Privacy-Preserving Cross-Carrier Auto Physical Damage Fraud Detection Using Non-IID Federated Learning, Dropout-Resilient Secure Aggregation, and Cross-Carrier Generalization

Main Article Content

Sai Charan Obuladinne
Manish Kumar

Abstract

Cross-carrier fraud detection can benefit from broader evidence than any single insurer observes, yet pooling claim-level records creates privacy and governance barriers. This study evaluates Heterogeneity-Aware Secure Federated Auto Physical Damage Fraud Detection (HA-SFAPD), a cross-carrier framework that combines local claim training, imbalance-aware optimization, heterogeneity-aware aggregation, dropout-resilient secure aggregation, and leave-one-carrier-out external validation. The completed evaluation comprised 60,000 de-identified Auto Physical Damage claims from five independent carriers, including 6,000 confirmed or investigated fraudulent claims and 54,000 non-fraudulent claims. Natural carrier partitions preserved unequal claim volumes, fraud prevalence, feature divergence, label divergence, and class imbalance. On the 6,000-claim primary holdout set, secure HA-SFAPD achieved PR-AUC 0.790, fraud recall 0.840, F1 0.764, MCC 0.739, ROC-AUC 0.940, and specificity 0.960. PR-AUC improved by 0.150 over carrierlocal learning and by 0.090 over FedAvg. Secure aggregation changed PR-AUC by only -0.002 relative to the unsecured configuration while adding 0.83 s mean round latency. Leave-one-carrier-out evaluation produced mean PR-AUC 0.727 versus 0.611 for single-carrier transfer. At 40% carrier dropout, resilient aggregation completed 96% of rounds versus 52% without recovery, while PR-AUC remained 0.770. The findings show that the evaluated framework preserved claim locality, improved fraud-sensitive performance under non-IID carrier distributions, transferred to unseen carriers, and quantified the operational cost of secure collaboration.

Article Details

Section
Articles