Subspace Differential Privacy for Unbiased Data Release

This direction is merged into another direction: post-processing bias. Its last version stays readable and citable here.

Version 1 of 19 papers15 conceptsfrozen 2026-10-04

Differential Privacy (DP) is the mathematical gold standard for protecting individual privacy in large datasets. By adding carefully calibrated noise to statistical queries, DP ensures that the presence or absence of any single individual does not significantly alter the output. However, in practice, raw noisy data is often unusable. For example, the U.S. Census Bureau must release data that satisfies "invariants"—hard constraints such as "the total population of a state must be exact" or "counts cannot be negative."

The traditional way to handle these constraints is through post-processing: first add noise, then use optimization to "nudge" the noisy results back into a feasible region that satisfies the constraints. While this makes the data look clean, researchers like Ferdinando Fioretto, Pascal Van Hentenryck, and Cuong Tran have shown that this "cleaning" step introduces a hidden "fairness tax."

The Geometry of Bias

The fundamental problem with post-processing is that it is not a neutral operation. In Post-processing of Differentially Private Data: A Fairness Perspective, Keyu Zhu and colleagues demonstrate that mathematical projections used to restore consistency systematically shift data points toward the center of the feasible region.

Imagine a small sub-population count that is naturally close to zero. If DP noise makes that count negative, a non-negativity constraint will push it back to zero. Because the noise is only "corrected" when it goes below zero but left alone when it goes above, the average value of the released data becomes higher than the true value. This systematic bias disproportionately affects minority groups and small geographic areas, leading to what Fioretto and Van Hentenryck describe in their survey as a distortion that can be more damaging than the original privacy-preserving noise itself.

Moving Noise into the Subspace

Instead of fixing the data after it has been broken by noise, a new framework called Subspace Differential Privacy proposes a different approach: design the noise so that it never violates the constraints in the first place.

In Subspace Differential Privacy, authors Jie Gao, Fang-Yi Yu, and Ruobin Gong argue that constraints define a "feasible subspace" within the high-dimensional space of all possible data. If a constraint requires the total population to remain constant, any noise added must sum to zero.

By injecting noise only into the "null space" of the constraints, the mechanism ensures that the mandated totals remain perfectly accurate while the individual components remain private. Because the noise is balanced within the subspace from the start, the resulting data is statistically unbiased, allowing downstream researchers to perform valid statistical analysis without worrying about the artifacts of post-hoc optimization.

Handling Discrete Counts

Real-world data like the Census consists of integers, not continuous numbers. This adds a layer of complexity: how do you add noise that is both restricted to a subspace and results in whole numbers?

Prathamesh Dharangutte, Jie Gao, and Ruobin Gong addressed this in Integer Subspace Differential Privacy. They utilize discrete lattice mathematics to ensure the noise respects both the linear invariants (like fixed totals) and the requirement that counts must be integers. This eliminates the need for rounding—another form of post-processing that typically introduces bias—and ensures that small sub-populations are not consistently undercounted or overcounted.

The Importance of Raw Noisy Data

The shift toward understanding these biases has led to a re-evaluation of how public agencies release data. As discussed by Jörg Drechsler and Weijie J Su in their foreword on Census Noisy Measurement Files, there is a growing demand for "Noisy Measurement Files" (NMFs).

NMFs are the raw, "messy" outputs of DP before any consistency constraints are applied. While they may contain negative numbers or inconsistent totals, they are statistically "pure." For high-precision scientific or demographic analysis, these raw outputs are often superior to "cleaned" data because their error distribution is transparent and predictable, enabling accurate uncertainty quantification that is otherwise impossible with biased, post-processed datasets.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_subspace_differential_privacy_for_unbias_dc1fb3a6,
  title = {Subspace Differential Privacy for Unbiased Data Release},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subspace-differential-privacy-for-unbiased-data-release/rdr_dc1fb3a69721487cbf177555b9514da0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_subspace_differential_privacy_for_unbias_dc1fb3a6_v1,
  title = {Subspace Differential Privacy for Unbiased Data Release},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subspace-differential-privacy-for-unbiased-data-release/rdr_dc1fb3a69721487cbf177555b9514da0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}