Abstract Employment inclusion for people with disabilities remains a persistent challenge across developed nations, with employment rates 24 percentage points below general population levels in Europe. Traditional centralized employment matching systems face privacy constraints, data sovereignty requirements, and limited collaborative learning capabilities. This paper presents a comprehensive privacy-preserving federated learning framework for multi-regional disability employment matching that enables collaborative model training while maintaining data locality and regulatory compliance. Our approach combines ensemble-based LightGBM federation with parameter-level MLP federation, achieving 0.9011 F1-score (LightGBM) and 0.7881 F1-score (MLP with privacy preservation) across five Italian employment centers in Veneto region. The privacy framework integrates differential privacy with RDP composition (ε = 1.0, δ = 10⁻⁶), Shamir’s 3-of-5 secret sharing for secure aggregation, and blockchain anchoring for long-term integrity verification. Empirical evaluation demonstrates minimal performance degradation under federated constraints (0.0005 F1-score loss) while providing formal privacy guarantees. The system reduces manual processing time from 30–60 min to under 5 min per candidate with sub-100 ms response times suitable for real-world deployment. Ongoing pilot deployment at partner employment centers validates practical applicability for European disability employment services. The research contributes to trustworthy AI frameworks for sensitive social applications requiring regulatory compliance, algorithmic fairness, and privacy preservation.
Kuznetsov et al. (Wed,) studied this question.