ProF Framework Guarantees Fairness in DNNs with 95% Generalization
New technique repairs bias using interval bound propagation and MILP solvers
Deep neural networks frequently exhibit discriminatory biases, but existing fairness repair methods lack provable guarantees and often fail on unseen data. To address this, researchers from Zhejiang University have developed ProF (Provable Fairness Repair), a framework that leverages interval bound propagation—a verification technique—to soundly capture model outputs over neighborhoods around biased samples. These bounds guide a unified constraint-solving formulation translated into a Mixed-Integer Linear Programming (MILP) problem, solvable by standard solvers. The result is a repaired model with guaranteed fairness on the entire set around each biased sample.
ProF was evaluated on four benchmark datasets and demonstrated remarkable performance: 95.93% fairness generalization on full datasets, 93.16% on the entire input space, and roughly 90% improvement in fairness metrics. The framework is flexible, supporting multiple sensitive attributes and practical fairness definitions. Accepted at ASE 2025 and available open-source, ProF represents a significant step toward trustworthy AI deployment where biased behavior can be provably eliminated.
- ProF uses interval bound propagation to soundly capture model outputs over entire neighborhoods around biased samples
- Formulated as a Mixed-Integer Linear Programming (MILP) problem solvable by off-the-shelf solvers
- Achieves up to 95.93% generalization on full datasets and 93.16% on entire input space, with ~90% fairness improvement
Why It Matters
Provable fairness guarantees remove data-centric limitations, enabling trustworthy AI in high-stakes domains like hiring and lending.