A Privacy-Preserving Data Pipeline Architecture for Federated and Generative AI Collaboration

Cross-organizational collaboration in artificial intelligence (AI) demands pipelines that protect privacy while maintaining scalability and trust. Centralized aggregation of sensitive data is increasingly untenable under modern regulations such as GDPR and HIPAA, and most existing federated or differentialprivacy systems remain siloed or limited to homogeneous, trusted environments. This paper presents a modular, cloud-agnostic architecture that embeds privacy enforcement and governance directly into federated AI pipelines. The proposed framework introduces a metadata-driven privacy controller that dynamically coordinates differential-privacy budgets, encrypted computation, and secure aggregation across heterogeneous organizations. It further incorporates adaptive trust management and gradient-validation mechanisms to mitigate malicious or compromised participants. A multi-node prototype evaluation using synthetic yet heterogeneous healthcare data demonstrates up to 75 % reduction in analytics latency, over 70 % lower infrastructure cost through serverless orchestration, and resistance to membership-inference attacks while preserving model utility (AUC$\approx 0.93$). The architecture is intended as a reusable reference design that also generalizes to unstructured data via privacy-preserving feature embeddings. These results highlight a practical foundation for scalable, privacy-first AI collaboration across institutional and data-modality boundaries.