Research map / Learning theory and algorithms / Fairness
Secure and Confidential Certificates of Online Fairness
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 OATH, a zero-knowledge protocol for confidential online group fairness certification that uses commit-online/prove-offline cut-and-choose auditing to scale to large query volumes.
Problem
Verifying fairness of a black-box deployed ML model confidentially is either unreliable (offline certificates on static data that miss distribution shift or model switching) or unscalable (ZKP per client query takes minutes).
Approach
Clients send hash commitments to their queries and outputs to the auditor during service; the provider then proves group demographic parity over all queries in ZK and opens only a group-balanced random sample for validity checks, giving a probabilistic bound on fairness manipulation with constant-size sampling and no client-facing ZKP.
Key results
- Service phase takes 0.22-0.25 s per query across five fairness datasets, with audit of 10^6 queries taking about 83-85 min for logistic regression.
- With nu=1000 sampled per group, any epsilon>=0.01 deviation evades detection with probability at most 6.65e-3; nu=3800 reduces this to 5.34e-9.
- On ResNet-101 (42.5M parameters), OATH achieves 4.4 s amortized runtime per query, with only 0.23 s requiring the client online, claimed to be a thousand-fold client-facing improvement over baselines in three of four categories.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OATH's probabilistic audit instead of verifying inference for every client query | Online group fairness certification · Neural network models | runtime | Helped |
| Use OATH's commit-online, prove-offline service phase rather than client-facing ZKPs | Online fairness certification · Client-facing service phase | runtime | Helped |
| Increase the audit sample parameter from \nu=1000 to \nu=3800 | OATH probabilistic audit · Fairness-gap deviation \epsilon \geq 0.01 | probability of evading detection: \num{6.65e-3} -> \num{5.34e-9} | Helped |
Limitations
- Assumes no collusion among provider, auditor, and clients and a random oracle model.
- Blame attribution requires clients to reveal data to the auditor; ZK alternative left to future work.
- Baseline runtimes for neural networks are estimated from prior ZKP inference timings rather than measured end-to-end.
Experimental setup
Datasets: COMPAS Crime Default Credit Adult German Credit
Benchmarks: COMPAS Crime Default Credit Adult German Credit
Metrics: runtime per query (s) audit runtime (min) detection probability of manipulation
Baselines: Black-Box audit Confidential-PROFITT FairProof Mystique-based ZKP fairness baseline
Builds on: Mystique FairProof Confidential-PROFITT EMP-toolkit IT-MAC ZKP
Closest papers by meaning
- FairProof : Confidential and Certifiable Fairness for Neural Networks (ICML 2024 · similarity 0.86)
- Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · similarity 0.81)
- An Information-Flow Perspective on Algorithmic Fairness (AAAI 2024 · similarity 0.81)
- Provable Robustness against Backdoor Attacks via the Primal-Dual Perspective on Differential Privacy (NeurIPS 2026 · similarity 0.81)
- Distributed Differentially Private Data Analytics via Secure Sketching (ICML 2025 · similarity 0.81)
- Confidential Guardian: Cryptographically Prohibiting the Abuse of Model Abstention (ICML 2025 · similarity 0.81)
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.