Stack Multiple Shallow Autoencoders into a Strong One: A New Reconstruction-Based Method to Detect Anomaly

Abstract

Anomaly detection methods based on deep learning typically utilize reconstruction as a proxy task. These methods train a deep model to reconstruct the input from high-level features extracted from the samples. The underlying assumption of these methods is that a deep model trained on normal data would produce higher reconstruction error for abnormal input. But this underlying assumption is not always valid. Because the neural networks have a strong capacity to generalize, the deep model can also reconstruct the unseen abnormal input well sometimes, leading to a not prominent reconstruction error for abnormal input. Hence the decision-making process cannot distinguish the abnormal samples well. In this paper, we stack multiple shallow autoencoders (StackedAE) to enlarge the difference between reconstructions of normal and abnormal inputs. Our architecture feeds the output reconstruction of prior AE into the next one as input. For abnormal input, the iterative reconstruction process would gradually enlarge the reconstruction error. Our goal is to propose a general architecture that can be applied to different data types, e.g., video and image. For video data, we further introduce a weighted loss to emphasize the importance of the center frame and its near neighbors because it is unfair to treat all frames in a 3D convolution frame cuboid equally. To understand the effectiveness of our proposed method, we test on video datasets UCSD-Ped2, CUHK Avenue, and the image dataset MNIST. The results of the experiments demonstrate the effectiveness of our idea.

The research was partly supported by the National Natural Science Foundation of China (No. 61775139), Shanghai Science and Technology Innovation Action Plan (No. 20JC1416503), and Shanghai Key Research Laboratory of NSAI.