Unifying Privacy Guarantees Across Differential Privacy Variants

29 concepts27 papers1 proposal

At its core, differential privacy (DP) provides a mathematical guarantee that an algorithm's output does not reveal whether any single individual's data was included in the training set. To formalize this guarantee, researchers measure the "privacy loss" using divergences between probability distributions.

However, the choice of divergence metric has led to a fragmented ecosystem of privacy definitions. While pure differential privacy ($\epsilon$-DP) uses a strict max-divergence, it is often too pessimistic for iterative algorithms like deep learning. To track privacy loss more precisely over thousands of training steps, researchers introduced alternative frameworks: approximate DP $(\epsilon, \delta)$, Rényi Differential Privacy (RDP), Zero-Concentrated Differential Privacy (zCDP), and Gaussian Differential Privacy (GDP).

The central challenge in modern privacy research is that an algorithmic mechanism tightly analyzed under one framework does not automatically transfer its guarantees to another. This fragmentation forces researchers to re-prove foundational properties from scratch for every new privacy variant, creating a critical need for privacy framework generalization.

The Friction of Conversion Loss

A tempting solution to this fragmentation is simply to use analytical conversion formulas to translate guarantees from one framework to another. For instance, one might analyze an algorithm using RDP and then convert the final bound back into standard $(\epsilon, \delta)$-DP. However, these mathematical translations are rarely lossless.

This conversion loss has severe practical consequences. In the context of adversarial robustness, Andrew Craig Cullen demonstrates that while specialized frameworks like RDP can yield exceptionally tight pointwise certification bounds against data poisoning attacks, translating these gains back into the more general approximate DP framework incurs a mathematical penalty that fails to fully preserve the certified radii.

Similarly, information-theoretic analyses of statistical costs often lack a single, unified lens. As shown by John C. Duchi et al., evaluating the fundamental "privacy tax" on high-dimensional estimation remains difficult to compare across different relaxations without a unified framework that spans multiple privacy definitions.

Algorithmic Primitives Trapped in Silos

Many of the most powerful algorithmic building blocks in differential privacy are currently locked into specific definitions, limiting their compatibility with modern machine learning workflows.

Subsampling and Selection

Randomly selecting a subset of data inherently amplifies privacy because an adversary cannot be certain if a specific individual's data was even used. Borja Balle et al. developed a unified profile-based framework to compute exact amplification bounds for standard DP. However, translating these tight bounds to RDP or zCDP remains an open challenge, creating a bottleneck for deep learning pipelines. This limitation is echoed by Lynn Chua et al., whose novel balls-and-bins sampling strategy for DP-SGD is currently restricted to standard DP definitions.

A similar silo exists for selection mechanisms. Choosing the top-$k$ elements privately typically requires an expensive iterative "peeling" process. To bypass this, Gang Qiao et al. introduced the Oneshot Laplace Mechanism, proving it achieves tight approximate DP bounds without composition. Yet, their proof techniques are specialized for approximate DP, leaving compatibility with concentrated or Rényi DP unresolved.

Optimization and Sensitivity

Optimization strategies face identical hurdles. Optimal local mechanisms, such as the staircase mechanisms designed by Pramod Viswanath et al., are optimized for local DP but do not seamlessly extend to approximate global DP.

On the utility side, instance-optimal mechanisms developed by Hilal Asi et al. via approximate inverse sensitivity are tailored for pure DP, requiring further generalization to unlock tighter bounds for complex datasets under RDP. Even fundamental theoretical equivalences, such as the link between private distribution learning and sample compression established by Shai Ben-David et al., remain restricted to pure DP.

Composition, Auditing, and Vulnerability Bounds

The fragmentation of privacy frameworks also complicates how we track privacy over time (composition), how we protect specific data properties, and how we empirically verify guarantees.

+-----------------------------------------------------------------------+
|                      The Generalization Gap                           |
+-----------------------------------------------------------------------+
|  Foundational Primitives             Modern Deep Learning             |
|  (Pure / Approximate DP)             (Rényi / Gaussian DP)            |
|                                                                       |
|  - Martingale Filters   ======X====>  - Incompatible with GDP         |
|  - Sequential Auditing  ======X====>  - Incompatible with f-DP        |
|  - Inference Bounds     ======X====>  - Loose under Conversion        |
+-----------------------------------------------------------------------+

Dynamic Tracking and Composition

When running multiple private queries, an analyst must track the cumulative privacy loss. Justin Whitehouse et al. introduced fully-adaptive privacy filters and odometers using advanced martingale concentration, allowing analysts to alter privacy parameters dynamically on the fly. However, these filters lack support for GDP or $f$-DP, hindering their use in deep learning.

Furthermore, Salil P. Vadhan et al. uncovered unexpected complexities when moving from pure to approximate DP under concurrent composition, leaving the safety properties of modern variants like zCDP in multi-user environments uncertain.

Attack and Vulnerability Bounds

Theoretical limits on privacy attacks are similarly siloed. Membership inference attack bounds derived by Franziska Boenisch et al. and label inference leakage bounds by Ruihan Wu et al. are optimized for standard DP. To provide tighter guarantees for modern optimization algorithms, these frameworks must be generalized to RDP or GDP.

This extends to specialized settings as well: the label differential privacy framework for binary response models developed by Ce Zhang et al., the "accuracy-first" post-processing immunity framework by Antti Koskela et al., and the private multi-armed bandit algorithms by Achraf Azize et al. all face open questions regarding how their guarantees translate to alternative privacy metrics.

Bridging the Divide

To resolve this fragmentation, the research community is actively building unifying bridges and automated tools designed to operate natively across diverse privacy definitions.

One prominent approach focuses on creating flexible measurement frameworks. Yi Liu et al. introduced a theoretical toolkit that acts as a bridge to Gaussian Differential Privacy. By providing a unified transformation framework, their method allows for direct visual comparisons and translations between Concentrated, Rényi, and standard DP paradigms without sacrificing the tightness of the bounds.

Concurrently, researchers are turning to automated verification to bypass manual mathematical proofs entirely. Lisa Oakley et al. leverage weighted model counting and probabilistic programming to automatically synthesize tight privacy and accuracy bounds. While initially focused on specific definitions, expanding these automated systems to natively support RDP and approximate DP represents the next frontier in privacy framework generalization, promising a future where privacy guarantees are universally compatible.

Key Foundational Work

Privacy Amplification by Subsampling: Tight Analyses via Couplings and Divergences (2018) — contributed 1 concept elements to this direction.
Extremal Mechanisms for Local Differential Privacy (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_privacy_framework_generalization_ecc5574f,
  title = {privacy framework generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-framework-generalization/rdr_ecc5574feb0946be82c2d0b4cc56d55f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_privacy_framework_generalization_ecc5574f_v1,
  title = {privacy framework generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-framework-generalization/rdr_ecc5574feb0946be82c2d0b4cc56d55f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake