AudiFair: Privacy-Preserving Framework for Auditing Fairness

ICLR 2026 Conference Submission16000 Authors

Keywords: fairness, zero-knowledge proofs, decision trees

Abstract:

Ensuring fairness in AI is challenging, especially when privacy concerns prevent access to proprietary models and training data.
We propose a cryptographic framework for auditing fairness without requiring model disclosure.

Unlike existing solutions---which either do not capture attack vectors enabling dishonest model providers to manipulate a dataset to pass audits unfairly, or require involving real-world model users to protect against dishonest behaviors---our framework realizes the following properties simultaneously for the first time:
1) Model Privacy: Proprietary model details remain hidden from verifiers.
2) Dishonest Provider Robustness: Even if model providers are dishonest, a verifier can statistically attest to the fairness of the model without involving real-world users.
3) Test Data Transparency: Test data for auditing is generated in a transparent and accountable way, preventing dishonest parties from manipulating it.
We achieve these goals by carefully orchestrating cryptographic commitments, coin tossing, and zero knowledge proofs, and we report the empirical performance for auditing private decision tree models.
Our solution is highly communication-efficient, delivering a significant improvement (~200,000x for a 30k-sized dataset) over the current state-of-the-art methods.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 16000