RAFA-BioAuth: Risk-Adaptive and Fairness-Aware Biometric Authentication for Secure Mobile Banking

This research examines RAFA-BioAuth, a risk-adaptive, fairness-aware framework for mobile banking incases of presentations and facial occlusions. The proposed solution combines aspects of: Identity Similarity, PassivePresentation Attack Detection (PAD), Asymmetric Financial-Risk Estimation and Fairness Regularization at theIdentity Level. For evaluation purposes, all benchmark datasets were split into subject disjoint training, validation, andtesting sets. The threshold values from the validation set were used. Bootstrap resampling was employed to estimate thevariance. Monte Carlo simulations were performed to estimate the risk. The results showed that identity verification(AUC = 0.548) and PAD (AUC ≈ 0.55) were poor. Comparing RAFA to the AND rule resulted in FAR = 0.20, FRR =0.31, and expected risk = 340. On the other hand, the AND rule had lower FAR of 0.09, but increased FRR to 0.78.Finally, a conservative end-to-end approach produced an FRR of 0.686. Thus, our results indicate trade-offs rather thanproduction-readiness as we did not perform deployment, cross-device, or cross-dataset validations on our solutions. Wepresent the contributions of this research as being an interpretable integration and not a new algorithm.