Development of an Enhanced Federated Feedforward Deep Learning Framework for Semantic Communication in Cross Institutional Research Collaboration

Abstract

Intelligent communication frameworks that guarantee data privacy, scalability, and effective knowledge exchange across diverse environments are becoming more and more necessary for cross-institutional research collaboration. In order to facilitate privacy-preserving cooperation between several institutions managing multimodal research data, this study suggests a distributed semantic communication platform built on Federated Feed-Forward Neural Networks (FFNNs). Graph neural networks (GNNs) and convolutional neural networks (CNNs), which have been extensively documented in the literature for semantic communication tasks but have drawbacks in distributed, multimodal, and enterprise-level deployments, are compared. Unlike CNNs and GNNs, the proposed FFNN model is implemented using C# and ML.NET, ensuring compatibility with enterprise systems, maintainability, and scalable federated training. To send meaningful data representations instead of raw datasets, the platform incorporates semantic encoding. This lowers communication costs, facilitates low-latency interaction, and protects data security. The FFNN based system obtains an accuracy of 93.8%, precision of 92.4%, recall of 94.1%, and an F1-score of 93.2% across structured, textual, and experimental datasets, according to experimental evaluation conducted under simulated federated conditions. Despite non-IID data distributions, convergence is reached in 11 federated rounds, with a 42% reduction in communication overhead. In comparison to CNN- and GNN-based methods, the results show that the suggested semantic enabled FFNN framework provides a better balance of accuracy, communication efficiency, scalability, and privacy preservation, making it a workable and efficient solution for safe, extensive, cross-institutional research collaboration.