acceptodds
Under review as a conference paper at ICLR 2027

Metric-Normalized Posterior Leakage (mPL): Attacker-Aligned Privacy for Joint Consumption

Abstract

Metric differential privacy (mDP) extends local differential privacy (LDP) by scaling noise to semantic distance, but many machine learning (ML) systems are consumed under joint observation, where model-agnostic, per-record guarantees can miss leakage from evidence aggregation. We introduce metric-normalized posterior leakage (mPL), an attacker-aligned, distance-calibrated measure of posterior-odds shift induced by releases, and show that for single or independent releases, uniformly bounding mPL is equivalent to mDP. Under joint observation, however, satisfying mDP may still leave mPL high because learned aggregators compound evidence across correlated items. To make control practical, we formalize probabilistically bounded mPL (PBmPL), which limits how often mPL may exceed a target budget, and we operationalize it via Adaptive mPL (AmPL), a trust-and-verify framework that perturbs, audits with a learned attacker, and adapts parameters (with optional Bayesian remapping) to balance privacy and utility. In a word-embedding case study, a learned joint adversary reveals mPL violations under joint consumption despite per-record mDP perturbations, whereas AmPL substantially lowers the frequency of such violations with low utility loss, indicating PBmPL as a practical, auditable protection for joint-consumption settings.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, Privacy-Preserving Machine Learning, Information Leakage, Adversarial Inference, Metric Differential Privacy, Text Perturbation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bBBiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes metric-normalized posterior leakage (mPL), which allows to solve some of the issues of metric differential privacy (mDP). In particular, the work focuses on joint-prior Bayesian attackers, where the joint observations of the samples are not independent and this allows an attacker to have a more effective attack in the mDP settings. The work proposes 2 new privacy definitions, which are bounded mPL, probabilistically bounded mPL (PBmPL) and additionally the work defines Adaptive mPL (AmPL) as a new mechanism for these settings. To validate their framework, they evaluate their framework on text embeddings to protect personally identifiable information (PII) and potentially identifiable information (PoII) and on Breast Cancer Wisconsin.

Critical strengths
  • The work proposes privacy metrics that overcome some of the limitations of mDP. In particular, the work focuses on the case where the records given to the mechanisms are correlated.
  • The introduction of the AmPL framework is a new and interesting privacy preserving strategy.
Critical weaknesses
  • It is unclear in which realistic cases this vulnerability practically applies. mDP is often used for geo-indistinguishability or in the text case to protect tokens or attributes. Given the very small reported violation ratios (< 10−610^-6), I am not sure how significant this issue is in practice and if commonly used mDP mechanisms suffer from this issue. Can you give some practical cases in which this new approach is useful and practical?
  • The AmPL framework seems to be using no provable protection against joint leakage. From my understanding the AmPL framework doesn’t imply bounded mLP or PBmPL. Additionally, the main privacy evaluations are done using posterior leakage violation ratio, which can be a reasonable additional metric, but I don’t think it is aligned with any standard privacy attacks, such as reconstruction or attribute inference (these appear only in the appendix and are not compared between EM and AmPL). Finally, there is no comparison and discussion with similar notions of privacy leakage and their corresponding mechanism (e.g. Pufferfish).
Questions for authors
  • How would this framework perform under realistic attackers? Overall the empirical evaluation is limited to 3 settings and 1 mechanism (the Exponential Mechanism), however, a much wider range of mechanisms exists. Additionally as the current framework doesn’t give provable protection against joint leakage. It would be interesting to understand how it compares with the broader set of empirical defenses.
  • What does bounded mPL add beyond a metric Pufferfish instantiation?
Reviewer_zghbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

Metric differential privacy (mDP) does not capture the inference risk that an adversary incurs by aggregating multiple correlated records. The paper proposes metric-normalized posterior leakage (mPL) to address this gap, measuring the risk as the metric-normalized change in posterior odds, and shows that uniformly bounding mPL is equivalent to mDP for single or independent releases, whereas this guarantee no longer holds once records are dependent. Since imposing the worst-case requirement under joint observation would require excessive noise, the paper introduces probabilistically bounded mPL (PBmPL), which bounds the probability of exceeding the mPL budget under a specified audit distribution. Finally, the paper presents an adaptive perturbation framework, Adaptive mPL (AmPL), which outputs a set of calibrated perturbation parameters and balances privacy against utility.

Critical strengths

The paper identifies a scope gap that is widely overlooked in metric-privacy practice: mDP's per-record guarantee does not extend to joint observation. When records within the same data stream are correlated, joint observation yields additional leakage, and mDP provides no guarantee for such leakage. To address this gap, the paper proposes a measure that characterizes it and presents a workable implementation framework.

Critical weaknesses

The organization of the paper makes it difficult to follow. Figures 2 and 3 largely restate the adjacent text rather than helping the reader grasp the pipeline — in the case of Figure 2, the caption itself is written as a paragraph of method prose. For Section 3, I suggest adding an algorithm table that summarizes the AmPL procedure, so that the pipeline can be followed directly.

Questions for authors

1.Are there other metrics and frameworks, comparable in spirit to mPL, that address the joint-observation setting? If such approaches exist, please incorporate them into the baseline experiments for comparison.

2.Within the AmPL framework, why is Step 4 (Bayesian Remapping) not integrated into Step 3, so that the parameter α is optimized in a unified manner?

Reviewer_iK2bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Metric differential privacy (mDP) generalizes local differential privacy (LDP) to metric spaces: the closer two inputs are, the weaker the required indistinguishability. However, mDP constrains each release individually, whereas an adversary can aggregate multiple correlated releases and reason about posterior beliefs. Per-release guarantees may therefore fail to reflect the true privacy risk.

The paper proposes metric-normalized posterior leakage (mPL) to quantify this leakage. To relax the worst-case requirement, it introduces probabilistically bounded mPL (PBmPL), which limits the probability that mPL exceeds the budget under a specified audit distribution. Building on this, it proposes Adaptive mPL (AmPL), which perturbs tokens of different sensitivity levels with different strengths, audits the result with a learned attacker, and uses the audit feedback to adjust the perturbation strengths. In embedding experiments on text datasets, the authors report that AmPL reduces the mPL violation ratio with small utility loss.

Critical strengths
  • The paper addresses an important problem: per-release guarantees no longer reflect the actual risk when releases are dependent.
  • Attacker-in-the-loop calibration is a practical and relatively new idea, and the authors construct examples to evaluate the proposed framework.
Critical weaknesses
  • Prior work has studied this problem, but it is discussed only briefly, and the experiments do not include baselines from it.
  • PBmPL is only an empirical audit against one particular attacker. It relies on assumptions (e.g., that the fitted prior is close to the true distribution) and lacks a rigorous theoretical guarantee.
Questions for authors
  • Could the authors compare with existing solutions to the same problem? Examples include a group privacy baseline that scales down \epsilon, Pufferfish-based mechanisms for correlated data [1], or an adaptation of the idea in [2] to this setting.
  • If the real data do not follow the structure assumed by the attacker, how many true violations would the audit miss? Could the authors test this on synthetic data whose structure differs from the attacker's prior?
  • Could the authors plot utility–violation curves for baselines, to show which achieves a lower violation ratio at the same utility?

[1] Song, Shuang, Yizhen Wang, and Kamalika Chaudhuri. "Pufferfish privacy mechanisms for correlated data." Proceedings of the 2017 ACM International Conference on Management of Data. 2017.

[2] Liu, Changchang, Supriyo Chakraborty, and Prateek Mittal. "Dependence makes you vulnberable: Differential privacy under dependent tuples." NDSS. Vol. 16. 2016.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.