In the real world, a large number of multivariate time series data are generated by Internet of Things systems, which are composed of many connected sensing devices. Therefore, it is impractical to consider only a single univariate time series for decision-making. High-dimensional time series decrease the performance of traditional anomaly detection methods. Moreover, many previously developed methods capture temporal correlations instead of spatial correlations. Therefore, it is necessary to learn the temporal and spatial correlations between different time series and timestamps. In this paper, to achieve improved anomaly detection performance for multivariate time series, we propose a novel architecture based on a graph attention network (GAT) with multihead dynamic attention (MDA). This framework simultaneously learns the dependencies between sensors in both the temporal and spatial dimensions. To tackle the overfitting problem in autoencoder (AE)-based methods, we propose a hybrid approach that combines a novel generative adversarial network (GAN) architecture as a reconstruction model with a multilayer perceptron (MLP) as a prediction-based model to detect anomalies together. The detection framework proposed in this paper is called the HAD-multihead dynamic GAT (MDGAT). Extensive experiments on different public benchmarks demonstrate the superior performance of HAD-MDGAT over state-of-the-art methods.INDEX TERMS Multivariate time series, graph attention network, anomaly detection, deep generative model, gated recurrent unit.

Federated generative adversarial networks are designed to collaborate across the communication and privacy-constrained edge servers participating in training. However, in the Internet of Things scenario, local updates uploaded by edge servers can lead to the risk of privacy breaches. Gradient-sanitized-based approaches can transmit sanitized sensitive data with strict privacy guarantees, but gradient clipping and perturbation severely degrade convergence performance. In this paper, our proposed algorithm enhances the privacy of terminated raw data through differential privacy before it is transmitted to the edge server. The edge server trains the local generator and discriminator using the perturbed data, which provides privacy guarantees for the gradient attack on the FedGAN without compromising the gradient accuracy. The results of the experimental evaluation show that the algorithm generates images with slightly better quality than that generated by the gradient-sanitized-based approaches while maintaining privacy.