ARAE: Adversarially robust training of autoencoders improves novelty detection

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Autoencoders (AEs) are commonly used for novelty detection, assuming they reconstruct normal data well but fail on anomalies.
• This assumption is often violated as AEs can reconstruct anomalous data by learning low-level features.

Purpose of the Study:

• To propose a novel training algorithm for AEs to enhance novelty detection capabilities.
• To address the limitation of AEs reconstructing anomalous data by promoting the learning of semantically meaningful features.

Main Methods:

• Developed a new AE training algorithm that encourages learning significant features.
• Incorporated adversarial robustness by stabilizing the AE's bottleneck layer against perturbations.
• The approach is general and applicable to various AE-based methods.

Main Results:

• The proposed AE method significantly improves novelty detection performance.
• Outperforms or is competitive with state-of-the-art methods on four benchmark and two medical datasets.
• Achieved these results with a simpler architecture compared to existing approaches.

Conclusions:

• The novel training algorithm effectively enhances AE-based novelty detection.
• Adversarial robustness is a key factor in learning semantically meaningful features for anomaly detection.
• The method offers a more robust and efficient solution for identifying novel data points.