Federated learning (FL) enables collaborative model training without sharing any raw data. However, real-world deployments in domains such as finance are limited by factors such as statistical heterogeneity, adversarial manipulation, privacy leakage, and low interpretability. Existing work address these challenges but are largely studied in isolation, with assessment usually limited to predictive performance. This paper proposes TAP-FL (Trust-Aware Privacy-Preserving Federated Learning), a unified federated learning framework designed to perform an extensive assessment in realistic conditions. TAP-FL uses non-IID data distributions with multiple adversarial attack settings, local differential privacy and a trust-aware aggregation mechanism which works based on update’s behavioral consistency and deviation. A multi-dimensional evaluation setup is added to collectively measure predictive performance, attack success, privacy leakage, and as well as SHAP-based analysis to measure explanation stability and attribution drift. The framework is further applied to cross-domain credit risk assessment task, combining heterogeneous models via a meta-learning formulation. Experimental results demonstrate that training stability is mostly influenced by adversarial updates, and privacy mechanisms create regulated uncertainty with minimal effect on inference risk. TAP-FL achieves consistent resilience against different types of attacks and has a consistent balance between utility and privacy, while maintaining interpretability.