Spot The Odd One Out: Regularized Complete Cycle Consistent Anomaly Detector GAN

TL;DR

RCALAD tackles high variance in anomaly detection by enforcing complete cycle consistency with a joint discriminator $D_{xxzz}$ and biasing reconstructions toward the normal data manifold via a supplementary distribution $\sigma( ext{x})$. Two anomaly scores, $A_{fm}$ and $A_{all}$, leverage joint and individual discriminator signals to robustly separate anomalies from normal data across tabular and image domains. Empirical results on six datasets show improved accuracy and reduced class-wise variance, with statistical tests supporting significance on several benchmarks. The work provides open-source code and introduces a principled framework for regularized cycle-consistent anomaly detection in GANs.

Abstract

This study presents an adversarial method for anomaly detection in real-world applications, leveraging the power of generative adversarial neural networks (GANs) through cycle consistency in reconstruction error. Previous methods suffer from the high variance between class-wise accuracy which leads to not being applicable for all types of anomalies. The proposed method named RCALAD tries to solve this problem by introducing a novel discriminator to the structure, which results in a more efficient training process. Additionally, RCALAD employs a supplementary distribution in the input space to steer reconstructions toward the normal data distribution, effectively separating anomalous samples from their reconstructions and facilitating more accurate anomaly detection. To further enhance the performance of the model, two novel anomaly scores are introduced. The proposed model has been thoroughly evaluated through extensive experiments on six various datasets, yielding results that demonstrate its superiority over existing state-of-the-art models. The code is readily available to the research community at https://github.com/zahraDehghanian97/RCALAD.

Figures (5)

• Figure 1: The information of complete cycle consistency in the proposed model (left side) and using the variables of input data space and latent space in the cycle consistency of the ALAD network (right side).
• Figure 2: Effect of the presence of supplementary distribution in the model training process. These figures indicate the trained models. On the left side, there is no supplementary distribution in the training phase, and on the right side, we had samples from supplementary distribution in training. In this figure, X represents the input data space and Z represents the latent space. Samples are mapped from the latent space to the input data space by the generator $G$, and the encoder $E$ reverses the mapping. Green circles show normal samples, red crosses represent abnormal samples, and blue stars represent samples generated by the supplementary distribution $\sigma(\mathrm{x})$. The turquoise-colored line shows the value of the abnormality score. As can be seen in Figure \ref{['fig2']}, if samples from the supplementary distribution $\sigma(\mathrm{x})$ are not present (on the left side of the figure), the abnormality score for the abnormal sample is lower than when these samples(blue stars) are used used (right side). In other words, when there is no supplementary distribution in the training process, after training, the model may reconstruct anomalous samples in a good way, which leads to a small difference between anomalous sample and its reconstruction, and as a result, we will have a low anomaly score, but on the right, the samples from supplementary distribution have biased the model towards the normal manifold. In this case, the model tries to reconstruct an anomalous sample near the normal data manifold, which is far from the anomalous data manifold that leads to a high anomaly score for abnormal data.
• Figure 3: Overall structure of the RCALAD model.
• Figure 4: Reconstruction of normal and abnormal inputs on CIFAR-10 dataset. The first row is normal inputs and the second row is their reconstruction. The third row are anomalies and the fourth row is corresponding reconstruction for anomalies.
• Figure 5: Reconstruction of normal and abnormal inputs on SVHN dataset. The first row is normal inputs and the second row is their reconstruction. The third row are anomalies and the fourth row is corresponding reconstruction for anomalies.