sensitive data of data institutions usually exist in the form of isolated islands, which makes it difficult to fuse data under the premise of protecting privacy and can not train high-performance anomaly detection models. Secondly, the data of different data institutions are statistically heterogeneous. In the personalized data scenario, the performance of using a unified anomaly detection model is poor. This paper proposes a personalized federal anomaly detection framework FedPAD(federated personalized anomaly detection) for multivariate time series data. Based on the federal learning architecture, FedPAD aggregates data under the premise of protecting privacy, and then constructs a relatively personalized model by fine tuning. Experiments on NASA spacecraft dataset show that FedPAD can achieve accurate and personalized anomaly detection, and the average F1 score is increased by 6.9% compared with the benchmark method.

Key words: multivariable time series, anomaly detection, federated learning, personalization