Research map / Federated learning and privacy / Differential privacy
Finding ε and δ of Traditional Disclosure Control Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Derives (ε, δ)-DP bounds for randomized versions of traditional statistical disclosure control methods (cell suppression and swapping) and compares their privacy, error, and fairness against Laplace and discrete Gaussian mechanisms on US ACS data.
Problem
Traditional SDC systems (cell suppression, swapping) lack formal privacy guarantees, and it is unclear how their privacy, accuracy, and fairness compare with differential privacy.
Approach
Introduces minimal randomization into cell suppression (noisy threshold via Laplace noise) and swapping (permute-and-flip donor selection) so their outputs closely resemble the originals while allowing derivation of (ε, δ)-DP bounds, plus bias, variance, and fairness analyses.
Key results
- DP cell suppression and DP swapping have very large delta values (e.g., delta 0.99 at epsilon 0.5-4), showing SDC methods give weak formal privacy guarantees.
- At epsilon=0.5 on MA data, Laplace yields L1 error 763.77 and fairness 3.65 versus DP-Swap error 3462.16 and fairness 1774.72.
- Classical Laplace and discrete Gaussian mechanisms achieve higher classification accuracy than SDC-based mechanisms in most settings, close to the non-private baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the Laplace mechanism instead of DP cell suppression | Histogram data release · Massachusetts ACS | error ∥B(M)∥1: 154.78 vs 510.18 | Helped |
| Use the Laplace mechanism instead of DP swapping | Histogram data release · Massachusetts ACS | error ∥B(M)∥1: 154.78 vs 3653.04 | Helped |
| Use the Laplace mechanism instead of the discrete Gaussian mechanism | Histogram data release · Massachusetts ACS | error ∥B(M)∥1: 154.78 vs 436.92 | Helped |
| Use the Laplace mechanism instead of DP swapping | Histogram data release · Massachusetts ACS | fairness ∥B(M)∥⇌: 0.90 vs 1946.84 | Helped |
| Use the Laplace mechanism instead of DP cell suppression | Histogram data release · Massachusetts ACS | variance ∥V(M)∥∞: 2.06 vs 2.22 | Hurt |
| Use Laplace or discrete Gaussian mechanisms rather than SDC mechanisms | Data release · Massachusetts ACS | privacy δ: 0 vs 0.99∗ | Helped |
| Use Laplace or discrete Gaussian mechanisms rather than SDC mechanisms | Logistic regression trained on private data · Massachusetts ACS classification | ROC AUC score | Helped |
| Use DP swapping instead of Laplace for subgroup marginal release | Two-way marginal query · National ACS, Black men | histogram error | Hurt |
| Use Laplace instead of discrete Gaussian for subgroup marginal release | Two-way marginal query · National ACS, White men and Black men | error disparity between groups | Helped |
| Increase ϵ | DP data release methods · Massachusetts and Texas ACS | Unique Exact Match score | Hurt |
Surprising findings
- Traditional SDC mechanisms can induce much higher fairness violations than classical DP mechanisms.
- Cell suppression attains near-perfect scores on the unique exact match privacy metric because suppressed rows are left untouched, an artifact of the metric.
Limitations
- The derived delta values are large, so the DP SDC variants offer weak formal guarantees and are not strong DP analogs of the originals.
- Evaluation is limited to histogram release and logistic regression on ACS data, with one empirical privacy metric.
- Analysis covers only one version of each SDC method, and composition with post-processing is left to future work.
Experimental setup
Datasets: 2019 ACS Diverse Community Data Excerpts (Massachusetts, Texas, National PUMAs)
Benchmarks: 2019 ACS Diverse Community Data Excerpts
Metrics: epsilon-delta privacy bound l1 histogram error alpha-fairness violation of bias variance ROC AUC for logistic regression classification unique exact match (empirical privacy)
Baselines: Laplace mechanism discrete Gaussian mechanism traditional cell suppression traditional swapping
Builds on: differential privacy permute-and-flip Laplace mechanism discrete Gaussian mechanism post-processing fairness
Closest papers by meaning
- Comparing Population Means Under Local Differential Privacy: With Significance and Power (AAAI 2018 · similarity 0.85)
- A Simple and Practical Method for Reducing the Disparate Impact of Differential Privacy (AAAI 2024 · similarity 0.85)
- Setting $\varepsilon$ is not the Issue in Differential Privacy (NeurIPS 2025 · similarity 0.84)
- Mean Estimation in the Add-Remove Model of Differential Privacy (ICML 2024 · similarity 0.83)
- Prior-itizing Privacy: A Bayesian Approach to Setting the Privacy Budget in Differential Privacy (NeurIPS 2024 · similarity 0.83)
- Auditing $f$-differential privacy in one run (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.