Federated Contrastive Feature Encoding for Privacy-Preserving and Communication-Efficient Healthcare Intelligence

Federated learning (FL) has developed into a fundamental technique for training in healthcare while maintaining patient privacy. However, challenges exist with FL, including communication costs, non-Identically Distributed (non-IID) distributions of data from clients, and information loss from compression methods such as the Discrete Cosine Transform (DCT). This limits diagnostic accuracy, notably in applications that are as important as medical image analysis. To overcome these challenges, the proposed Federated Contrastive Feature Encoding (FedCFE) is a communication-efficient and privacy-preserving federated learning framework based on the learned contrastive feature compression. Initially, data were collected from the COVID-19 Radiography Dataset. Further, data are preprocessed utilizing resizing, normalization, and data augmentation, when appropriate, to ensure meaningful input and that the representation generalizes better during the compression phase. A shared ResNet-18 encoder is employed to extract high-level semantic features from the input. Next, higher-level features were compressed using a projection head. Contrastive learning ensures that the embeddings for all samples within the same disease class remain close together while separating the embeddings for other classes, which it sends back to the clients for local fine-tuning. The proposed FedCFE attained better results, when compared to ModelContrastive Federated Learning-Compressive Sensing (MCFLCS) in accuracy (99.87 %), precision (99.48 %), recall (98.89 %) and F1- Score (99.03 %) respectively.