Anomaly detection of multivariate time series (MTS) is crucial in industrial intelligent systems. To address the challenges of absence of anomaly labels, fast inference time, multi-source and multi-modality in anomaly detection, researchers have primarily investigated unsupervised reconstruction-driven methods. However, the existing reconstruction-driven methods mainly focus on minimizing reconstruction errors while neglecting the importance of training methods that increase errors between normal and abnormal classes. Furthermore, accurately constructing the feature space of normal and abnormal classes during the reconstruction process remains a challenge. In this paper, we propose an innovative model, namely the confidence adversarial autoencoder (CAAE). The proposed CAAE combines a confidence network, based on window credibility judgment, with an autoencoder to provide credibility support for anomaly detection. We further introduce fake labels to provide the confidence network with a discriminative knowledge for identifying reconstructed data. Additionally, we implement the confidence adversarial training method to generate fake labels to construct an adversarial loss aiming to expand the decision boundary of anomaly scores. Extensive experimental results on publicly available time series datasets are provided to demonstrate the efficiency of our proposed CAAE. It reveals that excellent generalization ability and superior average performance are achieved on different datasets compared with the state-of-the-art methods.