Can Untrained Neural Networks Detect Anomalies?

Anomaly detection (AD) plays a crucial role in identifying unusual data patterns indicative of potential issues or opportunities. Recent data-driven AD models require extensive training for satisfactory performance. This study explores the potential of untrained neural networks (UNNs) for AD tasks. UNNs are used for nonlinear random projection. The anomaly scores are derived from the randomly mapped features using the Mahalanobis distance. We conducted a series of experiments on 12 tabular and two image datasets, comparing the performance of UNNs with 12 established AD models, including state-of-the-art deep learning approaches. Our results demonstrate that UNNs can achieve competitive AD performance without training, which also underscores the importance of training to ensure higher performance beyond the untrained baseline. In addition, the proposed approach offers advantages in terms of time, computational costs, and accessibility, making it a compelling alternative for various applications.