Full search page →

Randomized smoothing in practice

Optional · supplementary note 5 of 6 for DLCV 14 · Robustness: Distribution Shift, Adversarial Examples and Anomalies. Nothing in the main text depends on it; citation numbers refer to the reference list at the bottom of this page.

Navigate: this page and this track: Navigate: this page and this track

On this page

Supplementary notes · 6

  1. Effective robustness and accuracy on the line
  2. Test-time entropy minimization
  3. Domain-adversarial training and the gradient-reversal layer
  4. Attack settings that decide an evaluation
  5. Randomized smoothing in practice
  6. PatchCore step by step

Paper guides · 3

  1. Madry et al., “Towards Deep Learning Models Resistant to Adversarial Attacks” (ICLR 2018)
  2. Athalye et al., “Obfuscated Gradients Give a False Sense of Security” (ICML 2018)
  3. Gulrajani and Lopez-Paz, “In Search of Lost Domain Generalization” (ICLR 2021)

Deep Learning for Computer Vision

  1. 02 · Deep Image Classification
View the whole course tree →

This note goes through the certification procedure of Cohen et al., explains why the certified radius is capped, and gives a tested snippet on digits.

The procedure. Given a base classifier ff, noise level σ\sigma, an input x\mathbf{x}, sample sizes n0n_0 and nn, and a failure probability α\alpha, the CERTIFY procedure is [31]:

  1. Draw n0n_0 noisy copies x+η\mathbf{x} + \boldsymbol{\eta}, classify them with ff, and take the most frequent class c^A\hat{c}_A (a guess at the top class).
  2. Draw nn fresh noisy copies and count how many are classified c^A\hat{c}_A; call the count kk.
  3. Compute a one-sided (1−α)(1 - \alpha) lower confidence bound pA‾\underline{p_A} on the binomial probability from kk successes in nn trials; Cohen et al. use the Clopper–Pearson bound.
  4. If pA‾>1/2\underline{p_A} > 1/2, return c^A\hat{c}_A and the radius R=σ Φ−1(pA‾)R = \sigma\, \Phi^{-1}(\underline{p_A}); otherwise abstain.

Step 4 uses the bound pB‾=1−pA‾\overline{p_B} = 1 - \underline{p_A} on all other classes, which turns the radius of the main text into σ2(Φ−1(pA‾)−Φ−1(1−pA‾))=σ Φ−1(pA‾)\frac{\sigma}{2}\big(\Phi^{-1}(\underline{p_A}) - \Phi^{-1}(1 - \underline{p_A})\big) = \sigma\, \Phi^{-1}(\underline{p_A}), because Φ−1(1−p)=−Φ−1(p)\Phi^{-1}(1 - p) = -\Phi^{-1}(p). With probability at least 1−α1 - \alpha, a returned class is what the smoothed classifier gg predicts everywhere within radius RR [31]. Using separate samples for steps 1 and 2 keeps the bound valid: selecting the class and estimating its probability from the same samples would bias the estimate.

Why the radius is capped. Even if all nn noisy copies agree (k=nk = n), the Clopper–Pearson lower bound is pA‾=α1/n\underline{p_A} = \alpha^{1/n}, so R≤σ Φ−1(α1/n)R \le \sigma\, \Phi^{-1}(\alpha^{1/n}). With n=2000n = 2000, α=0.001\alpha = 0.001 and σ=0.5\sigma = 0.5, this cap is about 1.351.35 (our calculation). Larger radii need many more samples, and the paper notes that RR grows only very slowly with nn [31].

Training the base classifier. The certificate holds for any ff, but it is useful only if ff classifies noisy inputs well, so ff is trained with Gaussian noise of the same σ\sigma added to its inputs.

Code · Python · 26 lines
import numpy as np
from scipy.stats import beta, norm
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.neural_network import MLPClassifier

X, y = load_digits(return_X_y=True)
X = X / 16.0
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, stratify=y, random_state=0)
rng = np.random.default_rng(0)
sigma = 0.5
Xnoisy = np.vstack([Xtr + rng.normal(0, sigma, Xtr.shape) for _ in range(4)])   # train f under noise
f = MLPClassifier(hidden_layer_sizes=(128,), alpha=1e-3, max_iter=200, early_stopping=True, random_state=0).fit(Xnoisy, np.tile(ytr, 4))

def certify(x, n0=100, n=2000, alpha=0.001):
    votes0 = np.bincount(f.predict(x + rng.normal(0, sigma, (n0, 64))), minlength=10)
    c = votes0.argmax()                                   # guess the top class with a small sample
    k = np.sum(f.predict(x + rng.normal(0, sigma, (n, 64))) == c)
    p_low = beta.ppf(alpha, k, n - k + 1)                 # one-sided Clopper-Pearson lower bound
    return (c, sigma * norm.ppf(p_low)) if p_low > 0.5 else (-1, 0.0)   # -1 means abstain

res = [certify(x) for x in Xte[:200]]
for r in [0, 0.25, 0.5, 0.75, 1.0]:
    ok = np.mean([(c == t) and (R >= r) for (c, R), t in zip(res, yte[:200])])
    print(f"certified accuracy at l2 radius {r:.2f}: {ok:.3f}")
print("abstained:", sum(c == -1 for c, _ in res), "of", len(res))

It prints:

Code · 6 lines
certified accuracy at l2 radius 0.00: 0.825
certified accuracy at l2 radius 0.25: 0.590
certified accuracy at l2 radius 0.50: 0.290
certified accuracy at l2 radius 0.75: 0.070
certified accuracy at l2 radius 1.00: 0.000
abstained: 32 of 200

(Our toy run.) Certified accuracy at radius rr is the fraction of test images that are certified with R≥rR \ge r and correctly classified. At r=0r = 0 it is the accuracy of the smoothed classifier counting abstentions as errors. To put the radii in scale: an ℓ2\ell_2 change of 0.5 on these [0,1][0, 1] images equals, for example, a change of 0.0625 in every one of the 64 pixels. Note that the noise is not clipped to [0,1][0, 1]; the theorem does not need it.

Edge cases. The certificate is for ℓ2\ell_2; an ℓ∞\ell_\infty ball of radius ε\varepsilon in dd dimensions contains ℓ2\ell_2 perturbations of length up to εd\varepsilon\sqrt{d}, so ℓ2\ell_2 certificates translate into weak ℓ∞\ell_\infty ones for large images. The guarantee is about gg, which is only available through sampling, so prediction is also randomized and costs nn forward passes per image.

References

  1. D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt and D. Song, “Natural Adversarial Examples,” CVPR, 2021. CVF
  2. P. W. Koh, S. Sagawa, H. Marklund, S. M. Xie, M. Zhang, A. Balsubramani et al., “WILDS: A Benchmark of in-the-Wild Distribution Shifts,” ICML, PMLR 139, 2021. PMLR
  3. J. Yang, K. Zhou, Y. Li and Z. Liu, “Generalized Out-of-Distribution Detection: A Survey,” International Journal of Computer Vision, vol. 132, no. 12, pp. 5635–5662, 2024. doi · arXiv
  4. D. Hendrycks and T. Dietterich, “Benchmarking Neural Network Robustness to Common Corruptions and Perturbations,” ICLR, 2019. arXiv
  5. D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo et al., “The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization,” ICCV, 2021. CVF
  6. A. Barbu, D. Mayo, J. Alverio, W. Luo, C. Wang, D. Gutfreund, J. Tenenbaum and B. Katz, “ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models,” NeurIPS, 2019. proceedings
  7. B. Recht, R. Roelofs, L. Schmidt and V. Shankar, “Do ImageNet Classifiers Generalize to ImageNet?,” ICML, PMLR 97, 2019. PMLR
  8. R. Taori, A. Dave, V. Shankar, N. Carlini, B. Recht and L. Schmidt, “Measuring Robustness to Natural Distribution Shifts in Image Classification,” NeurIPS, 2020. proceedings
  9. J. Miller, R. Taori, A. Raghunathan, S. Sagawa, P. W. Koh, V. Shankar, P. Liang, Y. Carmon and L. Schmidt, “Accuracy on the Line: on the Strong Correlation Between Out-of-Distribution and In-Distribution Generalization,” ICML, PMLR 139, 2021. PMLR
  10. D. Hendrycks, N. Mu, E. D. Cubuk, B. Zoph, J. Gilmer and B. Lakshminarayanan, “AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty,” ICLR, 2020. arXiv
  11. R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann and W. Brendel, “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness,” ICLR, 2019. arXiv
  12. A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal et al., “Learning Transferable Visual Models From Natural Language Supervision,” ICML, PMLR 139, 2021. PMLR
  13. A. Fang, G. Ilharco, M. Wortsman, Y. Wan, V. Shankar, A. Dave and L. Schmidt, “Data Determines Distributional Robustness in Contrastive Language Image Pre-training (CLIP),” ICML, PMLR 162, 2022. PMLR
  14. M. Wortsman, G. Ilharco, J. W. Kim, M. Li, S. Kornblith, R. Roelofs et al., “Robust Fine-Tuning of Zero-Shot Models,” CVPR, 2022. CVF
  15. S. Schneider, E. Rusak, L. Eck, O. Bringmann, W. Brendel and M. Bethge, “Improving robustness against common corruptions by covariate shift adaptation,” NeurIPS, 2020. proceedings
  16. D. Wang, E. Shelhamer, S. Liu, B. Olshausen and T. Darrell, “Tent: Fully Test-Time Adaptation by Entropy Minimization,” ICLR, 2021. arXiv
  17. Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand and V. Lempitsky, “Domain-Adversarial Training of Neural Networks,” Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016. JMLR
  18. K. Zhou, Z. Liu, Y. Qiao, T. Xiang and C. C. Loy, “Domain Generalization: A Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. doi · arXiv
  19. I. Gulrajani and D. Lopez-Paz, “In Search of Lost Domain Generalization,” ICLR, 2021. arXiv
  20. Y. Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. V. Dillon, B. Lakshminarayanan and J. Snoek, “Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift,” NeurIPS, 2019. proceedings
  21. C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow and R. Fergus, “Intriguing properties of neural networks,” ICLR, 2014. arXiv
  22. I. J. Goodfellow, J. Shlens and C. Szegedy, “Explaining and Harnessing Adversarial Examples,” ICLR, 2015. arXiv
  23. A. Madry, A. Makelov, L. Schmidt, D. Tsipras and A. Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks,” ICLR, 2018. arXiv
  24. F. Croce, M. Andriushchenko, V. Sehwag, E. Debenedetti, N. Flammarion, M. Chiang, P. Mittal and M. Hein, “RobustBench: a standardized adversarial robustness benchmark,” NeurIPS Datasets and Benchmarks Track, 2021. proceedings · leaderboard
  25. N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik and A. Swami, “Practical Black-Box Attacks against Machine Learning,” ACM Asia Conference on Computer and Communications Security (AsiaCCS), pp. 506–519, 2017. doi
  26. N. Carlini and D. Wagner, “Towards Evaluating the Robustness of Neural Networks,” IEEE Symposium on Security and Privacy (SP), pp. 39–57, 2017. doi
  27. D. Tsipras, S. Santurkar, L. Engstrom, A. Turner and A. Madry, “Robustness May Be at Odds with Accuracy,” ICLR, 2019. arXiv
  28. H. Zhang, Y. Yu, J. Jiao, E. P. Xing, L. El Ghaoui and M. I. Jordan, “Theoretically Principled Trade-off between Robustness and Accuracy,” ICML, PMLR 97, 2019. PMLR
  29. A. Athalye, N. Carlini and D. Wagner, “Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples,” ICML, PMLR 80, 2018. PMLR
  30. F. Croce and M. Hein, “Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks,” ICML, PMLR 119, 2020. PMLR
  31. J. Cohen, E. Rosenfeld and J. Z. Kolter, “Certified Adversarial Robustness via Randomized Smoothing,” ICML, PMLR 97, 2019. PMLR
  32. K. Eykholt, I. Evtimov, E. Fernandes, B. Li, A. Rahmati, C. Xiao, A. Prakash, T. Kohno and D. Song, “Robust Physical-World Attacks on Deep Learning Visual Classification,” CVPR, 2018. CVF
  33. T. B. Brown, D. Mané, A. Roy, M. Abadi and J. Gilmer, “Adversarial Patch,” arXiv:1712.09665, 2017. arXiv
  34. D. Hendrycks and K. Gimpel, “A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks,” ICLR, 2017. arXiv
  35. W. Liu, X. Wang, J. D. Owens and Y. Li, “Energy-based Out-of-distribution Detection,” NeurIPS, 2020. proceedings
  36. K. Lee, K. Lee, H. Lee and J. Shin, “A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks,” NeurIPS, 2018. proceedings
  37. Y. Sun, Y. Ming, X. Zhu and Y. Li, “Out-of-Distribution Detection with Deep Nearest Neighbors,” ICML, PMLR 162, 2022. PMLR
  38. D. Hendrycks, M. Mazeika and T. Dietterich, “Deep Anomaly Detection with Outlier Exposure,” ICLR, 2019. arXiv
  39. J. Zhang, J. Yang, P. Wang, H. Wang, Y. Lin, H. Zhang et al., “OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection,” arXiv:2306.09301, 2023. arXiv
  40. P. Bergmann, M. Fauser, D. Sattlegger and C. Steger, “MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,” CVPR, 2019. CVF
  41. L. Heckler-Kram, J.-H. Neudeck, U. Scheler, R. König and C. Steger, “The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection,” International Journal of Computer Vision, vol. 134, no. 4, 2026. doi · arXiv
  42. K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox and P. Gehler, “Towards Total Recall in Industrial Anomaly Detection,” CVPR, 2022. CVF
  43. J. Bitterwolf, M. Müller and M. Hein, “In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation,” ICML, PMLR 202, 2023. PMLR
  44. J. Liang, R. He and T. Tan, “A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts,” International Journal of Computer Vision, vol. 133, no. 1, pp. 31–64, 2025. doi · arXiv
  45. N. Akhtar, A. Mian, N. Kardan and M. Shah, “Advances in Adversarial Attacks and Defenses in Computer Vision: A Survey,” IEEE Access, vol. 9, pp. 155161–155196, 2021. doi · arXiv
  46. G. Pang, C. Shen, L. Cao and A. van den Hengel, “Deep Learning for Anomaly Detection: A Review,” ACM Computing Surveys, vol. 54, no. 2, pp. 1–38, 2021. doi · arXiv
  47. J. Liu, G. Xie, J. Wang, S. Li, C. Wang, F. Zheng and Y. Jin, “Deep Industrial Image Anomaly Detection: A Survey,” Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024. doi · arXiv
  48. Z. Wang, T. Pang, C. Du, M. Lin, W. Liu and S. Yan, “Better Diffusion Models Further Improve Adversarial Training,” ICML, PMLR 202, 2023. PMLR
  49. J. Yang, P. Wang, D. Zou, Z. Zhou, K. Ding, W. Peng et al., “OpenOOD: Benchmarking Generalized Out-of-Distribution Detection,” NeurIPS Datasets and Benchmarks Track, 2022. proceedings
  50. J. Jeong, Y. Zou, T. Kim, D. Zhang, A. Ravichandran and O. Dabeer, “WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation,” CVPR, 2023. CVF