Abstract
We examine machine learning models in a setup where individuals have the choice to share optional personal information with a decision-making system, as seen in modern insurance pricing models. Some users consent to their data being used whereas others object and keep their data undisclosed. In this work, we show that the decision not to share data can be considered as information in itself that should be protected to respect users' privacy. This observation raises the overlooked problem of how to ensure that users who protect their personal data do not suffer any disadvantages as a result. To address this problem, we formalize protection requirements for models which only use the information for which active user consent was obtained. This excludes implicit information contained in the decision to share data or not. We offer the first solution to this problem by proposing the notion of Protected User Consent (PUC), which we prove to be loss-optimal under our protection requirement. To learn PUC-compliant models, we devise a model-agnostic data augmentation strategy with finite sample convergence guarantees. Finally, we analyze the implications of PUC on a variety of challenging real-world datasets, tasks, and models.
Introduction: The Hidden Cost of Privacy
In many modern systems—from health insurance pricing to credit scoring—users are asked for "base features" (e.g., age, residence) and offered the choice to provide "optional features" (e.g., fitness app data, extra financial history). While this choice is presented as voluntary, the decision to share is rarely random.
If only the healthiest individuals share their fitness data to secure discounts, a standard machine learning model quickly learns that "missing data" is a proxy for high risk. Consequently, users who "prefer not to say" for privacy reasons are implicitly lumped in with high-risk groups and penalized. This creates a coercive environment where the choice to remain private results in a higher price or a rejected application. This paper formalizes a way to respect user consent so that non-sharers are treated based only on their provided data, not on the fact that they withheld the rest.
Defining Protected User Consent (PUC)
The authors propose a framework called Protected User Consent (PUC), built on a core requirement termed Availability Inference Restriction (AIR).
- AIR (The Rule): A model must not use the unavailability of a feature to make inferences. For a non-sharer, the model should behave as if it only ever had access to the base features.
- PUC (The Goal): A model that satisfies AIR while remaining "loss-optimal"—meaning it provides the best possible predictions for both sharers (using all data) and non-sharers (using only base data) without cross-contamination.
PUCIDA: A Model-Agnostic Solution
To achieve PUC, the researchers developed PUC-inducing Data Augmentation (PUCIDA). Instead of training on raw data, they create synthetic samples. For every instance where a user shared optional data, they create a "shadow" version where that data is missing. By carefully weighting these samples, they force the model to see the label distribution as independent of the sharing decision. This approach is model-agnostic, working effectively with Decision Trees, Neural Networks, and Gradient Boosted Trees alike.
Key Results
The researchers evaluated their approach using a mix of synthetic and real-world datasets, including medical, recidivism, and insurance data.
- The Penalty is Real: Without PUC, non-sharers faced significant disadvantages. In some datasets, full-feature models assigned outcomes up to 18% worse for non-sharers compared to a fair base model.
- PUCIDA Effectiveness: Applying the augmentation technique successfully vanished or significantly reduced this "privacy gap" across all tested datasets and classifiers.
- Theoretical Soundness: They proved that the "PUC-Gap" (the distance between a trained model and the ideal fair model) shrinks at a rate of $O(1/N)$, ensuring reliability as sample sizes grow.
- Sharers Still Benefit: Consenting users still received more accurate, personalized predictions, proving that systems can incentivize data sharing without punishing those who opt out. This confirms that privacy and performance are not a zero-sum game; a PUC-compliant model is guaranteed to be at least as accurate as a model that never asked for optional data.
Conclusion
By formalizing the "silence as a signal" problem, this work provides a path toward truly voluntary data sharing. PUC ensures that the decision to protect one's personal data does not lead to systematic disadvantages, maintaining the integrity of user consent in automated decision-making systems.
Summary updated