Backpropagated Gradient Representations for Anomaly Detection

Abstract

Learning representations that clearly distinguish between normal and abnormal data is key to the success of anomaly detection. Most of existing anomaly detection algorithms use activation representations from forward propagation while not exploiting gradients from backpropagation to characterize data. Gradients capture model updates required to represent data. Anomalies require more drastic model updates to fully represent them compared to normal data. Hence, we propose the utilization of backpropagated gradients as representations to characterize model behavior on anomalies and, consequently, detect such anomalies. We show that the proposed method using gradient-based representations achieves state-of-the-art anomaly detection performance in benchmark image recognition datasets. Also, we highlight the computational efficiency and the simplicity of the proposed method in comparison with other state-of-the-art methods relying on adversarial networks or autoregressive models, which require at least 27 times more model parameters than the proposed method.

A Unifying Review of Deep and Shallow Anomaly Detection

“…By adopting an adversarial loss to regularize and match the latent encoding distribution, AAEs can employ any arbitrary prior p(z), as long as sampling is feasible. Finally, other AE variants that have been applied to AD include RNN-based AEs [194], [231], [397], [398], convolutional AEs [54], AE ensembles [126], [398], and variants that constrain the gradients [399] or actively control the latent code topology [400] of an AE. AEs also have been utilized in two-step approaches that use AEs for dimensionality reduction and apply traditional methods on the learned embeddings [136], [401], [402].…”

CVAD: A generic medical anomaly detector based on Cascade VAE

“…Besides, there are many other approaches contributing to OOD detection, such as GradCon [17], generalized ODIN [11] and FSSD [13]. Please refer to the papers for more details.…”

Open-Set Recognition With Gradient-Based Representations

“…Apart from their original utility as a tool to search for a converged solution, gradients have been utilized for various purposes, including visualization [14,15,16] and adversarial attack generation [3,17]. Gradients have also been explored to obtain effective representations [18,19,20,21,12] for many applications including image quality and saliency estimation, and out-of-distribution/anomaly/novelty detection. However, the effectiveness of gradient-based representations has not been fully explored in the application of open-set recognition.…”