Hybrid Machine Learning for Privacy-Preserving Multi-Modal Data Analytics: An Application-Oriented Framework

Abstract

In the context of applications in healthcare, finance, and smart cities, the volume of sensitive information is increasing rapidly. The conventional centralized mechanisms pose challenges as ownership losses (unauthorized access), data leakage, and non-compliance with GDPR and HIPAA privacy regulations. To overcome these, two approaches are integrated with best practice components in the architecture, which progress towards a novel framework with modular, integrated FL-DP, such as Federated learning and Differential Privacy. In this, Federated learning uses decentralized training of the model among multiple entities without sharing raw data, whereas differential privacy injects calibrated noise to avoid member attacks and inference attacks for ensuring guaranteed privacy. The hybrid model uses attributes such as sensor data, clinical information, and contextual content by ensuring high modal utility. The hybrid approach provides a state-of-the-art framework for privacy-preserving that enables scalable, secure, and compliant data processing in a distributed environment. Both privacy and accuracy are achieved using client-level differential privacy with adaptive privacy budgeting, cross-modal feature fusion, and momentum-based federated optimization. Experimental results show that the hybrid model outperforms the benchmark methods, such as Centralized frameworks, Homomorphic encryption, Multi-party Computation (MPC), and Trusted execution environments. In addition, it supports multi-party collaboration, and heterogeneous data sources.