DHARMA : Distributed Hybrid Architecture for Responsible Model Alignment and Assured Privacy

The integration of Artificial Intelligence (AI) into healthcare diagnostics relies on the fusion of multimodal data such as dermoscopic images and Electronic Health Records. Yet, strict data protection regulations (e.g., GDPR, HIPAA) restrict data centralisation and force sensitive information to remain siloed. Vertical Federated Learning (VFL) offers a mechanism for collaborative model training across feature-partitioned datasets without sharing raw patient records. However, current VFL systems face a dual challenge: limited performance due to naive fusion strategies that overlook complex cross-modal interactions, and a growing security risk whereby the expressive representations needed for high accuracy increase vulnerability to privacy inference and adversarial manipulation.This dissertation addresses these challenges through the DHARMA framework (Distributed Hybrid Architecture for Responsible Model Alignment and Assured Privacy), which combines client-side disentangled representation learning with server-side Cross-Modal Transformer fusion. Empirical results show that DHARMA substantially reduces the performance gap between VFL and centralised oracle models. The dissertation then conducts a rigorous adversarial stress test including number of integrity and privacy attacks to reveal a core tension: richer embeddings improve diagnostics but also heighten susceptibility to leakage. Finally, the work proposes a theoretical defence blueprint based on Information Bottleneck and Game-Theoretic principles, demonstrating that targeted defences can suppress adversarial success to near-random levels while preserving diagnostic utility. Together, these contributions establish a systematic methodology for building trustworthy, high-performance federated learning systems for medical AI.