PP-MTAD: Privacy-Preserving and Efficient Multivariate Time Series Anomaly Detection

Minhua Su, Jia-Nan Liu, Kai He, Jia-Si Weng, Anjia Yang, Xueqiao Liu, Jian Weng

Abstract: Multivariate time series (MTS) data play a crucial role in decision-making across various fields, such as healthcare, industrial control, and economic analysis. However, MTS data often involve substantial sensitive information, and anomalies in such data can directly lead to system failures or decision-making errors, making anomaly detection under privacy protection a crucial task. Existing methods typically prioritize either privacy protection or anomaly detection, struggling to effectively balance the two while incurring significant overhead when handling high-dimensional data. To address these issues, we propose PP-MTAD, a privacy-preserving anomaly detection framework for MTS. The framework designs a series of cryptographic protocols for secure computation, enabling collaboratively confidential anomaly detection. To enhance detection accuracy and efficiency, we leverage Dynamic Time Warping (DTW) distance to measure the similarity between time series, and employ clustering to organize time series into groups, enabling more precise anomaly detection based on distance to cluster centroids. Security analysis demonstrates that our framework effectively ensures data privacy, and experimental results further validate its advantages in accuracy and computational efficiency.

External IDs:dblp:conf/cisc/SuLHWYLW25