SCC-VFL framework cuts unfair decisions 98% in distributed learning

When sensitive data lives across multiple institutions, ensuring that an individual’s outcome doesn’t arbitrarily change based on a protected attribute becomes a complex privacy‑fairness trade‑off. In a new paper accepted at ACM FAccT 2026, researchers from Virginia Tech and the U.S. Army Research Laboratory introduce SCC‑VFL (Selective Counterfactual Consistency for Vertical Federated Learning). The framework tackles individual fairness—i.e., per‑instance prediction consistency under counterfactual interventions on protected attributes—rather than group parity metrics. SCC‑VFL operates in a server‑centric architecture: it first uses a differentially private, graph‑free method to discover feature roles (non‑descendants, policy‑permitted mediators, and impermissible proxies) from only a private sketch of the sensitive attribute. Then it generates masked counterfactuals that edit only mediators while fixing non‑descendants and suppressing proxy leakage. Finally, a server‑side consistency loss penalizes prediction changes that would result from altering the protected attribute via impermissible pathways.

The empirical results are striking. Across three real‑world datasets in credit, healthcare, and criminal justice, SCC‑VFL maintains or improves predictive accuracy while reducing decision flip rates by up to 98% compared to strong baselines like adversarial debiasing or fairness‑aware federated learning. It also lowers attribute‑inference attack success rates and improves robustness to distribution shifts. The method achieves favorable utility‑fairness‑privacy trade‑offs without requiring centralized data or exposing sensitive attributes across parties. For organizations deploying vertical federated learning in regulated domains—such as banks sharing transaction data, hospitals pooling patient records, or criminal justice agencies coordinating risk assessments—SCC‑VFL offers a practical path to individual fairness without sacrificing privacy or accuracy.