Old is Gold: Redefining the Adversarially Learned One-Class Classifier Training Paradigm (2004.07657v4)

Abstract: A popular method for anomaly detection is to use the generator of an adversarial network to formulate anomaly scores over reconstruction loss of input. Due to the rare occurrence of anomalies, optimizing such networks can be a cumbersome task. Another possible approach is to use both generator and discriminator for anomaly detection. However, attributed to the involvement of adversarial training, this model is often unstable in a way that the performance fluctuates drastically with each training step. In this study, we propose a framework that effectively generates stable results across a wide range of training steps and allows us to use both the generator and the discriminator of an adversarial model for efficient and robust anomaly detection. Our approach transforms the fundamental role of a discriminator from identifying real and fake data to distinguishing between good and bad quality reconstructions. To this end, we prepare training examples for the good quality reconstruction by employing the current generator, whereas poor quality examples are obtained by utilizing an old state of the same generator. This way, the discriminator learns to detect subtle distortions that often appear in reconstructions of the anomaly inputs. Extensive experiments performed on Caltech-256 and MNIST image datasets for novelty detection show superior results. Furthermore, on UCSD Ped2 video dataset for anomaly detection, our model achieves a frame-level AUC of 98.1%, surpassing recent state-of-the-art methods.

• The paper redefines the discriminator's role by training it to evaluate reconstruction quality instead of distinguishing between real and fake data.
• The paper employs a dual-phase training process using an earlier generator to stabilize GAN training and enhance anomaly detection performance.
• The paper introduces a pseudo-anomaly module that generates synthetic anomalies, significantly improving one-class classifier robustness.

Analysis of "Old is $\mathbf{\mathcal{G}^{old}$: Redefining the Adversarially Learned One-Class Classifier Training Paradigm"

The paper "Old is $\mathbf{\mathcal{G}^{old}$: Redefining the Adversarially Learned One-Class Classifier Training Paradigm" by Zaheer et al. addresses a significant challenge in anomaly detection using Generative Adversarial Networks (GANs). Specifically, it proposes a novel method to stabilize the training of one-class classifiers by altering the role of GAN components, markedly improving the detection of anomalies.

Key Contributions and Methodology

1. Redefinition of the Discriminator Role: The central novelty of the paper lies in modifying the discriminator's typical role in GANs. Instead of distinguishing real from fake data, the discriminator is trained to differentiate between high and low-quality reconstructions produced by the generator. This shift allows the discriminator to more effectively identify anomalies based on reconstruction quality differences.
2. Generator Stabilization Through Dual-Phase Training: The paper introduces a two-phase training approach:
   • Phase One: Standard adversarial training involving both the generator and discriminator. The generator is trained to minimize reconstruction errors while the discriminator distinguishes normal from poorly reconstructed data.
   • Phase Two: Focuses on optimizing the discriminator's ability to discern reconstruction quality. Here, low-quality examples are generated using a fixed, earlier version of the generator ($\mathcal{G}^{old}$), allowing enhanced training stability and anomaly detector robustness.
3. Pseudo-Anomaly Module: This module generates synthetic anomalies that mimic the appearance of reconstructed anomalous inputs. By training the discriminator with these synthetic anomalies, the network becomes adept at recognizing those subtle deviations indicative of real anomalies.

Experimental Validation

The proposed framework exhibits prominent performance improvements and consistency across various image and video datasets:

• Caltech-256 and MNIST: The model outperforms previous methods on standard benchmarks for outlier detection, demonstrating robustness across various settings with differing ratios of outliers.
• UCSD Ped2: The method achieves a frame-level Area Under the Curve (AUC) of 98.1%, surpassing state-of-the-art techniques. This performance gain is attributed to the reduction of the typical instability observed in standard GAN training due to the proposed structural modifications in discriminator training (such as utilizing pseudo-anomaly data).

Implications and Future Directions

The introduction of a discriminator trained on quality reconstruction paradigms heralds a significant advancement in one-class classification problems within anomaly detection. This approach illustrates how GANs can be tailored to improve detection rates while stabilizing on typically complex and unstable training regimes. The proposed methodology shows promise not only in static image data but also in dynamic video contexts where temporal coherence and anomaly detection are paramount.

Future work could expand upon this framework by exploring different architectures for $\mathcal{G}^{old}$ generation, potentially leveraging advancements in ensemble methods or dynamic training adjustments that account for varying data distribution during training. Additionally, the integration of more complex synthetic anomaly generation strategies may further enhance anomaly detection capabilities in diverse applications beyond traditional image and video data.

Conclusion

This paper contributes a substantive methodological advancement in the field of anomaly detection using GANs. By redefining how generators and discriminators interact within one-class classifiers, the authors provide a pathway for more robust anomaly detection tools, potentially transforming applications in security, surveillance, and beyond.