Research map / Learning theory and algorithms / Prediction and decision losses
I Prefer Not to Say: Protecting User Consent in Models with Optional Personal Data
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.
Introduces Protected User Consent (PUC) and an model-agnostic PUC-inducing data augmentation (OFFIDA) so that models with optional personal data do not infer information from the decision to withhold data, while remaining loss-optimal and not worse than the base model.
Problem
Machine learning models with optional features can implicitly infer information from the unavailability of optional data, violating non-sharers' consent and disadvantaging users who protect their data (e.g., insurance pricing with optional fitness data).
Approach
Defines Protected User Consent (PUC) as the conditional expectation given base features and only the consented (available) optional features, marginalizing over whether other optional features are shared; achieves it via data augmentation that randomly drops available optional features and reweights instances by 2^|available|, so missingness carries no label information.
Key results
- Without PUC, full-feature random forests give non-sharers up to 18% lower positive-outcome rates than base models (e.g., income change -18.8%), while OFFIDA reduces this gap to about 1% or less on classification tasks.
- Exhaustive OFFIDA improves over the base model on all six datasets in the r-dimensional strategic setting (e.g., insurance MSE 271.72 base vs 242.99 fixed-size and 260.77 exhaustive) while staying moderately worse than full-feature models.
- PUC-gaps converge to zero at roughly O(1/N) on synthetic data for models such as Random Forest and MLP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply OFFIDA instead of the base model | PUC-compliant model · Diabetes, strategic withholding | misclassification error: 29.30% -> 26.61% | Helped |
| Apply OFFIDA instead of the base model | PUC-compliant model · COMPAS, strategic withholding | misclassification error: 42.89% -> 40.85% | Helped |
| Apply OFFIDA instead of the base model | PUC-compliant model · Adult, strategic withholding | misclassification error: 16.05% -> 15.94% | Helped |
| Apply OFFIDA instead of the base model | PUC-compliant model · Income, strategic withholding | MSE × 100: 85.07 -> 80.22 | Helped |
| Apply OFFIDA instead of the base model | PUC-compliant model · California, strategic withholding | MSE × 100: 15.62 -> 14.79 | Helped |
| Apply OFFIDA instead of the base model | PUC-compliant model · Insurance, strategic withholding | MSE × 100: 262.43 -> 254.35 | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · Diabetes | misclassification error: 34.01% -> 31.44% | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · COMPAS | misclassification error: 44.54% -> 41.47% | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · Adult | misclassification error: 13.41% -> 12.84% | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · Water | misclassification error: 10.97% -> 10.00% | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · Colic | misclassification error: 15.05% -> 11.34% | Helped |
| Use the full-feature model rather than OFFIDA | Random Forest · Income | MSE × 100: 110.73 -> 109.11 | Helped |
Surprising findings
- Full-feature models violate consent by exploiting missingness: non-sharers can be disadvantaged by up to 18% even though they provided no extra data.
- Under PUC, the decision maker can benefit from optional data (predictive non-degradation) without penalizing non-sharers.
Limitations
- Assumes users act either purely strategically or purely on privacy grounds; heterogeneous user preferences not modeled.
- Exhaustive augmentation grows dataset size by 2^r factor, and the finite-sample bound scales with 2^r and feature space size.
- Experiments use one or a few optional features with synthetic availability mechanisms for most datasets.
Experimental setup
Datasets: Diabetes COMPAS UCI Adult water treatment horse colic California housing income (ACS/folktables) insurance claims synthetic binary dataset synthetic continuous dataset
Benchmarks: Diabetes COMPAS UCI Adult water treatment horse colic California housing income insurance
Metrics: misclassification error MSE PUC-gap squared average positive prediction rate for non-sharers
Baselines: base model (base features only) full-feature model with zero imputation conditional statistical parity model fixed-size OFFIDA exhaustive OFFIDA
Builds on: Rastegarpanah et al. data minimization with optional features conditional statistical parity random forest
Closest papers by meaning
- Setting $\varepsilon$ is not the Issue in Differential Privacy (NeurIPS 2025 · similarity 0.84)
- MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective (ICML 2024 · similarity 0.84)
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.84)
- Improved Bounds for Pure Private Agnostic Learning: Item-Level and User-Level Privacy (ICML 2024 · similarity 0.84)
- PASS: Private Attributes Protection with Stochastic Data Substitution (ICML 2025 · similarity 0.84)
- It's My Data Too: Private ML for Datasets with Multi-User Training Examples (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.