Dynamic Subgroup Adaptive Secure Multi-Modal Heterogeneous Federated Learning (DSAS-MHFL)

Abstract

Federated Learning (FL) enables remote clients to collaborate on model training while keeping their data private. They do this without sharing raw data. However, current FL systems have significant limitations in real-world scenarios. They assume data distributions are the same, that client participation is constant, and that aggregation algorithms are rigid. When clients have different types of data, such as medical images, financial transactions, retail data, and healthcare records, these limitations become even more serious. In this study, propose a framework for privacy-preserving collaborative intelligence in diverse distributed settings. This framework is called Dynamic Subgroup Adaptive Secure Multi-Modal Heterogeneous Federated Learning (DSAS-MHFL). DSAS-MHFL brings together several techniques within a single architecture. The proposed strategy is based on prototype-based few-shot learning, drift-aware federated aggregation, and adaptive Gaussian noise masking for differential privacy. It also uses Multi-Head Self-Attention to obtain semantic representations and adaptive feature-space alignment to create a shared 64-dimensional embedding space. We calculate client model drift in real time to improve convergence stability and reduce utility loss. This calculation guides both noise scaling and aggregate weighting. The Heart Disease UCI, Credit Card Fraud Detection, Online Shoppers' Intention, and Chest X-ray Pneumonia datasets are used to evaluate the system. The results show that DSAS-MHFL is a scalable foundation for diverse real-world federated intelligence. It achieves a classification accuracy of over 91%, maintains stable convergence in federated settings, supports efficient cross-modal semantic matching, and performs well in few-shot classification, even with limited labeled data.