CoChat

Private Prediction Sets

Open access

The DOI matches its Crossref record; 78 of the 82 references it cites were found in a registry or an index; none of the 15 sources its claims cite is checked for support yet.

Abstract

In real-world settings involving consequential decision-making, the deployment of machine learning systems generally requires both reliable uncertainty quantification and protection of individuals’ privacy. We present a framework that treats these two desiderata jointly. Our framework is based on conformal prediction, a methodology that augments predictive models to return prediction sets that provide uncertainty quantification—they provably cover the true response with a user-specified probability, such as 90%. One might hope that when used with privately trained models, conformal prediction would yield privacy guarantees for the resulting prediction sets; unfortunately this is not the case. To remedy this key problem, we develop a method that takes any pretrained predictive model and outputs differentially private prediction sets. Our method follows the general approach of split conformal prediction; we use holdout data to calibrate the size of the prediction sets but preserve privacy by using a privatized quantile subroutine. This subroutine compensates for the noise introduced to preserve privacy in order to guarantee correct coverage. We evaluate the method on large-scale computer vision data sets.

Overview

CoChat Summary

TL;DR

The paper develops a differentially private version of split conformal prediction that turns any pretrained predictive model plus a calibration set into prediction sets with finite-sample coverage and ε-differential privacy. It shows the cost of privacy in calibration is mild, with coverage upper bound scaling as Õ((nε)^−1) versus O(n^−1) for nonprivate conformal prediction.

Key findings

  • The proposed private prediction-set procedure is ε-differentially private and satisfies coverage \(1-\alpha \le P\{Y \in C(X)\} \le 1-\alpha + O((n\epsilon)^{-1})\) (Theorem 1 informal preview).
  • A simplified upper bound gives \(P\{Y \in C(X)\} \le 1-\alpha + O(\log(n\epsilon/\alpha)/(n\epsilon))\), compared with \(1-\alpha + O(n^{-1})\) for nonprivate conformal prediction (Corollary 1).
  • In CIFAR-10 experiments with \(n=5000\), \(\epsilon=8\), \(\delta=1e-5\), and \(\alpha=0.1\), the nonprivate model had 73% accuracy and the private model 67%; private conformal prediction caused only a minuscule increase in set size, while private model training caused a larger set-size increase.
  • On ImageNet with a nonprivate ResNet-152, \(n=30000\) calibration points and 20000 evaluation images, coverage was nearly constant over three orders of magnitude of bin numbers and showed little sensitivity to \(\epsilon\); on CoronaHack, using a nonprivate ResNet-18 with 75% accuracy, median coverage was 90.4% at \(\alpha=0.1\).

Methods

  • The method follows split conformal prediction: compute conformity scores \(s_i = S_f(X_i,Y_i)\) on held-out calibration data, choose a cutoff \(\hat{s}\), and output \(C(x)=\{y:S(x,y)\le \hat{s}\}\).
  • Privacy is achieved by replacing the ordinary conformal quantile with a privatized quantile subroutine: scores are discretized into bins and the exponential mechanism is applied, extending the private median routine of Feldman and Steinke.
  • Experiments evaluate image classification on CIFAR-10, ImageNet, and CoronaHack; the score used is \(S(x,y)=1-f(x)_y\), and private model training on CIFAR-10 uses DP-SGD via Opacus with \(\epsilon=8\), \(\delta=1e-5\).

Limitations

  • The paper states that the coverage upper bound requires weak regularity on the score distribution, such as a continuous distribution with bounded density and no point masses; the lower bound holds for any distribution, and the paper notes the assumption could be enforced with tiebreaking noise.
  • The paper states that private model training causes a substantial accuracy decrease and larger prediction sets, and concludes that the cost of privacy lies primarily in model fitting rather than private calibration.
  • Evident from the text: the empirical evaluation is limited to image classification data sets (CIFAR-10, ImageNet, CoronaHack), and the private model training experiment is only on CIFAR-10; the method also requires a held-out calibration set not used to train \(f\), and the full pipeline is differentially private only if the underlying model is also trained differentially privately.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1162/99608f92.16c71dad

Cites

78 of 78 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Modular data science.
Figure 1 Modular data science.CC-BY · From the PDF
Figure 2 Examples of private conformal prediction sets on COVID-19 data. We show three examples of lung X-rays taken from the CoronaHack data set (Perez et al., 2020) with their corresponding private prediction sets at α = 10% from a ResNet-18. All three patients had viral pneumonia (likely COVID-19). The classes in the prediction sets appear in ranked order according to the softmax score of the model; the center and right images are incorrectly classified if the predictor returns only the most likely class, but are correctly covered by the private prediction sets. See Experiment 4.4 for details.
Figure 2 Examples of private conformal prediction sets on COVID-19 data. We show three examples of lung X-rays taken from the CoronaHack data set (Perez et al., 2020) with their corresponding private prediction sets at α = 10% from a ResNet-18. All three patients had viral pneumonia (likely COVID-19). The classes in the prediction sets appear in ranked order according to the softmax score of the model; the center and right images are incorrectly classified if the predictor returns only the most likely class, but are correctly covered by the private prediction sets. See Experiment 4.4 for details.CC-BY · From the PDF
Figure 3 The private quantile ˜ q as n and ϵ grow. We demonstrate the adjusted quantile from (3) as n and ϵ increase, with automatically chosen values for m and γ described in Appendix C. As the number of samples grows and the privacy constraint relaxes, the procedure chooses a less conservative quantile, eventually approaching the limiting value 1 -α . The mild fluctuations in the curves are due to differing choices of m ∗ and γ .
Figure 3 The private quantile ˜ q as n and ϵ grow. We demonstrate the adjusted quantile from (3) as n and ϵ increase, with automatically chosen values for m and γ described in Appendix C. As the number of samples grows and the privacy constraint relaxes, the procedure chooses a less conservative quantile, eventually approaching the limiting value 1 -α . The mild fluctuations in the curves are due to differing choices of m ∗ and γ .CC-BY · From the PDF
Image 4
CC-BY · From the PDF
Figure 5 Coverage and set size for different values of m . We demonstrate the performance on Imagenet of private conformal prediction using a nonprivate ResNet-152 as the base model at α = 0 . 1 and ϵ = 1. The coverage is nearly constant over three orders of magnitude of bin numbers. All of the histograms on the right hand side are overlapping. See Section 4.2 for details.
Figure 5 Coverage and set size for different values of m . We demonstrate the performance on Imagenet of private conformal prediction using a nonprivate ResNet-152 as the base model at α = 0 . 1 and ϵ = 1. The coverage is nearly constant over three orders of magnitude of bin numbers. All of the histograms on the right hand side are overlapping. See Section 4.2 for details.CC-BY · From the PDF
Figure 6 Coverage and set size for different values of ϵ . We demonstrate the performance on ImageNet of private conformal prediction using a nonprivate ResNet-152 as the base model with α = 0 . 1. The coverage improves slightly for liberal (large) ϵ , although the cost of privacy is evidently very low. All of the histograms on the right hand side are overlapping. See Section 4.3 for details.
Figure 6 Coverage and set size for different values of ϵ . We demonstrate the performance on ImageNet of private conformal prediction using a nonprivate ResNet-152 as the base model with α = 0 . 1. The coverage improves slightly for liberal (large) ϵ , although the cost of privacy is evidently very low. All of the histograms on the right hand side are overlapping. See Section 4.3 for details.CC-BY · From the PDF
Figure 7 Coverage and set size on the CoronaHack data set. We demonstrate the performance on the CoronaHack data set of private conformal prediction using a nonprivate ResNet-18 as the base model with α = 0 . 1. The median coverage was 90 . 4%. See Section 4.4 for details.
Figure 7 Coverage and set size on the CoronaHack data set. We demonstrate the performance on the CoronaHack data set of private conformal prediction using a nonprivate ResNet-18 as the base model with α = 0 . 1. The median coverage was 90 . 4%. See Section 4.4 for details.CC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 1 Not valid · 9 Not checked · 58 Own results and interpretations

33 more claims

Create a free account to see all 68 claims and to verify the 15 references still open.

See all claims — free account
  • It may state two separate claims as one. 14
  • Probably not a substantial claim. 16
  • Borderline: one of the tests for a substantial claim is unclear. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 78 references

Research works printed but not linked: 1

  1. [2]2. Abowd, J. M. (2018). The US census bureau adopts differential privacy. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 2867-2867).

Cited by

See all 9

Metadata

Authors
Angelopoulos, A. N., Bates, S., Zrnic, T., & Jordan, M. I.
Affiliations
University of California, Berkeley (US)
Venue
Harvard Data Science Review
Publisher
MIT Press
Published
2022-04-28
Volume
4
Issue
2
ISSN
  • 2644-2353 (ISSN-L)
  • 2688-8513
OpenAlex
W3129027643
Cited by
8
References
82
Citations by year
2026: 3 · 2025: 1 · 2024: 2 · 2023: 1 · 2022: 1
FWCI
0.41
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Harvard Data Science Review (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
Keywords
conformal prediction, split conformal prediction, differential privacy, prediction sets, uncertainty quantification, privacy-preserving machine learning, coverage guarantees, computer vision

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Private Prediction Sets. https://app.cochat.ai/papers/10.1162/99608f92.16c71dad
The paper
Angelopoulos, A. N., Bates, S., Zrnic, T., & Jordan, M. I. (2022). Private Prediction Sets. Harvard Data Science Review. https://doi.org/10.1162/99608f92.16c71dad

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.