Privacy-Utility Tradeoffs in Federated Financial Analytics: An Optimization Framework

Cross-institutional financial analytics face fundamental challenges balancing privacy protection, model utility, and computational efficiency. This paper presents a comprehensive optimization framework addressing privacy-utility tradeoffs in federated learning for financial services. We propose adaptive privacy budget allocation mechanisms combined with a hybrid Trusted Execution Environment and Secure Multi-Party Computation protocols. Our framework targets KYC/AML workflows where regulatory compliance demands stringent data protection without sacrificing analytical AUC‑ROC. Experimental evaluation demonstrates superior performance across multiple financial datasets, achieving AUC-ROC = 0.867 at ε=2.0, while reducing per-round bandwidth costs by ~94% via gradient compression; TEE-assisted aggregation reduces compute/round-trip overhead rather than bandwidth. (achieving 3.21× speedup over a pure MPC-based secure aggregation baseline and reducing round time from 847s to 264s). The proposed approach ensures algorithmic fairness through demographic parity constraints and provides quantifiable privacy risk metrics aligned with commonly used industry thresholds and internal policy targets. Metric Convention: Unless otherwise specified, all performance metrics reported in this paper are AUC‑ROC; any occurrences labeled as 'AUC‑ROC' in results refer to AUC‑ROC for binary classification.