Introduction
As machine learning (ML) models are increasingly deployed in high-stakes domains such as credit scoring and hiring, ensuring algorithmic fairness has become a critical requirement. However, two major challenges impede the adoption of fairness audits: the need to protect proprietary model weights (confidentiality) and the need to guarantee that the model being audited is the same one used for real-world decisions (trust).
FairProof addresses these challenges by providing a system for generating mathematically certain "certificates of fairness" without revealing confidential model information. By integrating geometric fairness certification with Zero-Knowledge Proofs (ZKPs), FairProof allows a model owner to prove to a user that a specific decision was made fairly, ensuring model uniformity and privacy.
Methodology
The core of FairProof lies in its approach to Local Individual Fairness (IF). Rather than relying on global statistics, the system proves that for a specific query, small perturbations to sensitive attributes (e.g., race, gender) do not alter the model's output.
Fairness as Geometry
FairProof treats ReLU-based neural networks as a collection of convex polytopes. In this geometric view, the decision space is partitioned into regions where the model's behavior is linear. Fairness certification then becomes a search for the nearest "decision boundary"—the distance an input can move before its classification flips.
ZKP-Friendly Design
To make geometric certification compatible with the computational constraints of Zero-Knowledge Proofs, FairProof employs several optimizations:
- Lower Bound Approach: Instead of calculating exact distances, the system computes a "projection distance." This simpler arithmetic is significantly faster to verify within a ZKP while still providing a valid lower bound on fairness.
- Strategic Sub-functionalities: The proof generation is decomposed into five verifiable steps, such as verifying the initial polytope and checking facet distances. This "divide and conquer" strategy manages computational overhead.
- Representative Points: Model owners can pre-compute points within geometric regions offline, accelerating the online proof generation for user queries.
Evaluation and Results
The system was evaluated using standard fairness benchmarks, including the Adult, Default Credit, and German Credit datasets.
Key Findings
- Practical Performance: For the German Credit dataset, FairProof generates a verifiable certificate in approximately 1.17 minutes on a standard CPU.
- Communication Efficiency: The resulting certificates are succinct, averaging only 43.5 KB, which facilitates easy transmission over web protocols.
- Discrimination Detection: The system effectively distinguishes between fair models (trained with regularization) and unfair, overfitted models by reflecting differences in the certified fairness parameter ($\epsilon$).
- Optimization Gains: The use of offline pre-computation resulted in up to a 1.74x speedup in the online phase.
Discussion and Limitations
While FairProof demonstrates the feasibility of confidential fairness certification, several areas for future work remain. The geometric approach currently faces scalability challenges; the number of facets to check grows exponentially with network size, limiting its application to the small-to-medium networks typical of tabular data. Furthermore, the current implementation is restricted to fully-connected ReLU networks. Future research may explore verifying only the final decision-making layers of larger models or extending the protocol to architectures like Transformers and CNNs.
By forcing model owners to cryptographically "commit" to their model weights before deployment, FairProof eliminates the possibility of "model swapping," establishing a new standard for trust in automated decision-making.
Summary updated