Research map / Learning theory and algorithms / Fairness
FairProof : Confidential and Certifiable Fairness for Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FairProof, a system using zero-knowledge proofs over committed neural network weights to publicly verify a personalized local individual fairness certificate for small ReLU networks, implemented in Gnark.
Problem
Confidential ML models cannot be publicly audited for fairness; consumers need verifiable, personalized fairness guarantees without revealing model weights and while ensuring the same model is used for all users.
Approach
Reduce local individual fairness certification to pointwise l2 robustness computed over per-sensitive-value slices of the ReLU polytope complex, use a GeoCert-style iterative search with a projection-based lower bound that is ZKP-friendly, and prove each step of the search with zk-SNARKs over committed weights.
Key results
- On German (2,4) the online proof generation averages 1.17 minutes per data point on an Intel i9 CPU without parallelism, with a certificate proof size of about 43.5KB.
- Offline precomputation reduces online proof time by up to 1.74x (German (4,2): 4.90 min online vs 3.61 min offline).
- Verification takes seconds versus minutes for proof generation across all models; unfair models (more overfit) show lower certified epsilon than fair ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Perform selected computations in an offline phase (ablation) | FairProof · Fairness certificate proof generation | online proof generation time: largest being 1.74$\times$ | Helped |
| Increase weight decay regularization | fully connected ReLU networks · Adult, Default Credit, and German Credit | fairness parameter ε | Helped |
Limitations
- Only small fully-connected ReLU networks with two hidden layers are evaluated; scalability to large models is unclear, with traversed facets possibly huge.
- The certificate is a lower bound via projection distance rather than exact local individual fairness.
- Model commitment/uniformity is assumed to be handled by external commitment schemes and is out of scope.
Experimental setup
Datasets: Adult Default Credit German Credit
Benchmarks: Adult Default Credit German Credit
Metrics: proof generation time (min) verification time (s) proof size (KB) local individual fairness parameter epsilon number of traversed facets
Builds on: GeoCert zk-SNARK (Gnark) certified robustness via polytope partitions
Closest papers by meaning
- Certified but Private: Scalable Zero-Knowledge Proofs for the Formal Verification of Neural Networks (NeurIPS 2026 · similarity 0.88)
- Secure and Confidential Certificates of Online Fairness (NeurIPS 2025 · similarity 0.86)
- Can Private Machine Learning Be Fair? (AAAI 2025 · similarity 0.85)
- Disparate Impact on Group Accuracy of Linearization for Private Inference (ICML 2024 · similarity 0.84)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.84)
- PFGuard: A Generative Framework with Privacy and Fairness Safeguards (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.