Anomalous Pattern Detection in Activations and Reconstruction Error of Autoencoders

Celia Cintas, Skyler Speakman, Victor Akinwande, Srihari Sridharan, William Ogallo, Edward McFowland III

TL;DR: Unsupervised method to detect adversarial samples in autoencoder's activations and reconstruction error space

Abstract:

In real-world machine learning applications, large outliers and pervasive noise are commonplace, and access to clean training data as required by standard deep autoencoders is unlikely.
Reliably detecting anomalies in a given set of images is a task of high practical relevance for visual quality inspection, surveillance, or medical image analysis. Autoencoder neural networks learn to reconstruct normal images, and hence can classify those images as anomalous if the reconstruction error exceeds some threshold. In this paper, we proposed an unsupervised method based on subset scanning over autoencoder activations. The contributions of our work are threefold. First, we propose a novel method combining detection with reconstruction error and subset scanning scores to improve the anomaly score of current autoencoders without requiring any retraining. Second, we provide the ability to inspect and visualize the set of anomalous nodes in the reconstruction error space that make a sample noised. Third, we show that subset scanning can be used for anomaly detection in the inner layers of the autoencoder. We provide detection power results for several untargeted adversarial noise models under standard datasets.

Code: https://github.com/usersubsetscan/autoencoder_anomaly_subset

Keywords: unsupervised anomaly detection, adversarial attacks, autoencoders, subset scanning