GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training (1805.06725v3)

Abstract: Anomaly detection is a classical problem in computer vision, namely the determination of the normal from the abnormal when datasets are highly biased towards one class (normal) due to the insufficient sample size of the other class (abnormal). While this can be addressed as a supervised learning problem, a significantly more challenging problem is that of detecting the unknown/unseen anomaly case that takes us instead into the space of a one-class, semi-supervised learning paradigm. We introduce such a novel anomaly detection model, by using a conditional generative adversarial network that jointly learns the generation of high-dimensional image space and the inference of latent space. Employing encoder-decoder-encoder sub-networks in the generator network enables the model to map the input image to a lower dimension vector, which is then used to reconstruct the generated output image. The use of the additional encoder network maps this generated image to its latent representation. Minimizing the distance between these images and the latent vectors during training aids in learning the data distribution for the normal samples. As a result, a larger distance metric from this learned data distribution at inference time is indicative of an outlier from that distribution - an anomaly. Experimentation over several benchmark datasets, from varying domains, shows the model efficacy and superiority over previous state-of-the-art approaches.

GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

The paper "GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training" by Samet Akcay, Amir Atapour-Abarghouei, and Toby P. Breckon presents an innovative approach to anomaly detection leveraging the capabilities of Generative Adversarial Networks (GANs). The problem of anomaly detection, where the objective is to identify rare and unusual samples in highly biased datasets dominated by normal instances, is traditionally challenging due to the scarcity of abnormal class data. This research proposes a novel model termed "GANomaly," which adopts a semi-supervised learning paradigm to tackle these challenges.

Methodology

The proposed GANomaly model is centered around an adversarial training approach and utilizes an encoder-decoder-encoder network architecture within the GAN framework. The model consists of three primary sub-networks:

1. Generator Sub-network (G): This comprises an encoder-decoder architecture where the input image is encoded into a low-dimensional latent vector and subsequently reconstructed back into the image space. This process aids the model in learning the distribution of normal samples.
2. Encoder Sub-network (E): This encodes the generated images back into the latent space, ensuring that the latent vectors of the original and generated images (both normal) are similar.
3. Discriminator Sub-network (D): This standard GAN discriminator distinguishes between real and generated images.

The objective function integrates three loss functions to optimize the model:

• Adversarial Loss: Ensures the generator produces images that are indistinguishable from real images.
• Contextual Loss: Encourages the preservation of contextual information, minimizing the L1 distance between the input and generated images.
• Encoder Loss: Minimizes the difference between the latent vectors of the input and generated images.

During inference, the model flags anomalies based on the dissimilarity between the latent vectors of the input and generated images, quantified by the encoder loss.

Experimental Results

The GANomaly model was evaluated on several datasets including MNIST, CIFAR10, the University Baggage Anomaly (UBA) dataset, and the Full Firearm vs. Operational Benign (FFOB) dataset. Key findings include:

• MNIST and CIFAR10: GANomaly outperformed contemporary models such as AnoGAN and EGBAD in terms of Area Under the Curve (AUC) metrics, demonstrating superior capability in both datasets.
• UBA and FFOB: In real-world application datasets, GANomaly achieved higher AUC scores compared to AnoGAN and EGBAD, except in detecting knife anomalies where the performance was comparable. Specifically, GANomaly achieved an AUC of 0.666 in the UBA dataset and 0.882 in the FFOB dataset, underscoring its efficacy in practical anomaly detection tasks.

Implications and Future Work

The implications of this research are substantial for the field of anomaly detection, particularly within operational contexts such as security screening, where the ability to detect rare anomalies efficiently is critical. The GANomaly framework showcases significant advancements in computational efficiency and robustness, offering a practical solution that can be scaled across various domains.

Future work may explore incorporating recent advancements in GAN stability and optimization techniques to further enhance the robustness and generalizability of the model. Additionally, expanding the experimental scope to include more diverse and complex datasets could provide further validation and refinement of the GANomaly framework.

Conclusion

The GANomaly model presents a robust and efficient semi-supervised approach to anomaly detection, leveraging the strengths of adversarial training. Its superior performance across various datasets, both synthetic and real-world, highlights its potential as a leading methodology in anomaly detection tasks. This research paves the way for more sophisticated and adaptable models addressing the challenges of rare anomaly detection in biased datasets.