Back

Unsupervised Out-of-Distribution Detection with Diffusion Inpainting

Zhenzhen LiuJin Peng ZhouYufan WangKilian Q. Weinberger

article2023ICML79 citations

Introduces Lift, Map, Detect (LMD), an unsupervised out-of-distribution detection framework that identifies anomalous images by masking them and measuring reconstruction error after inpainting with a diffusion model trained only on in-domain data.

Listen

Machine learning systems deployed in high-stakes environments, such as medical diagnostics and criminal justice, rely on the assumption that incoming test data match the distribution of their training data. When exposed to unfamiliar or out-of-distribution inputs, these models can produce silent, misleading, or hazardous failures. Many existing safety techniques require labeled data or prior knowledge of potential anomalies, which is often impractical because anomalies in the real world are unpredictable and human labeling is expensive. Developing effective out-of-distribution detection that relies strictly on unlabeled in-domain data is therefore essential for reliable artificial intelligence deployment.

The article introduces and evaluates Lift, Map, Detect (LMD), a new unsupervised framework that identifies out-of-distribution images using generative diffusion models without requiring model retraining or labeled data.

The approach builds on the principle that diffusion models learn to map corrupted images back toward the underlying data manifold on which they were trained. Under LMD, an image is first lifted off its original distribution by applying a corrupting mask. The model then fills in the missing regions through an inpainting process. If the input belongs to the target domain, the model reconstructs it accurately; if it comes from an unfamiliar domain, the model attempts to force it into the target domain, creating significant visual errors. The distance between the original and reconstructed image is measured using a standard perceptual similarity metric. To reduce random variation, the framework performs multiple reconstruction attempts using an alternating checkerboard mask pattern and aggregates the results via a median score.

The evaluation demonstrates that LMD achieves top-tier performance across diverse benchmark datasets, attaining the highest overall average area under the ROC curve (0.907) compared to seven competitive baseline methods. LMD achieved standout results on difficult image pairs, improving detection accuracy by up to 10% over the best baseline on CIFAR-100 versus SVHN and reaching a 0.991 detection score on high-resolution image benchmarks. Ablation studies confirm that using an alternating checkerboard pattern and aggregating roughly ten reconstruction attempts per image consistently maximizes detection accuracy, whereas traditional single-region masks degrade performance.

These findings indicate that generative diffusion models can serve as robust safety guards for vision-based artificial intelligence systems without requiring expensive retraining or domain labels. By leveraging perceptual reconstruction errors rather than raw statistical likelihoods, LMD avoids common failure modes where generative models erroneously assign high confidence to abnormal inputs. This reduces operational and safety risks in production environments where abnormal data could otherwise bypass standard error checks.

Organizations seeking to implement LMD should adopt the alternating checkerboard masking strategy combined with perceptual similarity scoring. However, because diffusion-based inpainting requires multiple iterative sampling steps, LMD is computationally intensive and currently too slow for real-time applications requiring millisecond responses. Future engineering efforts should integrate emerging fast-sampling diffusion algorithms to reduce inference latency before deploying the system in time-critical operational workflows.

Cornell Universityout-of-distribution detectionout-of-distribution detectiondiffusion inpaintingunsupervised anomaly detection
arXiv: 2302.10326
Cover for Unsupervised Out-of-Distribution Detection with Diffusion Inpainting

Abstract

Unsupervised out-of-distribution detection (OOD) seeks to identify out-of-domain data by learning only from unlabeled in-domain data. We present a novel approach for this task – Lift, Map, Detect (LMD) – that leverages recent advancement in diffusion models. Diffusion models are one type of generative models. At their core, they learn an iterative denoising process that gradually maps a noisy image closer to their training manifolds. LMD leverages this intuition for OOD detection. Specifically, LMD lifts an image off its original manifold by corrupting it, and maps it towards the in-domain manifold with a diffusion model. For an out-of-domain image, the mapped image would have a large distance away from its original manifold, and LMD would identify it as OOD accordingly. We show through extensive experiments that LMD achieves competitive performance across a broad variety of datasets. Code can be found at https://github.com/zhenzhel/lift_map_detect.

Topics

out-of-distribution detection

Keywords

out-of-distribution detectiondiffusion inpaintingunsupervised anomaly detectiondiffusion modelsnovelty detectionreconstruction distance metricsrandom patch mask

Table of Contents

  • 1. Introduction
  • 2. Background
  • 3. Lift, Map, Detect
  • 4. Experiments
  • 4.1. Experiment Settings
  • 4.2. Implementation Details of LMD
  • 4.3. Experimental Results
  • 4.4. Ablation
  • 5. Discussion and Conclusion
  • 6. Acknowledgement
  • References

Knowls

  1. Knowl 1 — Lift, Map, Detect for unsupervised OOD detection

    model/method

    Lift, Map, Detect (LMD) detects out-of-distribution images using only unlabeled in-domain training images. A diffusion model is trained on the in-domain distribution and used as a mapping toward the in-domain image manifold. For a test image, LMD first lifts the image away from its original manifold by masking part of it, then maps the masked image back toward the learned in-domain manifold through diffusion inpainting. The perceptual distance between the original and inpainted images is used as the OOD score: a large distance indicates that the diffusion model could not reconstruct the image in its original vicinity and therefore suggests that the image is out-of-domain.

    The method relies on the hypothesis that an in-domain image can be plausibly restored after masking, whereas an out-of-domain image is mapped toward the in-domain manifold and receives a substantially different reconstruction. The diffusion model is used without retraining during detection.

  2. Knowl 2 — End-to-end LMD detection procedure

    algorithm

    For a test image xx, an in-domain diffusion model θin\theta_{\mathrm{in}}, a sequence of binary masks M1,…,MrM_1,\ldots,M_r, and a perceptual distance function, LMD returns one OOD score. In every mask, Mi=1M_i=1 denotes pixels retained from the original image and Mi=0M_i=0 denotes pixels to be inpainted. The default implementation uses LPIPS as the distance function and aggregates the distances with their median.

    Input: test image x, in-domain diffusion model θin, masks M1, ..., Mr, diffusion steps T
    Output: OOD score s(x)
    for i = 1 to r do
        xmasked ← x with pixels where Mi = 0 removed
        for t = T down to 1 do
            if t = T then
                xinpainted ← sample from the diffusion noise distribution at step T
            end if
            xorig_prime ← diffuse(x; θin) to step t - 1
            xinpainted ← denoise(xinpainted; θin) to step t - 1
            xinpainted ← xorig_prime · Mi + xinpainted · (1 - Mi)
        end for
        di ← LPIPS(x, xinpainted)
    end for
    return s(x) = median(d1, ..., dr)

    The inpainting loop preserves the appropriately diffused original pixels in unmasked regions while replacing masked regions with the diffusion model’s denoised samples. The resulting score is larger when the reconstructed image is perceptually more dissimilar to xx.

  3. Knowl 3 — Alternating checkerboard masking

    model/method

    LMD uses an alternating checkerboard N×NN\times N masking strategy to lift images while preserving enough spatial context for plausible in-domain inpainting. Each image is divided into an N×NN\times N grid of patches, and approximately half of the patches are masked in a checkerboard pattern. On the next reconstruction attempt, the masked and unmasked patches are exchanged. The default is an alternating checkerboard with N=8N=8.

    This design masks every image region across two attempts, so a distinguishing out-of-domain feature is less likely to remain visible in all reconstructions. It also avoids the two extremes in which a very small mask gives little separation between domains or a very large mask removes so much context that even in-domain reconstructions become unrelated to their originals.

  4. Knowl 4 — Median aggregation over multiple reconstructions

    model/method

    Because diffusion inpainting is stochastic, a single reconstruction can be atypically dissimilar for an in-domain image or atypically similar for an out-of-domain image. LMD therefore performs multiple independent lift-and-inpaint attempts for each test image, computes one reconstruction distance per attempt, and uses the median distance as the final OOD score. The paper’s default is r=10r=10 reconstructions per image with alternating checkerboard 8×88\times8 masks.

    In the experiments, increasing the number of attempts almost always improves ROC-AUC, with the largest gains occurring at small numbers of attempts and performance generally saturating around ten attempts. The authors also report in preliminary experiments that the median performs better than simple mean aggregation.

  5. Knowl 5 — Benchmark comparison across twelve dataset pairs

    data/table

    LMD was evaluated pairwise on CIFAR-10, CIFAR-100, and SVHN, and separately on MNIST, KMNIST, and FashionMNIST. Each detector was trained, when necessary, using the in-domain training set and evaluated on the full in-domain and out-of-domain test sets. The metric is ROC-AUC, where higher values are better. LMD used the same configuration for every pair: alternating checkerboard 8×88\times8 masks, LPIPS distance, and ten reconstructions per image. The comparison includes likelihood, input complexity, likelihood regret, a pretrained feature extractor with Mahalanobis distance, autoencoder reconstruction error with MSE, autoencoder Mahalanobis distance, and AnoGAN.

    Could not parse LaTeX table

    LMD obtains the best ROC-AUC on five of the twelve dataset pairs and has the highest average ROC-AUC, 0.907. Its strongest gains occur for CIFAR-10 versus SVHN, with ROC-AUC 0.992, and CIFAR-100 versus SVHN, with ROC-AUC 0.985.

  6. Knowl 6 — Mask-pattern ablation

    data/table

    The mask choice materially affects LMD. The ablation compares alternating checkerboards at three grid sizes, a fixed non-alternating checkerboard, a centered square mask covering one-fourth of the image, and a random patch mask covering half of an 8×88\times8 patch grid. Each image is reconstructed once, and ROC-AUC is reported for three in-domain/out-of-domain pairs.

    Could not parse LaTeX table

    The alternating 8×88\times8 checkerboard is the most consistently strong choice across the three pairs. The centered mask performs particularly poorly for MNIST versus KMNIST because it removes too much information. The alternating 4×44\times4 and 16×1616\times16 patterns can also lose performance when the masked regions are poorly matched to the image structure. The fixed 8×88\times8 checkerboard is close to the alternating version on these datasets, but alternating masks have the additional property that they cover the entire image over multiple attempts.

  7. Knowl 7 — Reconstruction-distance ablation

    data/table

    LMD was evaluated with four reconstruction-distance choices: pixelwise mean squared error (MSE), structural similarity index measure (SSIM), learned perceptual image patch similarity (LPIPS), and cosine distance between SimCLRv2 representations. The metrics were compared on three dataset pairs using ROC-AUC.

    Could not parse LaTeX table

    LPIPS is competitive on all three pairs and is the paper’s default distance metric. SimCLRv2 performs best on CIFAR-10 versus CIFAR-100, but its performance is lower than LPIPS on CIFAR-10 versus SVHN and KMNIST versus MNIST, showing that the most effective distance metric can depend on the domain.

  8. Knowl 8 — High-resolution qualitative and quantitative behavior

    empirical result

    For a higher-resolution test, the authors used CelebA-HQ as the in-domain dataset and ImageNet as the out-of-domain dataset. Because CelebA-HQ had no train/test split and its available checkpoint was trained on the full dataset, they used a pretrained FFHQ checkpoint to reduce potential memorization concerns. They randomly sampled 100 images from each dataset, standardized them to 256×256256\times256, performed one reconstruction per image, and used LPIPS as the distance metric.

    The resulting ROC-AUC was 0.991 with an alternating checkerboard 8×88\times8 mask, 0.994 with an alternating checkerboard 4×44\times4 mask, and 1.000 with a centered mask. In-domain face reconstructions were visually almost identical to their originals, whereas ImageNet reconstructions showed locally incoherent or face-like artifacts. The result supports the applicability of LMD beyond the small low-resolution datasets used in the main benchmark, although it is presented as a qualitative high-resolution demonstration.

  9. Knowl 9 — Alternative diffusion-and-denoising instantiation

    empirical result

    LMD’s lift-and-map framework can also be instantiated without masking and inpainting. In the alternative version, an image is lifted by applying the diffusion process to step t=500t=500 of a T=1000T=1000-step schedule, then mapped by denoising from that noisy state. The experiments used ten attempts per image and the median LPIPS reconstruction error as the OOD score.

    Could not parse LaTeX table

    Diffusion and denoising remains competitive, but masking and inpainting performs better on all three reported pairs. This experiment indicates that the broader lift-then-map principle is not restricted to one particular corruption and reconstruction mechanism.

  10. Knowl 10 — Computational limitation of vanilla LMD

    limitation

    Vanilla LMD inherits the high computational cost of diffusion sampling because each inpainting reconstruction requires many iterative denoising steps, and multiple reconstructions are used for every test image. The authors therefore judge LMD difficult to deploy for real-time OOD detection in its current form. They identify fast diffusion samplers and methods that skip or reduce sampling steps as a possible way to address this limitation, including samplers operating in roughly 10–20 steps.

Coverage note — No substantial contributed material was omitted; background diffusion equations and related-work discussion were excluded because they are not contributions of this paper.

References

  1. 1.Ahmadian, A. and Lindsten, F. Likelihood-free out-of-distribution detection with invertible generative models. In IJCAI, pp. 2119–2125, 2021.
  2. 2.Bergamin, F., Mattei, P.-A., Havtorn, J. D., Senetaire, H., Schmutz, H., Maaløe, L., Hauberg, S., and Frellsen, J. Model-agnostic out-of-distribution detection using combined statistical tests. In International Conference on Artificial Intelligence and Statistics, pp. 10753–10776. PMLR, 2022.
  3. 3.Bergman, L. and Hoshen, Y. Classification-based anomaly detection for general data. arXiv preprint arXiv:2005.02359, 2020.
  4. 4.Bishop, C. M. Novelty detection and neural network validation. IEE Proceedings-Vision, Image and Signal processing, 141(4):217–222, 1994.
  5. 5.Chen, T., Kornblith, S., Swersky, K., Norouzi, M., and Hinton, G. E. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243–22255, 2020.
  6. 6.Choi, H., Jang, E., and Alemi, A. A. Waic, but why? generative ensembles for robust anomaly detection. arXiv preprint arXiv:1810.01392, 2018.
  7. 7.Clanuwat, T., Bober-Irizar, M., Kitamoto, A., Lamb, A., Yamamoto, K., and Ha, D. Deep learning for classical japanese literature. 2018.
  8. 8.Denouden, T., Salay, R., Czarnecki, K., Abdelzad, V., Phan, B., and Vernekar, S. Improving reconstruction autoencoder out-of-distribution detection with mahalanobis distance. arXiv preprint arXiv:1812.02765, 2018.
  9. 9.Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks, 2014. URL https://arxiv.org/abs/1406.2661.
  10. 10.Graham, M. S., Pinaya, W. H., Tudosiu, P.-D., Nachev, P., Ourselin, S., and Cardoso, M. J. Denoising diffusion models for out-of-distribution detection. arXiv preprint arXiv:2211.07740, 2022.
  11. 11.Hamet, P. and Tremblay, J. Artificial intelligence in medicine. Metabolism, 69:S36–S40, 2017.
  12. 12.Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016.
  13. 13.Hendrycks, D., Mazeika, M., and Dietterich, T. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018.
  14. 14.Hendrycks, D., Mazeika, M., Kadavath, S., and Song, D. Using self-supervised learning can improve model robustness and uncertainty. Advances in neural information processing systems, 32, 2019.
  15. 15.Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020.
  16. 16.Huang, R., Geng, A., and Li, Y. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 34:677–689, 2021.
  17. 17.Karras, T., Aila, T., Laine, S., and Lehtinen, J. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. URL http://arxiv.org/abs/1710.10196.
  18. 18.Karras, T., Laine, S., and Aila, T. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4401–4410, 2019.
  19. 19.Kingma, D. P. and Dhariwal, P. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018.
  20. 20.Kirichenko, P., Izmailov, P., and Wilson, A. G. Why normalizing flows fail to detect out-of-distribution data. Advances in neural information processing systems, 33:20578–20589, 2020.
  21. 21.Krizhevsky, A. Learning multiple layers of features from tiny images. Technical report, 2009.
  22. 22.LeCun, Y., Cortes, C., and Burges, C. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010.
  23. 23.Lee, K., Lee, K., Lee, H., and Shin, J. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems, 31, 2018.
  24. 24.Li, D., Chen, D., Goh, J., and Ng, S.-k. Anomaly detection with generative adversarial networks for multivariate time series. arXiv preprint arXiv:1809.04758, 2018.
  25. 25.Liang, S., Li, Y., and Srikant, R. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017.
  26. 26.Liu, L., Ren, Y., Lin, Z., and Zhao, Z. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022.
  27. 27.Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., and Van Gool, L. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11461–11471, 2022.
  28. 28.Maaløe, L., Fraccaro, M., Lievin, V., and Winther, O. Biva: A very deep hierarchy of latent variables for generative modeling. Advances in neural information processing systems, 32, 2019.
  29. 29.Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.-Y., and Ermon, S. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021.
  30. 30.Morningstar, W., Ham, C., Gallagher, A., Lakshminarayanan, B., Alemi, A., and Dillon, J. Density of states estimation for out of distribution detection. In International Conference on Artificial Intelligence and Statistics, pp. 3232–3240. PMLR, 2021.
  31. 31.Nalisnick, E., Matsukawa, A., Teh, Y. W., Gorur, D., and Lakshminarayanan, B. Do deep generative models know what they don’t know? arXiv preprint arXiv:1810.09136, 2018.
  32. 32.Nalisnick, E. T., Matsukawa, A., Teh, Y. W., and Lakshminarayanan, B. Detecting out-of-distribution inputs to deep generative models using a test for typicality. 2019.
  33. 33.Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y. Reading digits in natural images with unsupervised feature learning. 2011.
  34. 34.Nichol, A. Q. and Dhariwal, P. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pp. 8162–8171. PMLR, 2021.
  35. 35.Ren, J., Liu, P. J., Fertig, E., Snoek, J., Poplin, R., Depristo, M., Dillon, J., and Lakshminarayanan, B. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32, 2019.
  36. 36.Rigano, C. Using artificial intelligence to address criminal justice needs. National Institute of Justice Journal, 280:1–10, 2019.
  37. 37.Ruff, L., Vandermeulen, R. A., Gornitz, N., Binder, A., Muller, E., Müller, K.-R., and Kloft, M. Deep semi-supervised anomaly detection. arXiv preprint arXiv:1906.02694, 2019.
  38. 38.Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. doi: 10.1007/s11263-015-0816-y.
  39. 39.Sakurada, M. and Yairi, T. Anomaly detection using autoencoders with nonlinear dimensionality reduction. In Proceedings of the MLSDA 2014 2nd workshop on machine learning for sensory data analysis, pp. 4–11, 2014.
  40. 40.Salimans, T., Karpathy, A., Chen, X., and Kingma, D. P. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. arXiv preprint arXiv:1701.05517, 2017.
  41. 41.Schlegl, T., Seebock, P., Waldstein, S. M., Schmidt-Erfurth, U., and Langs, G. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pp. 146–157. Springer, 2017.
  42. 42.Sehwag, V., Chiang, M., and Mittal, P. Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051, 2021.
  43. 43.Serra, J., Álvarez, D., Gómez, V., Slizovskaia, O., Núñez, J. F., and Luque, J. Input complexity and out-of-distribution detection with likelihood-based generative models. arXiv preprint arXiv:1909.11480, 2019.
  44. 44.Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp. 2256–2265. PMLR, 2015.
  45. 45.Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019.
  46. 46.Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.
  47. 47.Tack, J., Mo, S., Jeong, J., and Shin, J. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33:11839–11852, 2020.
  48. 48.Wang, H., Li, Z., Feng, L., and Zhang, W. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4921–4930, 2022.
  49. 49.Wang, Z., Simoncelli, E. P., and Bovik, A. C. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pp. 1398–1402. Ieee, 2003.
  50. 50.Watson, D., Chan, W., Ho, J., and Norouzi, M. Learning fast samplers for diffusion models by differentiating through sample quality. In International Conference on Learning Representations, 2022.
  51. 51.Xia, Y., Cao, X., Wen, F., Hua, G., and Sun, J. Learning discriminative reconstructions for unsupervised outlier removal. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1511–1519, 2015.
  52. 52.Xiao, H., Rasul, K., and Vollgraf, R. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. CoRR, abs/1708.07747, 2017. URL http://arxiv.org/abs/1708.07747.
  53. 53.Xiao, Z., Yan, Q., and Amit, Y. Likelihood regret: An out-of-distribution detection score for variational auto-encoder. Advances in neural information processing systems, 33:20685–20696, 2020.
  54. 54.Xiao, Z., Yan, Q., and Amit, Y. Do we really need to learn representations from in-domain data for outlier detection? arXiv preprint arXiv:2105.09270, 2021.
  55. 55.Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., and Hu, H. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9653–9663, 2022.
  56. 56.Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, Y., Zhang, W., Cui, B., and Yang, M.-H. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022.
  57. 57.Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 586–595, 2018.
  58. 58.Zhou, C. and Paffenroth, R. C. Anomaly detection with robust deep autoencoders. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 665–674, 2017.
  59. 59.Zong, B., Song, Q., Min, M. R., Cheng, W., Lumezanu, C., Cho, D., and Chen, H. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International conference on learning representations, 2018.

Citation

MLA
Liu, Z., et al. “Unsupervised Out-of-Distribution Detection with Diffusion Inpainting”. International Conference on Machine Learning, vol. 202, 2023, pp. 22528–38, https://proceedings.mlr.press/v202/liu23bd.html.
APA
Liu, Z., Zhou, J. P., Wang, Y., & Weinberger, K. Q. (2023). Unsupervised Out-of-Distribution Detection with Diffusion Inpainting. International Conference on Machine Learning, 202, 22528–22538. https://proceedings.mlr.press/v202/liu23bd.html
Chicago
Liu, Z., J. P. Zhou, Y. Wang, and K. Q. Weinberger. 2023. “Unsupervised Out-of-Distribution Detection with Diffusion Inpainting”. International Conference on Machine Learning 202: 22528–38. https://proceedings.mlr.press/v202/liu23bd.html.
Harvard
Liu, Z. et al. (2023) “Unsupervised Out-of-Distribution Detection with Diffusion Inpainting”, International Conference on Machine Learning. PMLR, pp. 22528–22538. Available at: https://proceedings.mlr.press/v202/liu23bd.html.
Vancouver
1. Liu Z, Zhou JP, Wang Y, Weinberger KQ (2023) Unsupervised Out-of-Distribution Detection with Diffusion Inpainting. In: International Conference on Machine Learning. PMLR, pp 22528–22538

BibTeX

@InProceedings{pmlr-v202-liu23bd,
  title = 	 {Unsupervised Out-of-Distribution Detection with Diffusion Inpainting},
  author =       {Liu, Zhenzhen and Zhou, Jin Peng and Wang, Yufan and Weinberger, Kilian Q},
  booktitle = 	 {Proceedings of the 40th International Conference on Machine Learning},
  pages = 	 {22528--22538},
  year = 	 {2023},
  editor = 	 {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan},
  volume = 	 {202},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {23--29 Jul},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v202/liu23bd/liu23bd.pdf},
  url = 	 {https://proceedings.mlr.press/v202/liu23bd.html},
  abstract = 	 {Unsupervised out-of-distribution detection (OOD) seeks to identify out-of-domain data by learning only from unlabeled in-domain data. We present a novel approach for this task – Lift, Map, Detect (LMD) – that leverages recent advancement in diffusion models. Diffusion models are one type of generative models. At their core, they learn an iterative denoising process that gradually maps a noisy image closer to their training manifolds. LMD leverages this intuition for OOD detection. Specifically, LMD lifts an image off its original manifold by corrupting it, and maps it towards the in-domain manifold with a diffusion model. For an OOD image, the mapped image would have a large distance away from its original manifold, and LMD would identify it as OOD accordingly. We show through extensive experiments that LMD achieves competitive performance across a broad variety of datasets. Code can be found at https://github.com/zhenzhel/lift_map_detect.}
}
Metadata:DOI registry

Access the Paper

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

Open PDF
License: https://creativecommons.org/licenses/by/4.0/