Financial fraud remains a critical challenge for digital banking, requiring detection solutions that ensure both scalability and data privacy. Traditional centralized approaches face limitations due to security risks and system bottlenecks. This paper proposes FedFraud, a novel federated learning framework that detects fraudulent transactions without sharing raw data. FedFraud introduces two key innovations: (i) a trust-aware client aggregation mechanism that assigns weights based on update reliability, and (ii) an asynchronous communication protocol enabling clients to contribute updates independently. Evaluated on the Credit Card Fraud Detection dataset under a non-IID setup, FedFraud achieves an F1-score of 0.90 and AUC of 0.96, outperforming FedAvg and state-of-the-art methods like FedGAT and FL-BGAT. It also reduces privacy leakage to 1.5% and limits gradient reconstruction success to 15%, compared to 35% for FedAvg. Scalability tests show stable convergence with up to 100 clients, establishing FedFraud as an effective and privacy-preserving solution for decentralized financial fraud detection.