Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Dependency Triad: A Metric to Quantify the Dependencies Between Attributes for Local Differential Privacy

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that three parameters—α, β, δ—suffice to summarize pairwise dependency information for local differential privacy, producing a constant-time conservative upper bound on correlation-induced leakage.

desk verdict The three-parameter compression is a real idea, but the shipped algorithm's core privacy guarantee is unproven because the LP-based beta is never shown to match the theorem's exact-calibration beta. read the letter →

arxiv 2608.03737 v1 pith:JKCGUAHB submitted 2026-08-04 cs.CR cs.ITmath.IT

Sandaru Jayawardana , Sennur Ulukus , Ming Ding , Kanchana Thilakarathna This is my paper
classification cs.CRcs.ITmath.IT
keywords DependencyTriadlocaldifferentialprivacycorrelation-inducedleakagebudgetcalibrationdistributionaluncertaintyconservativeestimatormultidimensionaldataconstant-timeCPLestimation
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

Local differential privacy promises that a user's released data leaks at most ε, but when attributes are correlated the release of one attribute can leak information about another, and measuring that correlation-induced leakage normally requires the true joint distribution and scales poorly with alphabet size. This paper tries to remove both barriers. It argues that for any pair of attributes, the dependency information relevant to this leakage—including uncertainty about the prior distribution—can be compressed into three scalars, α, β, and δ, which it calls the Dependency Triad. From those scalars a closed-form estimate of worst-case pairwise leakage is evaluated in constant time, and the paper proves it is an upper bound that never understates the risk. The payoff is practical: privacy-budget calibration becomes a one-time parameter computation followed by instant evaluation at any budget, and imperfect prior knowledge can be folded into the parameters through an entry-wise uncertainty matrix.

What carries the argument

The carrying object is the ratio vector Q=G⊘G′, the element-wise division of the two conditional distributions. The proof first relaxes all feasible likelihood ratios to the extreme interval [e^{-γ}, e^α] and solves the resulting linear-fractional problem, obtaining a closed-form bound (Theorem IV.1) that depends only on the extremes; a sparsity correction δ is then introduced (Theorem IV.3). The calibrated lower endpoint e^{-β} replaces e^{-γ} by matching the estimate to the exact CPL* at a chosen budget ε0, and Theorem IV.4 shows the resulting estimate is an upper bound for all larger budgets. Uncertainty in the known distribution enters as an entry-wise matrix Δ that widens α, β, δ to the

What would settle it

Take a pair of conditional distributions and an uncertainty matrix Δ, solve the exact maximization of Σ_{i∈S}(G_i−G′_i) over all feasible S,G,G′ by enumeration, and compare it with φ returned by the relaxed LP in Algorithm 1 (line 18). If φ is smaller than the exact maximum, compute DT's CPL estimate for ε>ε0 and compare with CPL* obtained by exhaustive search over S; an estimate below CPL* disproves the upper-bound claim for the algorithm as delivered.

Watch

Extended reading notes

Core claim

Given two correlated attributes X_k and X̂ with conditional probability vectors G and G′, the exact correlation-induced privacy leakage CPL* is the optimum of a ratio of linear forms over all subsets of X̂'s alphabet. The paper's central discovery is that the part of that distribution needed to upper-bound CPL* survives in just three scalars: α, the log of the largest (uncertainty-corrected) likelihood ratio; δ, the mass sitting on entries whose conditional denominator can be zero; and β, a calibrated lower ratio level obtained by matching the closed-form estimate to the exact CPL* at one chosen privacy budget ε0. With (α,β,δ) in hand, CPL is given by a piecewise closed-form expression, and

Load-bearing premise

The upper-bound guarantee is proven for a β calibrated against the exact CPL* at ε0, but Algorithm 1 actually derives β from a relaxed linear program whose solution φ is asserted, not proven, to be an upper bound on the calibration statistic; if φ falls below the true maximum, the delivered estimate could understate the true leakage.

Editorial extensions

If this is right

  • A one-time offline computation of (α,β,δ) costs O(a²b^3.5) and occupies O(1) space; after that, CPL at any privacy budget is a constant-time lookup, making iterative privacy-budget search practical for high-cardinality attributes.
  • Pairwise DT estimates compose through the sequential-composition bound into a total-leakage estimate for multidimensional records, so the triad serves as the building block, not the endpoint.
  • When several prior distributions are available, using the range or standard deviation as Δ yields uncertainty-aware per-attribute budgets that deliberately sit below the nominal optimum, hedging against demographic shifts.
  • Across five real datasets, DT's normalized error against exact CPL* stays below 0.025 for ε∈[0.01,10] and below 0.003 for ε≤1, so the conservative guarantee is tight where LDP is typically deployed.
  • In attribute-inference attack experiments at matched utility, DT achieves attack resistance comparable to simple budget splitting while consuming a smaller total privacy budget.

Reading between the lines

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

  • If the compression is as general as claimed, the same ratio-extremes-plus-calibration scheme should extend to approximate (ε,δ)-LDP mechanisms by adding a second slack term for the δ-privacy parameter; the paper lists this as future work, and the existing proof structure suggests the natural form.
  • Because β is calibrated at one anchor budget ε0, tightness is guaranteed to degrade as ε moves far above ε0; a testable extension is a piecewise triad with multiple anchor budgets, which would tighten the bound in low-privacy regimes without breaking the upper-bound property.
  • The two-sided sparsity slacks (δ̃, δ̂) derived in the proof are collapsed into a single δ for simplicity; on sparse or asymmetric distributions, keeping them separate would likely tighten the estimate while preserving the bound.
  • The empirical role split—β dominates leakage in the strong-privacy regime and α in the weak-privacy regime—suggests the triad could serve as an interpretable pairwise feature for selecting which attribute pairs to collect jointly or which are most exposed to attribute inference.
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

4 major / 5 minor

Summary. The paper proposes the Dependency Triad (DT), a three-parameter summary (α, β, δ) of a pairwise conditional distribution, intended to quantify correlation-induced privacy leakage (CPL) under local differential privacy. The authors extend the CPL optimization of [14] to settings with distributional uncertainty, give Algorithm 1 for computing DT (using a relaxed linear program), and give Algorithm 2 for constant-time CPL estimation. The central claims are that CPL-relevant pairwise dependency information can be compressed into three parameters, that the resulting estimator is a conservative upper bound on the exact CPL under distributional uncertainty, and that it scales to high-cardinality and sparse distributions. Experiments on synthetic and five real datasets compare DT with HCC-2/LTM, demonstrate privacy-budget calibration, attribute-inference resistance, and utility gains over random sampling.

Significance. If the theoretical claims are made rigorous, this is a useful contribution: it addresses a real limitation of existing CPL algorithms, which require ground-truth distributions and repeated per-budget computation. The paper ships an artifact link, includes extensive experiments on real datasets, and the idea of an uncertainty-aware, low-dimensional summary of CPL is appealing. However, the central privacy guarantee currently rests on several unproved or under-specified analytic steps, most importantly the relationship between the relaxed LP in Algorithm 1 and the exact calibration theorem. The contribution is therefore promising but conditional.

major comments (4)
  1. [§V-B, Algorithm 1 lines 18–22 vs. Theorem IV.4] Theorem IV.4 proves the upper-bound property for a β calibrated using exact CPL* at ε0. In the deployment path, Algorithm 1 obtains β from φ produced by the relaxed LP at line 18 and the formula (16). The paper asserts that this LP is a polynomial-time upper bound to the exact MILP, but no proof is given that (i) the LP optimum is at least max_{S,G,G'} Σ_{i∈S}(G_i−G'_i) over the uncertainty set, and (ii) the map φ ↦ β, and then β ↦ CPL, is monotone nondecreasing. Without (i)–(ii), the delivered (α,β,δ) may understate CPL, violating the core 'never understate' guarantee. This needs an explicit lemma with proof.
  2. [§IV-B, Theorem IV.1 and Appendix C] The proof sketch asserts that the maximum CPL over all G,G' sharing ratio vector Q equals the maximum over two-point instances supported only on the extreme ratios e^α and e^{-γ}. Appendix C proves the value for a relaxed constraint set (27), but the reduction from the exact ratio-constrained set H to the extreme-ratio set ilde H is only asserted: intermediate ratios with positive mass cannot simply be set to zero without changing Q, and the argument appears to yield a supremum rather than an attained maximum. Since Theorem IV.4 inherits this result, the equality needs a rigorous proof or a relaxation argument that establishes the upper bound directly.
  3. [§IV-C, Theorem IV.8 and Appendix J, Corollary J.1] The uncertainty-aware claim relies on Corollary J.1, which states that taking componentwise maxima of two characterizations gives an upper bound, justified by an 'easy' monotonicity of CPL in α, β, and δ. No proof of this monotonicity is provided, and it is not immediate from Corollary IV.7, which only addresses γ. This monotonicity is load-bearing: it justifies using the upper-bound φ from the LP and the entry-wise uncertainty box Δ. A proof for each parameter is needed.
  4. [§V-B, Algorithm 1 line 18] The LP constraint contains the denominator M_i^+ + M_i^-. When M_i^+ = M_i^- = 0, the constraint is 0 ≤ z_i ≤ 0/0, which is undefined. This occurs when U_i = L'_i and U'_i = L_i, in particular for zero-probability symbols with zero uncertainty—exactly the sparse-distribution regime the paper claims to support. The pseudocode must define a convention (e.g., drop the constraint or set z_i = 0) and state it formally.
minor comments (5)
  1. [§V-B, Algorithm 1 line 16] The line 'δ←min(1,max(δ, δ))' appears to contain a typo or variable shadowing; the global δ and the loop-local δ are confused. Please rename and clarify.
  2. [Figure 3 caption] The caption says 'Figure 3a and Figure 3a'; the second reference should be Figure 3b.
  3. [§IV-B3, toy example] The toy example states α=1.50, γ=1.38, δ=0 for both scenarios; it would help to state explicitly that α and γ are computed for a fixed ordered pair (x,x') and how the orientation that maximizes leakage is selected.
  4. [Table IV] The 'Near monotone likelihood ratio' row reports mean CPL estimation error 7.9e−4 at ε=0.1 but 2.5e−1 at ε=1 with a large standard deviation; the text should comment on this non-monotonicity in ε, since DT is intended to be tight in high-privacy regimes.
  5. [Throughout] Several theorem references in the text (e.g., 'Theorem IV.8' in the proof of Theorem IV.8, 'Theorem F' in Appendix F) are inconsistent or refer to the appendix by its letter. Please normalize numbering.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: default DT's beta is fitted to CPL* at the anchor epsilon0->0, so low-epsilon agreement is by construction; the upper-bound theorem supplies independent content.

  1. fitted input called prediction [Section IV-B4 (Theorem IV.4 and Corollary IV.5); Section VI-C; Definition V.1 Eq. (16)]
    "We first calibrate β by equating the analytical CPL expression to the exact leakage CPL* at the calibration budget ε0 and solving for β. ... DT corresponds to the (α, β, δ) method calibrated at ε0 = 0."

    Theorem IV.4 defines β by inverting the estimator expression so that the estimated CPL equals CPL* at ε0. In the default case ε0→0 (Corollary IV.5), β is set from φ = max_{S,G,G'} Σ_{i∈S}(G_i−G'_i), which is exactly the derivative of CPL* at ε=0: d/dε ln((1+A(e^ε−1))/(1+B(e^ε−1)))|_{0} = A−B, maximized by S={i:G_i>G'_i}. Thus the DT estimate is forced to match CPL* (and its slope) at the anchor by construction; agreement near ε0 is tautological, not predictive. The upper bound for ε>ε0 is a separate theorem, so the circularity is partial, but validation in the low-ε regime partly reflects the fitted anchor.

full rationale

The central construction is a calibrated bound: β is deliberately chosen so the estimator interpolates the exact CPL* at the anchor budget ε0 (default ε0→0). This makes the estimator's low-ε behavior match CPL* by construction, so claims like 'DT consistently estimates CPL' in that regime are not independent confirmations. However, the paper's main privacy guarantee—that the estimate upper-bounds CPL* for all ε≥ε0—is a separate mathematical claim (Theorem IV.4) and does not reduce to the fit; the same holds for the uncertainty-aware α,β,δ in Theorem IV.8. The paper also relies on the same authors' [14] for the CPL* optimization baseline and Algorithm 3; this is load-bearing but is a concrete, checkable optimization rather than an unverified uniqueness theorem, so self-citation alone is not circular. A genuine correctness gap is that Algorithm 1's β uses a relaxed LP (line 18) asserted, but not proven, to upper-bound the exact MILP, and a possible 0/0 in M_i^+=M_i^-=0; these are omitted proofs/robustness issues, not circularity. Overall: one step of fitted-input-called-prediction at the calibration anchor, with independent theorem content elsewhere; score 4.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The DT parameters alpha and delta are genuine summary statistics, but beta is a calibrated parameter that is fitted in the theory to the exact CPL* at epsilon0. This makes the 'three-parameter summary' depend on a global LP over the full distribution and on a designer-chosen calibration point, increasing the circularity burden. The upper-bound guarantee of the implemented algorithm rests on an unproven equivalence between the LP output and the calibrated beta.

free parameters (3)
  • beta (calibrated lower-level parameter) = set by epsilon0; default epsilon0 -> 0
    beta is chosen so the estimator matches CPL* at the calibration budget epsilon0; the default DT therefore interpolates the exact leakage at epsilon = 0.
  • calibration budget epsilon0 = 0 (default)
    Designer-chosen point at which the estimator is forced to equal CPL*; controls tightness for epsilon > epsilon0.
  • uncertainty matrix Delta = heuristics: e * P_K (e in 0.1..0.2), range, or std
    No principled selection rule; experiments justify 10-20% naive margin and range/std for multi-source priors.
assumptions (5)
  • domain assumption The maximum CPL between two attributes can be written as maximizing a ratio of linear forms with coefficients e^epsilon and 1 over subsets S (Eq. 6).
    Inherited from prior work [14]; restricts the LDP mechanism to two-level perturbation for the worst case.
  • domain assumption Total privacy leakage of an attribute is upper bounded by epsilon plus the sum of pairwise CPLs (Eq. 3).
    Uses sequential composition; pairwise decomposition is the tractable building block.
  • ad hoc to paper The set H of all distributions sharing a likelihood-ratio vector Q has its maximum CPL attained by a two-point distribution supported on the extreme ratios e^alpha and e^-gamma.
    Central to Theorem IV.1; only a proof sketch is provided and the construction can be delicate when probabilities are zero.
  • ad hoc to paper The phi computed by the relaxed LP in Algorithm 1 yields a beta that satisfies the calibration upper-bound property of Theorem IV.4.
    Not proven in the paper; the LP is a heuristic relaxation and its objective does not obviously match the calibration statistic A-tilde minus B-tilde.
  • ad hoc to paper The true distribution lies within an entry-wise uncertainty box Delta around the known distribution.
    The choice of Delta (range, std, or percentage) is heuristic and not derived from a statistical model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dependency Triad: A Metric to Quantify the Dependencies Between Attributes for Local Differential Privacy." pith.science (2026). https://pith.science/paper/JKCGUAHB

@misc{pith2026260803737,
  author       = {Pith},
  title        = {Pith review of: Dependency Triad: A Metric to Quantify the Dependencies Between Attributes for Local Differential Privacy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKCGUAHB}},
  note         = {Machine review of arXiv:2608.03737}
}
read the original abstract

Collecting multidimensional user data is essential for extracting rich insights across various applications. Local Differential Privacy (LDP) has emerged as a de facto standard for mitigating privacy risks in such scenarios. A key challenge in privacy-preserving multidimensional data collection lies in inter-attribute dependencies, as they can inadvertently reveal correlated information and increase privacy vulnerabilities. Therefore, accurately measuring correlation-induced privacy leakage (CPL) is essential for privacy analysis and privacy-utility trade-off. However, existing CPL analysis solutions either require accurate prior knowledge or face scalability challenges for large numbers of attributes and high-cardinality attributes. These limit their practical applicability in real data. To address this research gap, we propose a novel metric, ``Dependency Triad'' (DT), which summarizes the pairwise dependency information relevant to CPL using three parameters and yields a \emph{constant-time} conservative estimator of pairwise CPL. DT explicitly models uncertainty in prior distributional knowledge through its parameters, delivering robust leakage estimates. Moreover, its robustness to sparse distributions makes it particularly suitable for high-cardinality attributes, while the pairwise formulation serves as a tractable building block for assessing total leakage in multidimensional settings. Extensive experiments on both synthetic and real datasets demonstrate that DT consistently estimates CPL across diverse dependency regimes and prior uncertainties.

Figures

Figures reproduced from arXiv: 2608.03737 by the authors.

Figure 1
Figure 1. Motivational example. ε denotes the maximum privacy leakage allowed for each attribute collected. Motivation Example - Suppose we conduct a state-wise household income survey in the United States under LDP. Data from an initial subset of states (e.g., five states) can be collected using correlation-agnostic baselines such as simple privacy–budget splitting (SPL) or random sampling [27]. As the survey progresses, we … view at source ↗
Figure 2
Figure 2. (a) An example for sorted (ascending order) ratio [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Results of the toy example. (a) LXˆ→Xk vs ε for Scenario 1. (b) LXˆ→Xk vs ε for Scenario 2. Proof sketch. For a fixed subset S, we aggregate the variables as g = P i∈S Gi and g ′ = P i∈S G′ i , reducing the orig￾inal multidimensional problem to a one-dimensional linear￾fractional optimization. The constraints provide two affine upper bounds on g in terms of g ′ , and monotonicity shows that the optimum occurs at a b… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Heatmap of the conditional distribution P(Xˆ | Xk) for the synthetic dataset, where both Xk and Xˆ have an alphabet size of 20. 4) Evaluation Metrics: We mainly use four metrics to evaluate DT: (i) TCPL, (ii) CPL estimation error, (iii) NMSE￾CPL, and (iv) total variati…
Figure 5
Figure 5. Figure 5: CPL analysis on two attribute synthetic datasets. Here, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 7
Figure 7. Figure 7: Interpreting CPL Through DT Parameters using all five [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: (a) Evaluation of DT under distributional uncertainty [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 10
Figure 10. Figure 10: TV distance (TVD) versus privacy budget ε for k-marginal estimation with k = 2, 3, 4. Lower TVD indicates better estimation utility. CPL-aware budget-tuning methods achieve substantially lower TVD than SPL and remain competitive with RS, with clearer advantages for hi…
Figure 11
Figure 11. Figure 11: Empirically validate the space and time complexities of DT. Here, space is measured with the Python object allocated [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Normalized over estimation and under estimation of [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Normalized overestimation and underestimation of estimated CPL with respect to ground truth [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 44 canonical work pages

  1. [14]

    The Hidden Cost of Correlation: Rethinking Privacy Leakage in Local Differential Privacy

    S. Jayawardana, S. Ulukus, M. Ding, and K. Thilakarathna, “The hidden cost of correlation: Rethinking privacy leakage in local differential privacy,” 2025. [Online]. Available: https://arxiv.org/abs/2508.12539

    work page Pith review arXiv 2025

  2. [26]

    Quantifying differential privacy under temporal correlations,

    Y . Cao, M. Yoshikawa, Y . Xiao, and L. Xiong, “Quantifying differential privacy under temporal correlations,” inICDE, 2017, pp. 821–832

    work page 2017

  3. [1]

    Rappor: Randomized aggre- gatable privacy-preserving ordinal response,

    U. Erlingsson, V . Pihur, and A. Korolova, “Rappor: Randomized aggre- gatable privacy-preserving ordinal response,” inProceedings of the 2014 ACM SIGSAC CCS, 2014

    work page 2014

  4. [2]

    L-srr: Local differential privacy for location-based services with staircase randomized response,

    H. Wang, H. Hong, L. Xiong, Z. Qin, and Y . Hong, “L-srr: Local differential privacy for location-based services with staircase randomized response,” in2022 ACM SIGSAC, ser. CCS ’22, p. 2809–2823

    work page

  5. [3]

    A survey of differential privacy-based techniques and their applicability to location-based services,

    J. W. Kim, K. Edemacu, J. S. Kim, Y . D. Chung, and B. Jang, “A survey of differential privacy-based techniques and their applicability to location-based services,”Computers Security, vol. 111, 2021

    work page 2021

  6. [4]

    PCKV: Locally differentially private correlated Key-Value data collection with optimized utility,

    X. Gu, M. Li, Y . Cheng, L. Xiong, and Y . Cao, “PCKV: Locally differentially private correlated Key-Value data collection with optimized utility,” inUSENIX Security 20, 2020

    work page 2020

  7. [5]

    Local differential privacy and its applications: A comprehensive survey,

    M. Yang, T. Guo, T. Zhu, I. Tjuawinata, J. Zhao, and K.-Y . Lam, “Local differential privacy and its applications: A comprehensive survey,” Computer Standards Interfaces, 2024

    work page 2024

  8. [6]

    Data privacy laws and their impact on financial technology companies: A review,

    A. T. Oyewole, B. B. Oguejiofor, N. E. Eneh, C. U. Akpuokwe, and S. S. Bakare, “Data privacy laws and their impact on financial technology companies: A review,”Computer Science & IT Research Journal, 2024

    work page 2024

Show all 50 references
  1. [7]

    Securema: protecting participant privacy in genetic association meta-analysis,

    W. Xie, M. Kantarcioglu, W. S. Bush, D. C. Crawford, J. C. Denny, R. Heatherly, and B. A. Malin, “Securema: protecting participant privacy in genetic association meta-analysis,”Bioinformatics, vol. 30 23, pp. 3334–41, 2014

    2014

  2. [8]

    Local privacy and statistical minimax rates,

    J. C. Duchi, M. I. Jordan, and M. J. Wainwright, “Local privacy and statistical minimax rates,” in2013 51st Annual Allerton Conference on Communication, Control, and Computing, 2013, pp. 1592–1592

    2013

  3. [9]

    Privacy at scale: Local differential privacy in practice,

    G. Cormode, S. Jha, T. Kulkarni, N. Li, D. Srivastava, and T. Wang, “Privacy at scale: Local differential privacy in practice,” ser. SIGMOD ’18. ACM, 2018, p. 1655–1658

    2018

  4. [10]

    Collecting telemetry data privately,

    B. Ding, J. Kulkarni, and S. Yekhanin, “Collecting telemetry data privately,” ser. NIPS’17, Red Hook, NY , USA, 2017, p. 3574–3583

    2017

  5. [11]

    Learning with privacy at scale,

    A. Inc., “Learning with privacy at scale,” https://machinelearning.apple. com/research/learning-with-privacy-at-scale, 2017

    2017

  6. [12]

    Calm: Consistent adaptive local marginal for marginal release under local differential pri- vacy,

    Z. Zhang, T. Wang, N. Li, S. He, and J. Chen, “Calm: Consistent adaptive local marginal for marginal release under local differential pri- vacy,” inACM SIGSAC Conference on Computer and Communications Security, ser. CCS ’18, 2018, p. 212–229

    2018

  7. [13]

    No free lunch in data privacy,

    D. Kifer and A. Machanavajjhala, “No free lunch in data privacy,” ser. SIGMOD ’11. ACM, 2011, p. 193–204

    2011

  8. [15]

    On the risks of collecting multidimensional data under local differential privacy,

    H. H. Arcolezi, S. Gambs, J.-F. Couchot, and C. Palamidessi, “On the risks of collecting multidimensional data under local differential privacy,”Proc. VLDB Endow., vol. 16, no. 5, p. 1126–1139, 2023

    2023

  9. [16]

    Random sampling plus fake data: Multidimensional frequency estimates with local differential privacy,

    H. H. Arcolezi, J.-F. Couchot, B. Al Bouna, and X. Xiao, “Random sampling plus fake data: Multidimensional frequency estimates with local differential privacy,” inProceedings of the 30th ACM CIKM, 2021, p. 47–57

    2021

  10. [17]

    Impact of prior knowledge and data correlation on privacy leakage: A unified analysis,

    Y . Li, X. Ren, S. Yang, and X. Yang, “Impact of prior knowledge and data correlation on privacy leakage: A unified analysis,”IEEE Transactions on Information Forensics and Security, vol. 14, no. 9, pp. 2342–2357, 2019

    2019

  11. [18]

    Collecting high-dimensional and correlation-constrained data with local differential privacy,

    R. Du, Q. Ye, Y . Fu, and H. Hu, “Collecting high-dimensional and correlation-constrained data with local differential privacy,” in2021 18th Annual IEEE SECON, 2021, pp. 1–9

    2021

  12. [19]

    Aaa: An adaptive mechanism for locally differentially private mean estimation,

    F. Wei, E. Bao, X. Xiao, Y . Yang, and B. Ding, “Aaa: An adaptive mechanism for locally differentially private mean estimation,”Proc. VLDB Endow., vol. 17, no. 8, p. 1843–1855, 2024

    2024

  13. [20]

    Local differential privacy for correlated location data release in its,

    K. M. Chong and A. Malip, “Local differential privacy for correlated location data release in its,”Computer Networks, vol. 255, p. 110830, 2024

    2024

  14. [21]

    Correlated differential privacy: Hiding information in non-iid data set,

    T. Zhu, P. Xiong, G. Li, and W. Zhou, “Correlated differential privacy: Hiding information in non-iid data set,”IEEE Transactions on Informa- tion Forensics and Security, vol. 10, no. 2, pp. 229–242, 2015

    2015

  15. [22]

    T. M. Cover and J. A. Thomas,Entropy, Relative Entropy, and Mutual Information. John Wiley and Sons, Ltd, 2005, ch. 2, pp. 13–55. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10. 1002/047174882X.ch2

    2005

  16. [23]

    User’s guide to correlation coefficients,

    H. Akoglu, “User’s guide to correlation coefficients,”Turkish Journal of Emergency Medicine, vol. 18, no. 3, pp. 91–93, 2018

    2018

  17. [24]

    Locally differentially private frequent itemset mining,

    T. Wang, N. Li, and S. Jha, “Locally differentially private frequent itemset mining,” in2018 IEEE Symposium on Security and Privacy (SP), 2018, pp. 127–143

    2018

  18. [25]

    The staircase mechanism in differential privacy,

    Q. Geng, P. Kairouz, S. Oh, and P. Viswanath, “The staircase mechanism in differential privacy,”IEEE Journal of Selected Topics in Signal Processing, vol. 9, pp. 1176–1184, 2015

    2015

  19. [27]

    Locally differentially private protocols for frequency estimation,

    T. Wang, J. Blocki, N. Li, and S. Jha, “Locally differentially private protocols for frequency estimation,” inUSENIX Security 17, 2017, pp. 729–745

    2017

  20. [28]

    ACS Supplemental Poverty Measure (SPM) Research Files: 2009–2019, 2021–2023,

    U.S. Census Bureau, “ACS Supplemental Poverty Measure (SPM) Research Files: 2009–2019, 2021–2023,” https://www.census. gov/data/datasets/time-series/demo/supplemental-poverty-measure/ acs-research-files.html, 2025

    2009

  21. [29]

    Dss 2016 dataset,

    A. U. R. I. N. (AURIN), “Dss 2016 dataset,” https://data.gov.au/data/ dataset/dss-payment-demographic-data

    2016

  22. [30]

    A comprehensive survey on local differential privacy toward data statistics and analysis,

    T. Wang, X. Zhang, J. Feng, and X. Yang, “A comprehensive survey on local differential privacy toward data statistics and analysis,”Sensors, vol. 20, no. 24, 2020

    2020

  23. [31]

    Computability of global solutions to factorable nonconvex programs: Part i—convex underestimating problems,

    G. P. McCormick, “Computability of global solutions to factorable nonconvex programs: Part i—convex underestimating problems,”Math- ematical Programming, vol. 10, no. 1, pp. 147–175, 1976

    1976

  24. [32]

    Boyd and L

    S. Boyd and L. Vandenberghe,Convex Optimization. Cambridge University Press, 2004

    2004

  25. [33]

    Deep learning face attributes in the wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” inICCV, December 2015

    2015

  26. [34]

    Becker and R

    B. Becker and R. Kohavi, “Adult,” UCI Machine Learning Repository, 1996, DOI: https://doi.org/10.24432/C5XW20

    1996 doi

  27. [35]

    Cardiovascular disease dataset,

    S. Sulianova, “Cardiovascular disease dataset,” https://www.kaggle.com/ datasets/sulianova/cardiovascular-disease-dataset, 2021

    2021

  28. [36]

    Macro f1 and macro f1,

    J. Opitz and S. Burst, “Macro f1 and macro f1,” 2021. [Online]. Available: https://arxiv.org/abs/1911.03347

    2021 arXiv

  29. [37]

    The theory of decision procedures for distri- butions with monotone likelihood ratio,

    S. Karlin and H. Rubin, “The theory of decision procedures for distri- butions with monotone likelihood ratio,”The Annals of Mathematical Statistics, vol. 27, pp. 272–299, 1956

    1956

  30. [38]

    Differential privacy,

    C. Dwork, “Differential privacy,” inAutomata, Languages and Program- ming, M. Bugliesi, B. Preneel, V . Sassone, and I. Wegener, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2006, pp. 1–12

    2006

  31. [39]

    Achieving k-anonymity privacy protection using gener- alization and suppression,

    L. Sweeney, “Achieving k-anonymity privacy protection using gener- alization and suppression,”Int. J. Uncertain. Fuzziness Knowl.-Based Syst., vol. 10, no. 5, p. 571–588, Oct. 2002

    2002

  32. [40]

    L-diversity: privacy beyond k-anonymity,

    A. Machanavajjhala, J. Gehrke, D. Kifer, and M. Venkitasubramaniam, “L-diversity: privacy beyond k-anonymity,” inICDE’06, 2006, pp. 24– 24

    2006

  33. [41]

    The algorithmic foundations of differential privacy,

    C. Dwork and A. Roth, “The algorithmic foundations of differential privacy,”Foundations and Trends® in Theoretical Computer Science, vol. 9, pp. 211–407, 2014

    2014

  34. [42]

    Enhancing correlated big data privacy using differential privacy and machine learning,

    S. Biswas, A. Fole, N. Khare, and P. Agrawal, “Enhancing correlated big data privacy using differential privacy and machine learning,”Journal of Big Data, vol. 10, no. 1, p. 30, 2023

    2023

  35. [43]

    Differentially private autocor- relation time-series data publishing based on sliding window,

    J. Zhao, S. Liu, X. Xiong, and Z. Cai, “Differentially private autocor- relation time-series data publishing based on sliding window,”Secur. Commun. Networks, vol. 2021, pp. 6 665 984:1–6 665 984:10, 2021

    2021

  36. [44]

    Ldp-based social content protection for trending topic recommendation,

    J. Wei, J. Li, Y . Lin, and J. Zhang, “Ldp-based social content protection for trending topic recommendation,”IEEE IoT Journal, vol. 8, no. 6, pp. 4353–4372, 2021

    2021

  37. [45]

    Tabtransformer: Tabular data modeling using contextual embeddings,

    X. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, “Tabtransformer: Tabular data modeling using contextual embeddings,” 2020. [Online]. Available: https://arxiv.org/abs/2012.06678 APPENDIX OPENSCIENCE Artifacts are available in https://github.com/ SandaruJayawardana/dependency-...

    2020 arXiv

  38. [46]

    Sinceα 1 ≥1andγ 1 ≥1, denominator is positive (i.e.,λ 1, λ2 >0)

    +e α1 (eγ1 −1). Sinceα 1 ≥1andγ 1 ≥1, denominator is positive (i.e.,λ 1, λ2 >0). Therefore, if and only ifα 1 ≥γ 1 thenU 1 −U 2 ≥0. This completes the proof. APPENDIXE PROOF OFTHEOREMIV.3 Proof.Based on new constraints in (9), we can formulate the updated optimization problem ...

  39. [47]

    ComputeV ⋆:We check the valid value range ofg ′ and howV(·)behaves withg ′ for g= eγ +ˆδ+g ′−1 eγ and g=e αg′ + ˜δ. •If g= eγ +ˆδ+g ′−1 eγ then V(g ′, δ, ε, α) =1 +e −γ(eγ + ˆδ+g ′ −1)(e ε −1) 1 +g ′(eε −1) .(41) Sincee −γ(eε −1)−(e ε −1)(1 +e −γ(eγ + ˆδ−1)(e ε −1)) = −e−γ(eε ...

  40. [48]

    Ifg ′ ∈[0, eγ −1−˜δeγ +ˆδ eα+γ −1 ]then g=e αg′ + ˜δ

  41. [49]

    Furthermore,e αg′ + ˜δ= eγ +ˆδ+g ′−1 eγ atg ′ = eγ −1−˜δeγ +ˆδ eα+γ −1

    Ifg ′ ∈[ eγ −1−˜δeγ +ˆδ eα+γ −1 ,1]then g= eγ +ˆδ+g ′−1 eγ . Furthermore,e αg′ + ˜δ= eγ +ˆδ+g ′−1 eγ atg ′ = eγ −1−˜δeγ +ˆδ eα+γ −1 . Therefore, we can computeV ⋆ based on the value of eγ −1−˜δeγ +ˆδ eα+γ −1 under three cases. Case 1 - ( eγ −1−˜δeγ +ˆδ eα+γ −1 ≤0)Sinceg ′ ≥0, ...

  42. [50]

    Interestingly, each of the above cases has attainableG andG ′

    Optimality ofV ⋆:Next, let us evaluate the optimality ofV ⋆. Interestingly, each of the above cases has attainableG andG ′. For example, consider the case eγ −1− ˜δeγ + ˆδ eα+γ −1 ≤0. In this scenario, we selectG ′ i for alli∈[b]\Sas G′ i =e γGi + ˆδi, such that P i∈[b]\S G′ i...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.