Pith. sign in

REVIEW 3 major objections 5 minor 14 references

Differentially Private Conformal Prediction via Quantile Binary Search

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that conformal prediction can be made differentially private by reading off the calibration quantile from a noisy binary search instead of sorting the scores.

desk verdict The core idea has merit, but the central privacy-utility guarantee and the empirical comparison rest on two unproven steps; the paper needs major theoretical repair. read the letter →

arxiv 2507.12497 v1 pith:SZH3BX74 submitted 2025-07-15 stat.ME cs.LGstat.APstat.COstat.ML

classification stat.MEcs.LGstat.APstat.COstat.ML
keywords differentialprivacyconformalpredictionzero-concentratedprivatequantileestimationrandomizedbinarysearchsetsuncertaintyquantificationcalibratednonconformityscores
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that conformal prediction can be made differentially private without inflating the reported quantile: instead of sorting the calibration scores and adding noise that must be corrected upward, one runs a noisy binary search on the score interval and releases the midpoint it converges to. The resulting procedure, P-COQS, is $\rho$-zCDP because the Gaussian noise added to each range count is scaled to the number of search iterations. The paper's main theoretical result, Theorem 2, bounds the coverage of the resulting prediction sets between $1-\alpha-\tau/(n_{\mathrm{cal}}+1)$ and $1-\alpha+(\tau+1)/(n_{\mathrm{cal}}+1)$, where $\tau$ is the rank error of the private quantile, which holds with probability at least $1-\beta$. If those guarantees hold, privacy-protected uncertainty sets can be produced in tens of binary-search iterations, at empirical coverage close to the nominal $1-\alpha$ level, with smaller sets than the leading exponential-mechanism alternative.

What carries the argument

The load-bearing object is the noisy range-count oracle $\mathrm{NoisyRC}([a,\mathrm{mid}], S)$ inside Algorithm 1, a fixed-iteration binary search over the score interval $[a,b]$ that replaces each exact count of calibration scores below a midpoint with a Gaussian-perturbed count of variance $\lceil\log_2((b-a)/\delta)\rceil/(2\rho)$. The variance is scaled to the number of search iterations so that, by sequential composition of zero-concentrated differential privacy, the entire search, and therefore the released quantile $q_{\mathrm{DP}}$ used as the conformal threshold, satisfies $\rho$-zCDP. The second ingredient is the rank error $\tau$ of Proposition 1, which is the quantity that carries the utility analysis: it measures, with probability at least $1-\beta$, how far the private quantile's rank can sit from the target rank $r = \lceil (1-\alpha)(n_{\mathrm{cal}}+1) \rceil$, and it is exactly the term that appears in both coverage bounds of Theorem 2.

What would settle it

Take the paper's own worked example, scores on $[0,1]$, $\delta = 10^{-10}$, $\rho = 0.1$, $\beta = 0.01$, $n_{\mathrm{cal}} = 3000$, $\alpha = 0.1$, and run P-COQS many times on simulated scores from a known distribution. The paper claims coverage between roughly $0.885$ and $0.915$ with probability at least $0.99$; check whether the empirical coverage across runs violates that band more often than $\beta$, and whether the released quantile's rank error exceeds $\tau \approx 45$ more often than $1\%$ of runs. A systematic violation would show that the rank-error bound has not transferred to the continuous-score binary search, and Theorem 2 would not hold as stated.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms: the randomized binary-search quantile method of the prior work the paper adapts carries over to continuous nonconformity scores, and that carry-over is enough to make conformal prediction private. Algorithm 1 narrows an interval $[a,b]$ on the score scale, and at each midpoint calls a noisy range count $\mathrm{NoisyRC}([a,\mathrm{mid}], S) = \mathrm{card}(S\cap[a,\mathrm{mid}]) + \mathcal{N}\bigl(0, \lceil\log_2((b-a)/\delta)\rceil/(2\rho)\bigr)$. Since each count has $\ell^1$-sensitivity $1$ and there are $N = \lceil\log_2((b-a)/\delta)\rceil$ iterations, composition of zero-concentrated differential privacy makes the whole search $\rho$-zCDP; Proposition 1 then asserts that the returned midpoint has rank error $\tau = \sqrt{\lceil\log_2 u\rceil/\rho \cdot \log(2\lceil\log_2 u\rceil/\beta)}$ with probability at least $1-\beta$, where $u = (b-a)/\delta$. Theorem 2 converts that rank error into two-sided coverage bounds for the prediction set $\{Y : s(X_{\mathrm{test}},Y,\hat{f}) \le q_{\mathrm{DP}}\}$, trading the exact $\ge 1-\alpha$ guarantee of non-private conformal prediction for a quantifiable approximation. The empirical claim that motivates the trade is that the method targets the nominal coverage across simulations, CIFAR-10, ImageNet, and CoronaHack while returning prediction sets that are smaller than the current private alternative.

Load-bearing premise

The entire coverage guarantee rests on one carried-over step: the claim that the noisy binary search returns a quantile whose rank is within $\tau$ of the target rank, which the paper's Proposition 1 asserts from the integer-domain algorithm it adapts, bounding the noise in each individual count without showing that the search decisions confine the final midpoint to that rank error.

Editorial extensions

If this is right

  • Releasing prediction sets through P-COQS carries $\rho$-zCDP with no additional privacy cost beyond the calibration-quantile search itself, since the prediction set is a post-processing of the private quantile.
  • Users can compute explicit coverage bounds for their own settings: with probability at least $1-\beta$, the true coverage lies between $1-\alpha-\tau/(n_{\mathrm{cal}}+1)$ and $1-\alpha+(\tau+1)/(n_{\mathrm{cal}}+1)$, so the worst-case coverage loss is known in advance rather than hidden behind an inflation hyperparameter.
  • The search runs in $N = \lceil \log_2((b-a)/\delta)\rceil$ iterations, which in the paper's benchmarks reduces conformal-prediction runtime by roughly two orders of magnitude relative to the binning-and-inflation alternative.
  • Because no quantile inflation is applied, prediction sets come out smaller: across simulations, CIFAR-10, ImageNet, and CoronaHack, the paper reports average set sizes closer to the non-private conformal baseline while empirical coverage targets $1-\alpha$.
  • Coverage slack $\tau/(n_{\mathrm{cal}}+1)$ shrinks as calibration size and privacy budget $\rho$ grow, so the approximate guarantee tightens exactly in the regimes where private conformal prediction is most useful.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the only data-touching component is the noisy count oracle, a natural extension is to swap that oracle for a local-DP or distributed count mechanism, which would extend private conformal prediction to federated calibration without re-deriving the coverage argument.
  • The two-sided coverage bound means a practitioner can detect wasteful over-coverage as well as under-coverage; the paper does not emphasize this, but the bound could be used as a diagnostic for automatic privacy-budget allocation across multiple calibration searches.
  • The rank error $\tau$ grows with the size of the score interval $[a,b]$, so clipping nonconformity scores to tighter bounds before the search would buy accuracy for free, a cheap modification the experiments do not explore.
  • The simulation evidence points to a small-sample regime (calibration sizes around $n \le 200$ with a DP-trained model) where P-COQS visibly under-covers; a hybrid that applies a small fixed inflation to the private quantile only in that regime would trade a little efficiency for the guaranteed lower bound.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes P-COQS, a differentially private conformal prediction method that computes the calibration quantile with an adaptation of Huang et al.'s binary search with noisy range counts. The authors state a zCDP guarantee and a rank-error bound (Proposition 1), derive approximate coverage bounds for the resulting prediction sets (Theorem 2), and compare P-COQS with the exponential-mechanism method of Angelopoulos et al. on simulated and benchmark data.

Significance. If the theoretical results were valid, P-COQS would be a valuable contribution: it has a simple zCDP composition argument, a fixed number of binary-search iterations, and the experiments suggest smaller and more informative prediction sets than ExponQ while empirically staying near the target coverage. The paper also provides a useful template for deriving approximate coverage bounds from a DP quantile subroutine. Unfortunately, the load-bearing rank-error transfer is false as stated, and the privacy-budget equivalence used to compare with ExponQ is incorrect; the main theoretical and comparative claims therefore need substantial reworking.

major comments (3)
  1. [Section 4, Proposition 1] The proof of Proposition 1 bounds the deviation of each noisy count by tau and takes a union bound over the N calls, but it never proves that a binary search in which every query has error at most tau returns a midpoint whose rank is within tau of r. The implication is false for continuous scores with ties. For example, with ncal=3000, alpha=0.1 (so r=2700), 2700 scores equal to 0 and 300 scores equal to 0.9, a=0, b=1, delta=10^-10, rho=100 and beta=0.01, we have N=34 and tau is approximately 1.7. Suppose every noisy count equals the true count minus 1; each per-query error is 1<tau, yet queries below 0.9 return 2699<r and move the left endpoint upward, while queries above 0.9 return 2999>=r and move the right endpoint downward. The final midpoint can then lie just above 0.9 with positive probability, in which case all 3000 scores are at or below qDP and the rank error is 300. Thus Proposition 1's rank-error statement is not merely unproved; it is false.
  2. [Section 4, Theorem 2] Theorem 2 is stated as a deterministic coverage inequality, but Proposition 1 guarantees the rank-error bound only on an event of probability at least 1-beta. The proof silently conditions on that event, and the line 'by Proposition 1 we have that Algorithm 1 outputs a private quantile qDP with rank error tau ensuring that sr-tau <= sr <= sr+tau' does not make sense as written. A correct statement would have to condition explicitly on the good-noise event and would carry an additional beta term for the failure event. As it stands, the theorem overstates the guarantee, and the subsequent discussion and Remark 2 treat the bounds as unconditional.
  3. [Section 3.2 and Section 5] The assertion that a mechanism satisfies (epsilon,0)-DP if and only if it satisfies epsilon-zCDP is incorrect. The standard relationship is that pure epsilon-DP implies (epsilon^2/2)-zCDP, and rho-zCDP implies approximate (epsilon,delta)-DP with epsilon = rho + 2*sqrt(rho log(1/delta)). Because ExponQ is pure epsilon-DP and P-COQS uses Gaussian noise analyzed in rho-zCDP, fixing epsilon_CP in Tables 1-4 and in the benchmark experiments does not hold the two methods to the same privacy guarantee. The comparative privacy-utility claims and the numerical example in Remark 2 therefore rest on an invalid equivalence.
minor comments (5)
  1. [Section 4, Equation (2) and surrounding text] The notation u = b-a/delta should be u = (b-a)/delta; as written, the formulas for N and tau are ambiguous and can be negative for small delta.
  2. [Section 4, Proof of Proposition 1] The tail bound is written with 'tau >= sigma sqrt(2 ln(2N/beta))' but Proposition 1 then treats tau as a fixed value; the proof should clarify whether tau is an upper bound or an equality for the reported numbers.
  3. [Section 4, Proof of Theorem 2] The expression 'sr-tau <= sr <= sr+tau' is nonsensical; the intended statement is presumably about the rank of qDP among the sorted scores, and it should be written correctly.
  4. [Section 3.1] The text defines ceil(v) as 'the smallest integer greater than x'; this should be 'greater than or equal to x'.
  5. [Throughout] There are several typos and garbled phrases, including 'respresents', 'prediciton', 'non-comformity', and 'Hence, the latter implies the former' in Section 3.2, which should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation rests on external DP-quantile results, not on self-citations or fitted inputs.

full rationale

I walked the claimed derivation chain. Proposition 1 is explicitly imported from an external source ('adapting directly from Huang et al. (2021), the following proposition holds'), and Huang et al. (2021) is not prior work by the present authors; no self-citation is load-bearing. Theorem 2 then converts the rank-error tau from Proposition 1 into coverage bounds via exchangeability of the non-conformity scores; this is an ordinary theorem/lemma implication rather than a definitional equivalence that assumes the conclusion. The noisy count function in (2) is not fitted to the target coverage or to any outcome; no model parameter is calibrated to the experimental results. The paper also repeatedly presents its weaker theoretical guarantee as a trade-off rather than as a forced conclusion. The genuine soft spot is that Proposition 1's proof bounds per-query count noise and then asserts, without a fully developed argument, that the binary search's final midpoint inherits a rank error tau; that is a correctness/rigor gap in the transfer from count errors to rank error, not a circularity. The coverage theorem openly conditions its stated bounds on Proposition 1, and the external provenance of that proposition is independent evidence under the rules. The empirical comparisons against ExponQ and non-private conformal prediction also make the experimental claims self-contained against external benchmarks. Accordingly, no enumerated circularity pattern is present, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method adds no new entities or fitted constants. Its guarantees rest on user-set parameters (interval, delta, beta, rho), exchangeability, and the transfer of Huang et al.'s rank-error analysis to the adapted binary search. The privacy-budget equivalence between pure DP and zCDP is an unsupported load-bearing assumption used only in the comparisons.

free parameters (4)
  • Search interval [a,b] = [0, 1] in the experiments
    Hand-set domain for the binary search; the quantile guarantee only holds if all scores are inside the interval.
  • Privacy parameter rho (or epsilon) = 0.1 to 10 in experiments
    User-selected privacy budget; the coverage error bound and noise scale depend on it.
  • Tolerance delta = 1e-10
    Stopping threshold that fixes the number of binary search iterations N and therefore the noise variance.
  • Failure probability beta = 0.01
    Used in Proposition 1 and Remark 2 to state the probability with which the rank-error and coverage bounds hold.
assumptions (5)
  • domain assumption Exchangeability of the calibration and test non-conformity scores, so the rank of stest among the ncal+1 scores is uniform on {1,...,ncal+1}.
    Used in the proof of Theorem 2 (Section 4); standard split-conformal assumption, not verified on the benchmark data.
  • domain assumption Huang et al. (2021) rank-error bound transfers unchanged to Algorithm 1 for real-valued scores with additive Gaussian noise and a midpoint-valued quantile.
    Proposition 1 is stated by direct adaptation; the proof only bounds count deviations and does not fully derive the rank error of the final midpoint.
  • standard math Gaussian mechanism zCDP guarantee and composition: each NoisyRC call with variance N/(2rho) is rho/N-zCDP and N adaptive calls compose to rho-zCDP.
    Invoked in the proof of Proposition 1, following Bun and Steinke (2016).
  • ad hoc to paper Pure epsilon-DP is equivalent to epsilon-zCDP under the paper's Definition 1.
    Asserted in Sections 3.2 and 5 to compare ExponQ and P-COQS under identical budgets; no proof is given and the stated equivalence is not standard.
  • domain assumption Non-conformity scores lie in a known user-specified interval [a,b] used as the binary search domain.
    Algorithm 1 requires bounds; in experiments [a,b]=[0,1] for hinge scores, but the paper claims generality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Differentially Private Conformal Prediction via Quantile Binary Search." pith.science (2026). https://pith.science/paper/SZH3BX74

@misc{pith2026250712497,
  author       = {Pith},
  title        = {Pith review of: Differentially Private Conformal Prediction via Quantile Binary Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZH3BX74}},
  note         = {Machine review of arXiv:2507.12497}
}
abstract

Most Differentially Private (DP) approaches focus on limiting privacy leakage from learners based on the data that they are trained on, there are fewer approaches that consider leakage when procedures involve a calibration dataset which is common in uncertainty quantification methods such as Conformal Prediction (CP). Since there is a limited amount of approaches in this direction, in this work we deliver a general DP approach for CP that we call Private Conformity via Quantile Search (P-COQS). The proposed approach adapts an existing randomized binary search algorithm for computing DP quantiles in the calibration phase of CP thereby guaranteeing privacy of the consequent prediction sets. This however comes at a price of slightly under-covering with respect to the desired $(1 - \alpha)$-level when using finite-sample calibration sets (although broad empirical results show that the P-COQS generally targets the required level in the considered cases). Confirming properties of the adapted algorithm and quantifying the approximate coverage guarantees of the consequent CP, we conduct extensive experiments to examine the effects of privacy noise, sample size and significance level on the performance of our approach compared to existing alternatives. In addition, we empirically evaluate our approach on several benchmark datasets, including CIFAR-10, ImageNet and CoronaHack. Our results suggest that the proposed method is robust to privacy noise and performs favorably with respect to the current DP alternative in terms of empirical coverage, efficiency, and informativeness. Specifically, the results indicate that P-COQS produces smaller conformal prediction sets while simultaneously targeting the desired coverage and privacy guarantees in all these experimental settings.

Figures

Figures reproduced from arXiv: 2507.12497 by the authors.

Figure 1
Figure 1. Results of Standard, ExponQ and P-COQS methods applied to the CIFAR-10 dataset. (a) Boxplots with coverage distributions of the three methods under the non-DP model (left) and DP model (right) scenarios; (b) Boxplots with distributions of average set sizes of the three methods under the non-DP model (left) and DP model (right) scenarios. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Results of ExponQ and P-COQS methods applied to the ImageNet dataset. (a) Median coverage (bar height), mean coverage (red dots) and 95% confidence intervals for coverage (red vertical whiskers) for each method under different privacy budgets ϵ (red-dashed horizontal line is the target coverage); (b) Average prediction set sizes (average set sizes in 100 splits): median (bar height), mean (red dots) and 95% confiden… view at source ↗
Figure 3
Figure 3. Results of Standard, ExponQ and P-COQS methods applied to the CoronaHack dataset. (a) Boxplots with coverage distributions of the three methods under the non-DP model (left) and DP model (right) scenarios; (b) Distribution of prediction set sizes where the y-axis represents the empirical proportion of sets of size 1, 2 or 3 (represented on the x-axis) for each method: median proportion (bar height), mean proportion … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

  1. [1]

    A. N. Angelopoulos and S. Bates. A gentle introduction to conformal prediction and distribution- free uncertainty quantification. arXiv preprint arXiv:2107.07511 ,

  2. [2]

    The numbers are metric averages over 1000 runs (per method and privacy budget) and in parentheses is the corresponding variance of the metrics. 19 Romanus and Molinari B.2 Sample size effect with DP model Sample Size Coverage Efficiency Informativeness Model Accuracy ExponQ P-COQS ExponQ P-COQS ExponQ P-COQS 100 0.9979 (0.0142) 0.9669 (0.0800) 1.9977 (0.0...

  3. [5]

    Papadopoulos, K

    H. Papadopoulos, K. Proedrou, V. Vovk, and A. Gammerman. Inductive confidence machines for regression. In Machine learning: ECML 2002: 13th European conference on machine learning Helsinki, Finland, August 19–23, 2002 proceedings 13 , pages 345–356. Springer,

  4. [11]

    Simulation Results With Non-DP Random Forest Model A.1 Effect of ϵCP with non-DP model The parameter settings are as in the non-DP NB model and the model’s accuracy is 81%

    Appendix A. Simulation Results With Non-DP Random Forest Model A.1 Effect of ϵCP with non-DP model The parameter settings are as in the non-DP NB model and the model’s accuracy is 81%. ϵconformal Coverage Efficiency Informativeness ExponQ P-COQS ExponQ P-COQS ExponQ P-COQS 0.1 0.9998 (0.0004) 0.9024 (0.0104) 1.9885 (0.0123) 1.2222 (0.0243)0.0115 (0.0123) ...

  5. [13]

    The numbers are metric averages over 1000 runs (per method and privacy budget) and in parentheses is the corresponding variance of the metrics

    with DP Random Forest model. The numbers are metric averages over 1000 runs (per method and privacy budget) and in parentheses is the corresponding variance of the metrics. Appendix C. Simulation Results With Non-DP Na ¨ ıve Bayes Model The parameter settings are as described in Section 5.1 and the accuracy of the model is 83%. C.1 Effect of Privacy Budge...

  6. [14]

    The numbers are metric averages over 1000 runs (per method and α value) and in parentheses is the corresponding variance of the metrics

    with non-DP Na ¨ ıve Bayes model (ϵf = 2). The numbers are metric averages over 1000 runs (per method and α value) and in parentheses is the corresponding variance of the metrics. C.4 Effect of Model Privacy Budget ( ϵf ) ϵf Coverage Efficiency Informativeness Model Accuracy ExponQ P-COQS ExponQ P-COQS ExponQ P-COQS 0.1 0.9218 (0.0115) 0.8994 (0.0101)1.65...

  7. [2002]

    Penso, B

    C. Penso, B. Mahpud, J. Goldberger, and O. Sheffet. Privacy-preserving conformal prediction under local differential privacy. arXiv preprint arXiv:2505.15721 ,

  8. [2015]

    Vovk and I

    V. Vovk and I. Petej. Venn-abers predictors. arXiv preprint arXiv:1211.0025 ,

Show all 14 references
  1. [2016]

    Gibbs and E

    I. Gibbs and E. Cand` es. Adaptive conformal inference under distribution shift. arXiv preprint arXiv:2106.00170,

  2. [2017]

    R. J. Tibshirani, R. F. Barber, E. J. Cand` es, and A. Ramdas. Conformal prediction under covariate shift. arXiv preprint arXiv:1904.06019 ,

  3. [2018]

    M. Li, M. Klamkin, M. Tanneau, R. Zandehshahvar, and P. Van Hentenryck. Conformal prediction with upper and lower bound models. arXiv preprint arXiv:2503.04071 ,

  4. [2020]

    Shokri, M

    R. Shokri, M. Stronati, C. Song, and V. Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP) , pages 3–18. IEEE,

  5. [2021]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pages 248–255. Ieee,

  6. [2025]

    J. C. P´ erez, J. M. C. Contreras, C. de Blas P´ erez, and F. L. Alvarez. Databiology lab coronahack: Collection of public covid-19 data. bioRxiv, pages 2020–10,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.