Observational Auditing of Label Privacy

TL;DR

This paper presents a new observational framework for auditing label privacy in machine learning, eliminating the need for dataset modifications and enabling scalable privacy evaluation in large systems.

Contribution

It introduces a novel observational auditing method that leverages data randomness, extending privacy auditing to labels without altering the dataset, supported by theoretical and experimental validation.

Findings

Effective in auditing label privacy on Criteo and CIFAR-10 datasets.

Eliminates the need for dataset modifications in privacy auditing.

Extends privacy auditing beyond membership inference to protected attributes.

Abstract

Differential privacy (DP) auditing is essential for evaluating privacy guarantees in machine learning systems. Existing auditing methods, however, pose a significant challenge for large-scale systems since they require modifying the training dataset -- for instance, by injecting out-of-distribution canaries or removing samples from training. Such interventions on the training data pipeline are resource-intensive and involve considerable engineering overhead. We introduce a novel observational auditing framework that leverages the inherent randomness of data distributions, enabling privacy evaluation without altering the original dataset. Our approach extends privacy auditing beyond traditional membership inference to protected attributes, with labels as a special case, addressing a key gap in existing techniques. We provide theoretical foundations for our method and perform experiments on Criteo and CIFAR-10 datasets that demonstrate its effectiveness in auditing label privacy guarantees. This work opens new avenues for practical privacy auditing in large-scale production environments.