Abstract
Anomaly detection is crucial for the stable operation of mechanical systems, securing financial transactions, and ensuring network security, among other critical areas. Presently, Generative Adversarial Networks (GANs)-based anomaly detection methods either require labeled data for semi-supervised learning or face challenges with low computational efficiency and poor generalization when dealing with complex distributions. Aim to address these limitations, we introduce a generative adversarial synthetic neighbors-based unsupervised anomaly detection (GASN) method. This method integrates generative adversarial networks and neighborhood analysis techniques, enhancing anomaly detection performance through a two-stage detection process. In the first stage, the generative adversarial networks are trained on original dataset that containing a small number of anomaly objects. To minimize errors, the generator focuses on modeling majority object distributions, thus mapping noise to synthetic data resembling normal objects. In the second stage, GASN employs neighborhood analysis techniques to compare the similarity between original and synthetic data, assigning an anomaly factor to each object. This approach allows GASN to sensitively detect subtle anomaly objects. Extensive experiments conducted on twelve public datasets with five state-of-the-art methods demonstrate that the proposed method improves the AUC by 9.93% over the second-best method, proving its effectiveness in anomaly detection.