Probabilistic Dataset Reconstruction from Interpretable Models

TL;DR

This paper introduces a new framework for probabilistic dataset reconstruction from interpretable models, enabling privacy risk assessment by quantifying information leaks across various model types.

Contribution

It generalizes previous reconstruction methods to handle different interpretable models and efficiently computes the uncertainty of reconstructions under realistic assumptions.

Findings

Optimal interpretable models leak less information than greedy ones.

The framework applies to decision trees and rule lists.

Uncertainty of reconstruction can be computed efficiently.

Abstract

Interpretability is often pointed out as a key requirement for trustworthy machine learning. However, learning and releasing models that are inherently interpretable leaks information regarding the underlying training data. As such disclosure may directly conflict with privacy, a precise quantification of the privacy impact of such breach is a fundamental problem. For instance, previous work have shown that the structure of a decision tree can be leveraged to build a probabilistic reconstruction of its training dataset, with the uncertainty of the reconstruction being a relevant metric for the information leak. In this paper, we propose of a novel framework generalizing these probabilistic reconstructions in the sense that it can handle other forms of interpretable models and more generic types of knowledge. In addition, we demonstrate that under realistic assumptions regarding the interpretable models' structure, the uncertainty of the reconstruction can be computed efficiently. Finally, we illustrate the applicability of our approach on both decision trees and rule lists, by comparing the theoretical information leak associated to either exact or heuristic learning algorithms. Our results suggest that optimal interpretable models are often more compact and leak less information regarding their training data than greedily-built ones, for a given accuracy level.