Scholarlink

<p>Anomaly detection in multivariate time series is critical for system status monitoring, particularly in healthcare surveillance. However, centralized anomaly detection in cloud-edge collaborative environments may raise privacy concerns. This paper proposes a Federated Aggregation Optimization-based Privacy-Preserving Anomaly Detection Framework for Time Series (FAPAD). The FAPAD framework mitigates the risk of privacy leakage by distributing data anomaly detection tasks across multiple edge nodes and deploying model anomaly detection and parameter aggregation modules on the cloud. Specifically, we integrate a deep anomaly detection method for multivariate time-series data in the anomaly detection module to ensure detection performance. This method employs an adversarial Transformer feature learning model to capture the feature representations of the time-series data. Anomaly scores and decision thresholds are then calculated using an anomaly interpretation method to enable effective differentiation between normal and anomalous data. Furthermore, the differential privacy technique is employed to protect the privacy of both the model and data during the parameter upload phase, thus preventing model inference attacks. In the parameter aggregation module, a federated aggregation optimization strategy is designed, which combines the <i>F</i>1 score and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(L_2\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mn>2</mn> </msub> </math></EquationSource> </InlineEquation> distance to assign aggregation weights to local model parameters, thereby reducing the impact of anomalous models on the global aggregation outcome. Extensive experiments conducted on four public datasets demonstrate the effectiveness of the proposed FAPAD framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FAPAD: a federated aggregation optimization-based privacy-preserving anomaly detection framework for time series

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Kejun Zhang,
  • Xinying Yu,
  • Jun Wang,
  • Wenbin Wang,
  • Meijiao Li,
  • Yucheng Yan,
  • Ying Zhang

摘要

Anomaly detection in multivariate time series is critical for system status monitoring, particularly in healthcare surveillance. However, centralized anomaly detection in cloud-edge collaborative environments may raise privacy concerns. This paper proposes a Federated Aggregation Optimization-based Privacy-Preserving Anomaly Detection Framework for Time Series (FAPAD). The FAPAD framework mitigates the risk of privacy leakage by distributing data anomaly detection tasks across multiple edge nodes and deploying model anomaly detection and parameter aggregation modules on the cloud. Specifically, we integrate a deep anomaly detection method for multivariate time-series data in the anomaly detection module to ensure detection performance. This method employs an adversarial Transformer feature learning model to capture the feature representations of the time-series data. Anomaly scores and decision thresholds are then calculated using an anomaly interpretation method to enable effective differentiation between normal and anomalous data. Furthermore, the differential privacy technique is employed to protect the privacy of both the model and data during the parameter upload phase, thus preventing model inference attacks. In the parameter aggregation module, a federated aggregation optimization strategy is designed, which combines the F1 score and \(L_2\) L 2 distance to assign aggregation weights to local model parameters, thereby reducing the impact of anomalous models on the global aggregation outcome. Extensive experiments conducted on four public datasets demonstrate the effectiveness of the proposed FAPAD framework.