FFV-Net: A Federated Fraud-Aware Multi-Task Deep Learning Framework for Cross-Institution Payment Fraud Detection and Dynamic Startup Valuation under Uncertainty

Abstract

Early-stage startup valuation is highly sensitive to credit risk signals and transaction-level anomalies, yet financial institutions typically operate in data-siloed environments due to privacy and regulatory constraints. This separation limits both cross-institution payment fraud detection performance and the reliability of startup valuation under uncertainty. Moreover, existing studies rarely model the transmission mechanism through which fraud risk influences enterprise credit profiles and valuation dispersion. To address these challenges, we propose FFV-Net (Federated Fraud-aware Valuation Network), a unified federated multi-task deep learning framework for cross-institution payment fraud detection and dynamic startup valuation under uncertainty. The framework consists of three hierarchical modules: a federated Transformer–Graph Neural Network hybrid for privacy-preserving fraud detection across heterogeneous institutions, a risk representation learning layer that transforms fraud probabilities into structured credit risk embeddings, and a Bayesian quantile-based valuation network that produces calibrated valuation intervals. By explicitly modeling uncertainty propagation from fraud volatility to valuation dispersion, FFV-Net provides a risk-aware decision-support mechanism for venture capital investment, financial supervision, and regulatory sandbox applications. Experimental results on multi-institution financial datasets show that FFV-Net achieves an AUC of 0.963, improving fraud detection performance by 2.2% over the strongest multi-task baseline and by 6.2% over single-task models. Meanwhile, it reduces valuation MAE from 0.149 to 0.128 (a 14.1% decrease) while increasing interval coverage to 0.93 with narrower prediction widths, demonstrating robust uncertainty-aware valuation under federated privacy constraints.