Research map / Learning theory and algorithms / Fairness
Fairness Shields: Safeguarding against Biased Decision Makers
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.
Introduces fairness shields, a symbolic run-time monitor that overrides a black-box classifier's decisions to guarantee bounded-horizon or periodic group fairness at minimal intervention cost.
Problem
Deployed AI decision-makers may be biased over short decision sequences even when long-run probabilistic fairness holds; regulators assess fairness over bounded horizons or periodically, and existing bias mitigation is design-time only.
Approach
A symbolic shield monitors a deployed black-box classifier's recommendations and intervenes at minimal cost so that empirical demographic parity or equal opportunity holds at a fixed horizon (FinHzn, via counter-based dynamic programming) or periodically (Static-Fair, Static-BW, Dynamic shields).
Key results
- Unshielded models violated bounded-horizon fairness in 44% of cases for DP and 65% for EqOpp, while FinHzn shielded models were always fair at the horizon.
- Dynamic shields achieved 100% fairness satisfaction in periodic settings for DP and EqOpp, versus 95.71% for Static-Fair and 83.1% for Static-BW (DP).
- Shield synthesis took about 1 second and 30 MB for DP and 1.5 seconds and 1.3 GB for EqOpp.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a FinHzn fairness shield to the unshielded classifier | ML classifiers trained with DiffDP, HSIC, LAFTR, PRemover, or ERM · Bounded-horizon DP | fairness violations: 44% | Helped |
| Add a FinHzn fairness shield to the unshielded classifier | ML classifiers trained with DiffDP, HSIC, LAFTR, PRemover, or ERM · Bounded-horizon EqOpp | fairness violations: 65% | Helped |
| Use Dynamic instead of Static-Fair shielding | ERM classifier · DP (all datasets and runs) | fairness satisfaction: 95.71% -> 100% | Helped |
| Use Dynamic instead of Static-BW shielding | ERM classifier · DP (all datasets and runs) | fairness satisfaction: 83.1% -> 100% | Helped |
| Use Dynamic instead of Static-BW shielding | ERM classifier · EqOpp (all datasets and runs) | fairness satisfaction: 56.4% -> 100 % | Helped |
| Train the classifier with a fairness mitigation method rather than ERM | FinHzn shielding · DP and EqOpp | utility loss | Helped |
| Decrease the bias threshold | FinHzn shielding · DP and EqOpp | utility loss | Hurt |
| Recompute the shield at the beginning of each period instead of reusing a static shield | Periodic fairness shielding · DP and EqOpp | fairness satisfaction | Helped |
| Use Static-BW instead of Dynamic shielding | Periodic shielding when assumptions are satisfied · DP and EqOpp | utility loss | Hurt |
Surprising findings
- Static-Fair shields' periodic-fairness assumption was never met in experiments, yet they still satisfied fairness in most runs (95.71% for DP).
Limitations
- Periodic guarantees hold only under balance/assumption conditions that may be violated; Static-Fair guarantees fail in general.
- Assumes the input distribution over decisions and intervention costs is known or learned, and uses only constant costs in experiments.
- Shield may become inactive or infeasible-prone for skewed group appearance orders; applies only to difference-of-ratios properties (not disparate impact).
Experimental setup
Datasets: Adult COMPAS German Credit Bank Marketing
Benchmarks: Adult COMPAS German Credit Bank Marketing
Metrics: demographic parity violation equal opportunity violation classification accuracy utility loss intervention cost
Baselines: ERM DiffDP HSIC LAFTR PRemover Static-Fair Static-BW
Builds on: demographic parity equal opportunity LAFTR DiffDP HSIC
Closest papers by meaning
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.87)
- Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.87)
- Intersectional Fairness via Mixed-Integer Optimization (NeurIPS 2026 · similarity 0.86)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.86)
- Identifying Reasons for Bias: An Argumentation-Based Approach (AAAI 2024 · similarity 0.86)
- Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.86)
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.