Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection

Abstract: Despite inherent ill-definition, anomaly detection is a research endeavor of great interest within machine learning and visual scene understanding alike. Most commonly, anomaly detection is considered as the detection of outliers within a given data distribution based on some measure of normality. The most significant challenge in real-world anomaly detection problems is that available data is highly imbalanced towards normality (i.e. non-anomalous) and contains a most a subset of all possible anomalous samples - hence limiting the use of well-established supervised learning methods. By contrast, we introduce an unsupervised anomaly detection model, trained only on the normal (non-anomalous, plentiful) samples in order to learn the normality distribution of the domain and hence detect abnormality based on deviation from this model. Our proposed approach employs an encoder-decoder convolutional neural network with skip connections to thoroughly capture the multi-scale distribution of the normal data distribution in high-dimensional image space. Furthermore, utilizing an adversarial training scheme for this chosen architecture provides superior reconstruction both within high-dimensional image space and a lower-dimensional latent vector space encoding. Minimizing the reconstruction error metric within both the image and hidden vector spaces during training aids the model to learn the distribution of normality as required. Higher reconstruction metrics during subsequent test and deployment are thus indicative of a deviation from this normal distribution, hence indicative of an anomaly. Experimentation over established anomaly detection benchmarks and challenging real-world datasets, within the context of X-ray security screening, shows the unique promise of such a proposed approach.

• The paper introduces a skip-connected encoder-decoder with adversarial training to accurately model normal data distributions and detect anomalies.
• It combines adversarial, contextual, and latent losses to enhance image reconstruction fidelity and improve detection performance.
• Experimental results on datasets like CIFAR-10 and X-ray screenings show superior performance, achieving AUC scores up to 0.953 in specific cases.

Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection

Introduction and Background

The research paper "Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection" addresses the intricate challenge of anomaly detection, a significant task within the domains of machine learning and visual scene understanding. Unlike typical supervised methods constrained by the lack of diverse anomalous samples, the study introduces an unsupervised approach leveraging an adversarially trained encoder-decoder neural network with skip connections. This method is crafted to precisely capture the normal data distribution, enabling the detection of deviations indicative of anomalies.

Methodology

The cornerstone of the proposed approach is an encoder-decoder architecture augmented with skip connections to efficiently model both high and low-dimensional feature spaces. The encoder down-samples the input image to a compact latent representation, while the decoder reconstructs the image from this latent space. The model employs an adversarial training scheme where the generator (G) and the discriminator (D) networks engage in a minimax game. The generator aims to reconstruct images as close to the originals as possible, while the discriminator distinguishes between real and generated images (Figure 1).

Figure 1: Overview of the proposed adversarial training procedure.

The inclusion of skip connections aids in preserving the spatial and multi-scale information necessary for high-fidelity reconstruction, enhancing the ability to differentiate between normal and anomalous samples. The proposed method uniquely binds adversarial loss with contextual and latent losses to ensure the network learns a robust distribution model.

Experimental Evaluation

The experimental evaluation encompasses standard datasets, including CIFAR-10 and various X-ray screening datasets like UBA and FFOB, demonstrating the model's versatility across different domains. The CIFAR-10 experiments adopt a one-class-vs-rest approach, revealing superior performance over other models such as AnoGAN, EGBAD, and GANomaly, particularly highlighting an AUC of 0.953 for the 'car' anomaly case (Figure 2).

Figure 2: AUC results for CIFAR-10 dataset. Shaded areas in the plot represent variations due to the use of 3 random seeds.

In the challenge of identifying anomalies in X-ray baggage screening datasets, the proposed model outperformed all baseline methods, demonstrating its capability to handle real-world complexity (Figures 5 and 6).

Figure 3: (a) Histogram of the normal and abnormal scores for the test data.

Figure 4: (b) t-SNE plot of 1000 subsampled normal and abnormal features extracted from the last convolutional layer of the discriminator.

The architecture's skip connections and discriminator-based latent feature extraction contribute to numerically superior results, manifesting the model's ability to detect subtle anomalies across diverse test settings.

Discussion and Implications

The deployment of skip connections confers the proposed model with robustness and adeptness in learning both global and local features of the normal class, instrumental in effective anomaly detection. Skip-GANomaly substantially advances the state-of-the-art by exploiting feature extraction from the discriminator, providing a comprehensive scaffold for training unsupervised models in high-stakes applications like security and surveillance.

Conclusion

This study's results highlight the potential of adversarially trained, skip-connected networks for unsupervised anomaly detection tasks, demonstrating significant improvements in detection accuracy. The methodology outlined opens avenues for applying anomaly detection to more intricate datasets and potentially integrating temporal data analysis. Future work should further explore high-resolution imagery and temporal sequences to broaden the effective application of the Skip-GANomaly approach.