A differentially private personalized federated learning model for cross-enterprise labor risk assessment

Cross-institutional labor-risk assessment involves sensitive records, class imbalance, and substantial distributional heterogeneity. This study presents DP-PFLR, a personalized federated-learning framework integrating Adaptive Privacy Budget Allocation (APBA), Risk-Aware Re-weighting (RAR), and Personalized Heterogeneity-aware Aggregation (PHA). APBA combines private adaptive clipping with an RDP-calibrated noise schedule; RAR combines focal loss with inverse-prevalence client weighting; and PHA retains personalized heads locally while aggregating the shared encoder. The framework provides record-level differential privacy for the server-visible training transcript, whereas local heads are not themselves differentially private. Experiments on Adult, IBM-HR, and OSHA evaluate income status, employee attrition, and establishment-level occupational risk under proxy cross-silo partitions. At the target budget (ε ,δ )=(4,10⁻⁵) , the maximum accounted privacy loss was 3.99. DP-PFLR achieved macro-F1 scores of 0.792, 0.695, and 0.650, respectively, exceeding the strongest evaluated private baseline on all three datasets. Paired tests, complete 2³ ablation, sensitivity analysis, and restrictive-privacy and robustness experiments further supported the framework’s privacy–utility performance. Nevertheless, these public benchmarks do not establish production effectiveness, complete Byzantine robustness, causal interpretation, or readiness for automated employment decisions. External validation, fairness auditing, secure governance, and human oversight remain necessary.