2021\useunderine

[1]\fnmWei \surZhu

[2]\fnmDongjin \surSong

1]\orgnameUniversity of Rochester, \orgaddress\stateNY, \countryUSA

2]\orgnameUniversity of Connecticut, \orgaddress\stateCT, \countryUSA

3]\orgnameNEC American Lab, \orgaddress\stateNJ, \countryUSA

We are pleased to submit a new manuscript entitled “Deep Federated Anomaly Detection for Multivariate Time Series Data” by Wei Zhu, Dongjin Song, Yuncong Chen, Wei Cheng, Bo Zong, Takehiko Mizoguchi, Cristian Lumezanu, Haifeng Chen, and Jiebo Luo for possible publication in Machine Learning.

We focus on federated anomaly detection in which multivariate time series data are heterogeneously distributed among different edge devices. We formally define and investigate the problem of federated unsupervised anomaly detection (FedUAD) for multivariate time series data. To handle this task, simply combining federated learning and unsupervised anomaly detection algorithms could lead to a series of problems. One of the main reasons is that each locally trained model may only partially cover the entire normal space. The federated learning algorithms, e.g., federated averaging operation, could produce a global model which may collapse to certain normal modes and even cover a certain part of the abnormal space wang2020federated. As a result, many anomalies could be mistreated as normal status and vice versa. We specially design Fed-ExDNN to overcome these problems. Then, we propose Fed-ExDNN to handle FedUAD, which consists of ExDNN for local anomaly detection and FedCC for model aggregation. On the edge device side, the Exemplar-based Deep Neural Network (ExDNN) can simultaneously learn local time series representations based on their compatibility with an exemplar module which is developed to capture potential normal patterns in the hidden feature space. On the server-side, FedCC could align and aggregate different local exemplar modules. ExDNN and FedCC work jointly to address the heterogeneous problem.

Our empirical studies on six public multivariate time series datasets demonstrate the effectiveness of the proposed ExDNN and Fed-ExDNN. These datasets are all publicly available and are commonly adopted as benchmarks for existing papers. According to the comprehensive experiments, our methods outperform most of the compared methods in terms of different evaluation metrics.

Our paper is closely related to several existing unsupervised anomaly detection zong2018deep and federated learning methods wang2020federated; mcmahan2017communication. However, the proposed method is novel either in anomaly detection or federated learning. Compared with existing anomaly detection methods zong2018deep, the ExDNN is developed based on an advanced clustering algorithm xie2016unsupervised. Moreover, we propose Deep Relation Preserving (DRP) to learn the representation of multivariate time series data in an unsupervised manner. We’d like to emphasize that ExDNN conducts clustering and representation learning simultaneously to mutually boost their performance for better anomaly detection. As for federated learning, we proposed Federated Constraint Clustering to seamlessly incorporate ExDNN for federated learning tasks.

Each of the authors confirms that this manuscript has not been previously published and is not currently under consideration by any other journal or conference. Additionally, all of the authors have approved the contents of this paper and have agreed to Machine Learning’s submission policies.

We believe that this manuscript is appropriate for publication in Machine Learning with the new problem formulation, effective mechanisms, appreciable performance improvements, and a good fit for machine learning topics. We have no conflicts of interest to disclose.