Back

GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

Samet AkcayAmir Atapour-AbarghoueiToby P. Breckon

article2018Asian Conference on Computer Vision1,773 citations

Presents an encoder-decoder-encoder conditional generative adversarial network that identifies unseen visual anomalies by measuring reconstruction discrepancies across both image and latent feature spaces.

Listen

In many critical security, biomedical, and industrial applications, identifying rare and unexpected threats is essential. Traditional automated inspection tools rely on large pools of labeled threat examples to train supervised models. However, in operational environments such as aviation and border checkpoint X-ray screening, dangerous objects occur rarely and continuously evolve, making comprehensive threat datasets unobtainable. While human operators adapt well to novel anomalies, automated systems struggle to flag items they have never seen before.

The article demonstrates an effective semi-supervised anomaly detection framework called GANomaly. The primary objective is to reliably detect unseen abnormal items by training solely on normal, non-anomalous imagery.

To achieve this, the authors designed a novel neural network architecture combining an encoder-decoder-encoder pipeline with an adversarial discriminator. The system compresses an input image into a compact vector representation, reconstructs the image, and then compresses that reconstructed image a second time. By training only on normal images, the model optimizes three joint losses: image realism, image contextual similarity, and latent feature distance. Because the model learns only the features of normal data, abnormal test images produce large discrepancies in the latent vector space, which serves as a direct anomaly score. The approach was evaluated on standard image benchmarks (MNIST and CIFAR-10) and two operational security datasets: the University Baggage Anomaly dataset (over 230,000 patches) and the UK government Full Firearm vs. Operational Benign dataset (over 72,000 full baggage scans).

The evaluation yielded several key findings. First, the proposed approach outperformed existing state-of-the-art anomaly detection models across benchmark and security datasets. On the full firearm security screening dataset, it achieved an Area Under the Curve (AUC) of 0.882, compared to 0.703 and 0.712 for prior generative adversarial models. Second, the model demonstrated substantial computational gains, running in roughly 2.5 to 2.8 milliseconds per image—more than three times faster than competing joint-training models and thousands of times faster than iterative optimization baselines. Third, feature analysis confirmed that the dual-encoder discrepancy effectively separates normal from abnormal samples even when visual reconstructions appear ambiguous.

These findings indicate that high-throughput screening operations can deploy effective automated threat screening without requiring exhaustive catalogs of dangerous items. The significant improvement in processing speed allows the model to perform real-time anomaly detection at operational line speeds without introducing bottlenecks. Furthermore, the framework lowers operational risks and data collection costs by eliminating the need to capture and label thousands of physical threat examples.

Organizations evaluating automated security inspection should consider piloting this architecture for screening workflows that encounter rare or evolving threats. Future development should incorporate modern adversarial training optimizations to further refine detection accuracy across complex threat categories.

A recognized limitation is that simple shapes, such as isolated knives, can result in higher false-positive rates due to model overfitting. Additionally, performance slightly declines when normal and abnormal classes share substantial visual overlap. Overall confidence in the system's operational viability remains high given its consistent statistical superiority and millisecond-level inference speed across extensive real-world security screening datasets.

Durham Universityanomaly detection methodscomputer visiongenerative adversarial networksGANomalysemi-supervised anomaly detectionadversarial training
arXiv: 1805.06725samet-akcay/ganomaly
  • Paper: Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery, Thomas Schlegl et al. (2017). This paper establishes the foundational paradigm of using generative adversarial networks and latent space mapping for unsupervised image anomaly detection that GANomaly directly builds upon and improves.
  • Paper: Adversarial Feature Learning, Jeff Donahue et al. (2016). It introduces bidirectional GANs for joint image generation and latent space inference, providing the core encoder-generator formulation leveraged by GANomaly.
  • Paper: Adversarially Learned Inference, Vincent Dumoulin et al. (2017). It details adversarially learned inference for mapping between data space and latent representations, establishing theoretical foundations for GANomaly's dual latent-data optimization.
  • Paper: Autoencoding beyond pixels using a learned similarity metric, Anders Boesen Lindbo Larsen et al. (2015). It demonstrates combining autoencoders and adversarial networks using learned discriminator feature representations to evaluate reconstruction fidelity rather than relying solely on pixel distances.
  • Paper: Conditional Generative Adversarial Nets, Mehdi Mirza et al. (2014). It introduces conditional generative adversarial networks, which provide the underlying conditional generation framework used by GANomaly to reconstruct images.
  • Paper: Generative Adversarial Networks, Ian J. Goodfellow et al. (2014). It introduces the fundamental minimax game and adversarial training framework underlying all GAN-based generative modeling.
  • Paper: Deep One-Class Classification, Lukas Ruff et al. (2018). It formalizes deep one-class classification principles and benchmarks for semi-supervised and unsupervised anomaly detection.
Cover for GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

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.

Topics

anomaly detection methodscomputer visiongenerative adversarial networksinferencesemi-supervised learningsupervised machine learningunsupervised learningdetection algorithmnetwork architecturelatent factor

Keywords

GANomalysemi-supervised anomaly detectionadversarial trainingadversarial autoencodernovelty detectionadversarial lossrepresentation learningvariational autoencoder

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Our Approach: GANomaly
  • Generative Adversarial Networks (GAN)
  • Adversarial Auto-Encoders (AAE)
  • GAN with Inference
  • 3.1 Proposed Approach
  • Problem Definition.
  • Ganomaly Pipeline.
  • 3.2 Model Training
  • Adversarial Loss.
  • Contextual Loss.
  • Encoder Loss.
  • 3.3 Model Testing
  • 4 Experimental Setup
  • MNIST.
  • CIFAR10.
  • University Baggage Anomaly Dataset — (UBA).
  • Full Firearm vs. Operational Benign — (FFOB).
  • 5 Results
  • 6 Conclusion
  • References

Knowls

  1. Knowl 1 — GANomaly Architecture

    model/method

    GANomaly is a semi-supervised anomaly detection framework based on a conditional generative adversarial network featuring an encoder-decoder-encoder pipeline alongside a discriminator. The architecture consists of three core sub-networks:

    1. Generator Autoencoder (GG): Composed of an encoder network GEG_E and a decoder network GDG_D. Given an input image x∈Rw×h×cx \in \mathbb{R}^{w \times h \times c} (with width ww, height hh, and cc channels), GEG_E applies convolutional layers with batch normalization and LeakyReLU activations to downscale and compress xx into a lower-dimensional bottleneck latent representation z=GE(x)∈Rdz = G_E(x) \in \mathbb{R}^d. The decoder GDG_D uses convolutional transpose layers, ReLU activations, batch normalization, and a final tanh⁡\tanh activation layer to upscale zz into a reconstructed image x^=GD(z)=GD(GE(x))∈Rw×h×c\hat{x} = G_D(z) = G_D(G_E(x)) \in \mathbb{R}^{w \times h \times c}.

    2. Latent Encoder (EE): Has identical layer architecture to GEG_E but with independently parameterized weights. It maps the reconstructed image x^\hat{x} back into the latent vector space to produce an encoded latent vector z^=E(x^)=E(G(x))∈Rd\hat{z} = E(\hat{x}) = E(G(x)) \in \mathbb{R}^d.

    3. Discriminator (DD): A Deep Convolutional GAN (DCGAN) classification network that accepts an image (xx or x^\hat{x}) to predict whether it is real or generated, and outputs intermediate feature layer activations f(⋅)f(\cdot) used for feature-matching loss.

    Training is performed exclusively on normal (non-anomalous) images so that the model learns the manifold of normal data. When presented with anomalous patterns at test time, the generator fails to faithfully reconstruct the anomalous features, leading to a large discrepancy between the input embedding zz and the re-encoded latent vector z^\hat{z}.

  2. Knowl 2 — GANomaly Multi-Objective Training Loss

    equation

    The GANomaly generator network G=GD(GE(⋅))G = G_D(G_E(\cdot)) and auxiliary latent encoder EE are trained jointly using a weighted composite loss function combining adversarial, contextual, and latent encoding objectives:

    L=wadvLadv+wconLcon+wencLenc\mathcal{L} = w_{adv}\mathcal{L}_{adv} + w_{con}\mathcal{L}_{con} + w_{enc}\mathcal{L}_{enc}

    where wadvw_{adv}, wconw_{con}, and wencw_{enc} are hyperparameter weights balancing the individual loss contributions (empirically set to wadv=1w_{adv} = 1, wcon=50w_{con} = 50, and wenc=1w_{enc} = 1).

    The individual loss components are defined as:

    • Adversarial Loss (Feature Matching) Ladv\mathcal{L}_{adv}: Measures the L2L_2 distance between intermediate discriminator feature representations f(⋅)f(\cdot) of real images x∼pXx \sim p_X and generated reconstructions G(x)G(x) to stabilize training:

    Ladv=Ex∼pX∥f(x)−f(G(x))∥2\mathcal{L}_{adv} = \mathbb{E}_{x \sim p_X} \|f(x) - f(G(x))\|_2

    • Contextual Loss Lcon\mathcal{L}_{con}: Measures the pixel-level L1L_1 reconstruction distance between the input image xx and the reconstructed image G(x)G(x) to preserve contextual structure without excessive blurring:

    Lcon=Ex∼pX∥x−G(x)∥1\mathcal{L}_{con} = \mathbb{E}_{x \sim p_X} \|x - G(x)\|_1

    • Latent Encoder Loss Lenc\mathcal{L}_{enc}: Measures the L2L_2 distance between the bottleneck latent vector GE(x)G_E(x) and the re-encoded latent vector E(G(x))E(G(x)) derived from the reconstructed image:

    Lenc=Ex∼pX∥GE(x)−E(G(x))∥2\mathcal{L}_{enc} = \mathbb{E}_{x \sim p_X} \|G_E(x) - E(G(x))\|_2

  3. Knowl 3 — GANomaly Inference and Anomaly Scoring Mechanism

    model/method

    During inference, an anomaly score is computed for a given test image x^∈Rw×h×c\hat{x} \in \mathbb{R}^{w \times h \times c} by measuring the L1L_1 distance between its bottleneck latent projection GE(x^)G_E(\hat{x}) and the re-encoded latent vector E(G(x^))E(G(\hat{x})) of the reconstructed image:

    A(x^)=∥GE(x^)−E(G(x^))∥1\mathcal{A}(\hat{x}) = \|G_E(\hat{x}) - E(G(\hat{x}))\|_1

    For a test dataset D^={x^1,x^2,…,x^N}\hat{\mathcal{D}} = \{\hat{x}_1, \hat{x}_2, \dots, \hat{x}_N\} yielding raw anomaly scores S={si=A(x^i)∣x^i∈D^}S = \{s_i = \mathcal{A}(\hat{x}_i) \mid \hat{x}_i \in \hat{\mathcal{D}}\}, the scores are scaled to the range [0,1][0, 1] via min-max normalization:

    si′=si−min⁡(S)max⁡(S)−min⁡(S)s'_i = \frac{s_i - \min(S)}{\max(S) - \min(S)}

    A test image is classified as anomalous if its anomaly score exceeds a chosen threshold ϕ\phi (i.e., A(x^)>ϕ\mathcal{A}(\hat{x}) > \phi or si′>ϕs'_i > \phi). Because GG and EE are trained only on normal data, anomalous input features cannot be effectively reconstructed or re-encoded, generating high discrepancy in the latent space.

  4. Knowl 4 — Anomaly Detection Performance on X-ray Security Screening Datasets

    data/table

    The anomaly detection effectiveness of GANomaly was evaluated against AnoGAN and EGBAD on two security X-ray datasets: the University Baggage Anomaly dataset (UBA; 230,275 sliding window 64×6464 \times 64 patches with 122,803 anomalies split into gun, gun-parts, and knife sub-classes) and the UK Government Full Firearm vs. Operational Benign dataset (FFOB; 67,672 normal operational scans and 4,680 concealed firearm threats resized to 64×6464 \times 64). Performance is measured by the Area Under the Receiver Operating Characteristic Curve (AUC):

    Method UBA FFOB
    Gun Gun-parts Knife Overall Full-weapon
    AnoGAN 0.598 0.511 0.599 0.569 0.703
    EGBAD 0.614 0.591 0.587 0.597 0.712
    GANomaly 0.747 0.662 0.520 0.643 0.882

    GANomaly outperforms both AnoGAN and EGBAD substantially on the firearm categories (UBA gun: 0.747 vs. 0.614; UBA gun-parts: 0.662 vs. 0.591; FFOB full-weapon: 0.882 vs. 0.712) and achieves the best overall AUC on both benchmarks. On the knife subclass, GANomaly obtains lower AUC (0.520) due to the geometric simplicity of knife shapes causing model overfitting and higher false-positive rates.

  5. Knowl 5 — Inference Runtime Performance Across GAN-Based Anomaly Detectors

    data/table

    Computational inference efficiency was measured in milliseconds (ms) per test image on an Intel Xeon E5-2630 v4 CPU with an NVIDIA GTX Titan X GPU across four evaluation benchmarks (MNIST at 32×3232 \times 32, CIFAR-10 at 32×3232 \times 32, UBA/DBA at 64×6464 \times 64, and FFOB at 64×6464 \times 64):

    Model MNIST CIFAR-10 UBA (DBA) FFOB
    AnoGAN 7120 7120 7110 7223
    EGBAD 8.92 8.71 8.88 8.87
    GANomaly 2.79 2.21 2.66 2.53

    AnoGAN requires over 7,100 ms7{,}100\text{ ms} per sample because it iteratively optimizes the latent vector via backpropagation at test time. EGBAD reduces runtime to approximately 8.7–8.9 ms8.7\text{--}8.9\text{ ms} by jointly mapping image and latent space. GANomaly achieves the fastest execution (2.21–2.79 ms2.21\text{--}2.79\text{ ms} per image), running more than 3×3\times faster than EGBAD and over 2,500×2{,}500\times faster than AnoGAN, with minimal runtime scaling even when image dimension and network capacity double from 32×3232 \times 32 to 64×6464 \times 64.

  6. Knowl 6 — Anomaly Detection Benchmark Results on MNIST and CIFAR-10

    empirical result

    GANomaly was benchmarked on MNIST (32×3232 \times 32) and CIFAR-10 (32×3232 \times 32) under a one-vs-all experimental protocol where one class at a time was designated as the anomalous class and the remaining nine classes served as normal training data (split 80% train / 20% test for normal samples):

    • MNIST: Across all ten digit classes (0–9), GANomaly consistently achieves higher AUC than AnoGAN, EGBAD, and standard Variational Autoencoders (VAE), reaching AUCs between ∼0.60\sim 0.60 and >0.90> 0.90 depending on the digit. Variational Autoencoders perform noticeably worse across all digit splits.
    • CIFAR-10: GANomaly outperforms both AnoGAN and EGBAD across all ten object classes (plane, car, bird, cat, deer, dog, frog, horse, ship, truck), achieving peak class AUCs around 0.65–0.680.65\text{--}0.68. The overall lower quantitative AUCs on CIFAR-10 relative to MNIST are attributed to high visual and semantic similarity between specific normal and abnormal pairs (e.g., plane vs. bird, cat vs. dog, horse vs. deer, and car vs. truck).

  7. Knowl 7 — Sensitivity to Latent Dimension and Loss Weighting Hyperparameters

    empirical result

    Empirical sensitivity analysis on GANomaly hyperparameters reveals distinct optima for latent space capacity and loss term weightings:

    • Latent Bottleneck Vector Size (dd): Evaluated across dimensions z∈{32,100,256,512,1024}z \in \{32, 100, 256, 512, 1024\} on MNIST with digit '2' as the anomaly class, model AUC peaks at d=100d = 100. Lower capacity (d=32d = 32) under-represents normal variation, while excessively high dimensionality (d≥512d \ge 512) allows anomalous samples to be partially reconstructed, degrading detection AUC.
    • Loss Term Weighting: Scanning individual weight ranges from 1 to 90 shows that setting wadv=1w_{adv} = 1, wcon=50w_{con} = 50, and wenc=1w_{enc} = 1 yields the maximum anomaly detection AUC. Heavily weighting the contextual reconstruction loss (wcon=50w_{con} = 50) ensures sharp image reconstruction fidelity while equal lower weighting on feature matching (wadv=1w_{adv} = 1) and latent representation alignment (wenc=1w_{enc} = 1) maintains training stability.

Coverage note — None was omitted; all contributed models, loss formulations, scoring mechanisms, experimental benchmarks, runtime figures, and ablation findings are covered.

References

  1. 1.OSCT Borders X-ray Image Library, UK Home Office Centre for Applied Science and Technology (CAST). Publication Number: 146/16 (2016)
  2. 2.Abdallah, A., Maarof, M.A., Zainal, A.: Fraud detection system: A survey. Journal of Network and Computer Applications 68, 90–113 (jun 2016). https://doi.org/10.1016/J.JNCA.2016.04.007, https://www.sciencedirect.com/science/article/pii/S1084804516300571
  3. 3.Ahmed, M., Mahmood, A.N., Islam, M.R.: A survey of anomaly detection techniques in financial domain. Future Generation Computer Systems 55, 278–288 (feb 2016). https://doi.org/10.1016/J.FUTURE.2015.01.001, https://www.sciencedirect.com/science/article/pii/S0167739X15000023
  4. 4.Ahmed, M., Naser Mahmood, A., Hu, J.: A survey of network anomaly detection techniques. Journal of Network and Computer Applications 60, 19–31 (jan 2016). https://doi.org/10.1016/J.JNCA.2015.11.016, https://www.sciencedirect.com/science/article/pii/S1084804515002891
  5. 5.Akcay, S., Kundegorski, M.E., Willcocks, C.G., Breckon, T.P.: Using deep convolutional neural network architectures for object classification and detection within x-ray baggage security imagery. IEEE Transactions on Information Forensics and Security 13(9), 2203–2215 (Sept 2018). https://doi.org/10.1109/TIFS.2018.2812196
  6. 6.An, J., Cho, S.: Variational autoencoder based anomaly detection using reconstruction probability. Special Lecture on IE 2, 1–18 (2015)
  7. 7.Arjovsky, M., Bottou, L.: Towards Principled Methods for Training Generative Adversarial Networks. In: 2017 ICLR (April 2017), http://arxiv.org/abs/1701.04862
  8. 8.Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: Proceedings of the 34th International Conference on Machine Learning. pp. 214–223. Sydney, Australia (06–11 Aug 2017), http://proceedings.mlr.press/v70/arjovsky17a.html
  9. 9.Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection. ACM Computing Surveys 41(3), 1–58 (jul 2009). https://doi.org/10.1145/1541880.1541882
  10. 10.Chen, X., Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., Abbeel, P.: InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets. In: Advances in Neural Information Processing Systems. pp. 2172–2180 (2016)
  11. 11.Creswell, A., Bharath, A.A.: Inverting the generator of a generative adversarial network (ii). arXiv preprint arXiv:1802.05701 (2018)
  12. 12.Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., Bharath, A.A.: Generative adversarial networks: An overview. IEEE Signal Processing Magazine 35(1), 53–65 (2018)
  13. 13.Dimokranitou, A.: Adversarial Autoencoders for Anomalous Event Detection in Images. Ph.D. thesis, Purdue University (2017)
  14. 14.Donahue, J., Kr¨ahenb¨uhl, P., Darrell, T.: Adversarial Feature Learning. In: International Conference on Learning Representations (ICLR). Toulon, France (apr 2017), http://arxiv.org/abs/1605.09782
  15. 15.Dumoulin, V., Belghazi, I., Poole, B., Mastropietro, O., Lamb, A., Arjovsky, M., Courville, A.: Adversarially learned inference. In: ICLR (2017)
  16. 16.Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672–2680 (2014)
  17. 17.Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. In: Advances in Neural Information Processing Systems. pp. 5767–5777 (2017)
  18. 18.Hasan, M., Choi, J., Neumann, J., Roy-Chowdhury, A.K., Davis, L.S.: Learning temporal regularity in video sequences. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 733–742 (2016)
  19. 19.Hodge, V., Austin, J.: A Survey of Outlier Detection Methodologies. Artificial Intelligence Review 22(2), 85–126 (oct 2004). https://doi.org/10.1023/B:AIRE.0000045502.10941.a9, http://link.springer.com/10.1023/B:AIRE.0000045502.10941.a9
  20. 20.Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Proceedings of the 32nd International Conference on Machine Learning. pp. 448–456. Lille, France (07–09 Jul 2015), http://proceedings.mlr.press/v37/ioffe15.html
  21. 21.Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5967–5976 (July 2017). https://doi.org/10.1109/CVPR.2017.632
  22. 22.Kinga, D., Adam, J.B.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (ICLR). vol. 5 (2015)
  23. 23.Kiran, B.R., Thomas, D.M., Parakkal, R.: An overview of deep learning based methods for unsupervised and semi-supervised anomaly detection in videos. Journal of Imaging 4(2), 36 (2018)
  24. 24.Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., Citeseer (2009)
  25. 25.LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist/
  26. 26.Lipton, Z.C., Tripathi, S.: Precise recovery of latent vectors from generative adversarial networks. In: ICLR Workshop (2017)
  27. 27.Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., Frey, B.: Adversarial autoencoders. In: ICLR (2016)
  28. 28.Markou, M., Singh, S.: Novelty detection: a reviewpart 1: statistical approaches. Signal Processing 83(12), 2481–2497 (dec 2003). https://doi.org/10.1016/J.SIGPRO.2003.07.018, https://www.sciencedirect.com/science/article/pii/S0165168403002020
  29. 29.Markou, M., Singh, S.: Novelty detection: a reviewpart 2:: neural network based approaches. Signal Processing 83(12), 2499–2521 (dec 2003). https://doi.org/10.1016/J.SIGPRO.2003.07.019, https://www.sciencedirect.com/science/article/pii/S0165168403002032
  30. 30.Medel, J.R., Savakis, A.: Anomaly Detection in Video Using Predictive Convolutional Long Short-Term Memory Networks. CoRR abs/1612.0 (dec 2016)
  31. 31.Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014)
  32. 32.Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch (2017)
  33. 33.Pimentel, M.A., Clifton, D.A., Clifton, L., Tarassenko, L.: A review of novelty detection. Signal Processing 99, 215–249 (2014)
  34. 34.Radford, A., Metz, L., Chintala, S.: Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. In: ICLR (2016)
  35. 35.Ravanbakhsh, M., Sangineto, E., Nabi, M., Sebe, N.: Training Adversarial Discriminators for Cross-channel Abnormal Event Detection in Crowds. CoRR abs/1706.0 (jun 2017), http://arxiv.org/abs/1706.07680
  36. 36.Rogers, T.W., Jaccard, N., Morton, E.J., Griffin, L.D.: Automated x-ray image analysis for cargo security: critical review and future promise. Journal of X-ray science and technology (Preprint), 1–24 (2016)
  37. 37.Sabokrou, M., Fathy, M., Hoseini, M., Klette, R.: Real-time anomaly detection and localization in crowded scenes. 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 56–62 (2015). https://doi.org/10.1109/CVPRW.2015.7301284, http://ieeexplore.ieee.org/document/7301284/
  38. 38.Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Advances in Neural Information Processing Systems. pp. 2234–2242 (2016)
  39. 39.Schlegl, T., Seeb¨ock, P., Waldstein, S.M., Schmidt-Erfurth, U., Langs, G.: Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 10265 LNCS, 146–147 (2017). https://doi.org/10.1007/978-3-319-59050-9 12
  40. 40.Zenati, H., Foo, C.S., Lecouat, B., Manek, G., Chandrasekhar, V.R.: Efficient gan-based anomaly detection. arXiv preprint arXiv:1802.06222 (2018)

Citation

MLA
Akcay, S., et al. “GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training”. arXiv, 2018, http://arxiv.org/abs/1805.06725v3.
APA
Akcay, S., Atapour-Abarghouei, A., & Breckon, T. P. (2018). GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training. arXiv. http://arxiv.org/abs/1805.06725v3
Chicago
Akcay, S., A. Atapour-Abarghouei, and T. P. Breckon. 2018. “GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training”. arXiv. http://arxiv.org/abs/1805.06725v3.
Harvard
Akcay, S., Atapour-Abarghouei, A. and Breckon, T.P. (2018) “GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1805.06725v3.
Vancouver
1. Akcay S, Atapour-Abarghouei A, Breckon TP (2018) GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training. arXiv

BibTeX

@article{akcay2018ganomaly,
  title = {GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training},
  author = {Akcay, Samet and Atapour-Abarghouei, Amir and Breckon, Toby P.},
  year = {2018},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1805.06725v3},
  eprint = {1805.06725}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF