Generative AI–Driven Trusted Credit Scoring with Threat-Aware Analytics for Project and Healthcare Data

Abstract

Credit scoring is a foundational component of modern financial systems: it determines who can borrow, at what cost, and under what conditions. Traditional credit scoring models rely on historical financial data and statistical models such as logistic regression or decision trees. However, as financial institutions seek greater accuracy, scalability, and robustness, generative artificial intelligence (AI) methods—especially Generative Adversarial Networks (GANs) and large language models (LLMs)—are emerging as promising tools. Yet, the adoption of generative AI in credit scoring raises critical concerns around trust: adversarial attacks, lack of interpretability, regulatory compliance, and computational efficiency.

 In this work, we propose a trusted generative AI framework for credit scoring that is threat-aware, explainable, and Apache-accelerated. First, we design a adversarially robust credit scoring model: we use GAN-based data augmentation to mitigate class imbalance in default versus non-default classes, while simultaneously analyzing potential adversarial vulnerabilities by generating counterfactual perturbations. Second, we integrate explainability modules using SHAP values and counterfactual explanations to provide both global and local interpretability of credit decisions. Third, we accelerate the training and inference pipeline using Apache Spark (or Apache Flink) to scale to large datasets typical of financial institutions, ensuring that risk scoring can be done in near real-time.

 We evaluate our framework on a realistic credit dataset (simulated or proprietary) and demonstrate that (a) generative augmentation improves predictive performance (AUC, F1) compared to baseline models, (b) adversarial counterfactuals help uncover fragile regions in the model decision space, and (c) explainability techniques provide human-comprehensible rationales aligned with regulatory needs. Furthermore, leveraging Apache Spark for distributed training significantly reduces latency and increases throughput, making the system practically deployable in enterprise settings.

 Our contributions are threefold: (1) a unified architecture combining generative AI, adversarial robustness, and interpretability; (2) an empirical evaluation showing trade-offs between performance, security, and explainability; (3) a scalable implementation leveraging Apache big-data frameworks demonstrating feasibility in production. We argue that this trusted generative AI approach can reconcile the power of advanced models with the transparency and regulation requirements of credit risk management, paving the way for more inclusive, secure, and efficient lending decisions.