[Paper Review] Memory Augmented Generative Adversarial Networks for Anomaly Detection

Ziyi Yang, Teng Zhang+2 more|arXiv (Cornell University)|Feb 7, 2020

Anomaly Detection Techniques and Applications4 citations

TL;DR

This paper proposes Memory Augmented Generative Adversarial Networks (MEMGAN), a novel anomaly detection framework that uses an external memory module to enhance GAN-based reconstruction and detection. By enforcing encoded normal data to lie within the convex hull of memory units—via a memory projection loss and mutual information regularization—MEMGAN achieves superior anomaly detection performance, with disentangled and interpretable memory units and state-of-the-art results on MNIST and CIFAR-10 datasets.

ABSTRACT

In this paper, we present a memory-augmented algorithm for anomaly detection. Classical anomaly detection algorithms focus on learning to model and generate normal data, but typically guarantees for detecting anomalous data are weak. The proposed Memory Augmented Generative Adversarial Networks (MEMGAN) interacts with a memory module for both the encoding and generation processes. Our algorithm is such that most of the extit{encoded} normal data are inside the convex hull of the memory units, while the abnormal data are isolated outside. Such a remarkable property leads to good (resp.\ poor) reconstruction for normal (resp.\ abnormal) data and therefore provides a strong guarantee for anomaly detection. Decoded memory units in MEMGAN are more interpretable and disentangled than previous methods, which further demonstrates the effectiveness of the memory mechanism. Experimental results on twenty anomaly detection datasets of CIFAR-10 and MNIST show that MEMGAN demonstrates significant improvements over previous anomaly detection methods.

Motivation & Objective

• To address the weak theoretical and empirical guarantees in existing deep generative models for anomaly detection, which often rely on reconstruction heuristics without explicit design for anomaly separation.
• To improve the interpretability and disentanglement of memory units in memory-augmented models by structuring them to form the boundary of a convex hull enclosing normal data.
• To provide a theoretically grounded mechanism where normal data are enclosed within the convex hull of memory units, while abnormal data lie outside, enabling strong detection guarantees.
• To empirically validate MEMGAN’s superiority over state-of-the-art methods, including DSVDD, ADGAN, and MEMAE, across diverse benchmark datasets.
• To analyze the robustness and sensitivity of MEMGAN to hyperparameters such as the number of memory units, ensuring practical deployability.

Proposed method

• MEMGAN employs a bidirectional GAN architecture with a generator $G$, encoder $E$, and discriminator $D_{{m{x}}{m{z}}}$, extended with an external memory module storing features of normal data.
• The model introduces a memory projection loss that ensures encoded normal data reside within the convex hull of memory units, while memory units are pushed to the boundary of this hull.
• A mutual information loss is incorporated to encourage disentangled and interpretable memory representations by maximizing information between memory units and encoded data.
• The memory module is updated via a differentiable attention mechanism, allowing gradient flow during end-to-end training.
• During inference, the decoder reconstructs data using weighted combinations of memory units, with anomaly scores derived from reconstruction error.
• Theoretical analysis shows that under proper hyperparameter settings, the optimal memory units form the vertices of a convex polytope enclosing the support of encoded normal data.

Experimental results

Research questions

• RQ1 How can memory-augmented GANs be designed to provide stronger theoretical guarantees for anomaly detection compared to standard autoencoder or GAN-based methods?
• RQ2 What is the geometric relationship between encoded normal data and memory units in a memory-augmented GAN, and how can it be enforced during training?
• RQ3 To what extent does the memory projection loss improve the disentanglement and interpretability of memory units compared to prior memory-augmented models?
• RQ4 How does MEMGAN’s performance compare to state-of-the-art methods like DSVDD and ADGAN across diverse datasets such as MNIST and CIFAR-10?
• RQ5 Is MEMGAN robust to variations in the number of memory units, and how sensitive is its performance to hyperparameter choices?

Key findings

• MEMGAN achieves the highest mean AUROC on seven out of ten MNIST digit classes and the best average AUROC across all ten classes, significantly outperforming MEMAE and DSVDD.
• On CIFAR-10, MEMGAN achieves performance on par with DSVDD and superior to ADGAN and ALAD, with the highest performance in three out of ten classes and the best average AUROC.
• Ablation studies confirm that the memory projection loss is critical: without it, memory units cluster in dense regions and fail to form a proper convex hull, degrading detection performance.
• MEMGAN demonstrates robustness to the number of memory units, with AUROC remaining stable across 25 to 200 units on MNIST, indicating insensitivity to this hyperparameter.
• Visualization of decoded memory units shows significantly higher quality and disentangled features compared to previous memory-augmented models, supporting the interpretability claim.
• Theoretical analysis confirms that the support of encoded normal data forms a convex polytope, and the optimal memory units are its vertices, providing a geometric foundation for strong anomaly detection guarantees.