Image generation and classification based on generative adversarial networks and federated learning for enhanced data privacy

Although deep learning-based approaches have achieved remarkable success in medical image classification, challenges such as data privacy, class imbalance, and distributed data heterogeneity remain significant obstacles. In healthcare applications, the sensitive nature of medical image data restricts centralized data collection, thereby limiting the applicability of conventional machine learning paradigms. In this context, Federated Learning has emerged as an effective privacy-preserving learning framework by enabling model training without sharing raw data. Unbalanced data distributions and class imbalance across clients often degrade the performance of federated learning models. In this thesis, a comprehensive federated learning framework enhanced by Generative Adversarial Networks is proposed to address these challenges. Synthetic medical images are generated using CGAN and DCGAN architectures on the MedMNIST dataset and integrated into the federated learning-based classification process to mitigate data imbalance issues. Experimental results demonstrate that Generative Adversarial Network based data augmentation improves classification performance under certain conditions. The findings indicate that the proposed approach provides an effective and scalable solution for privacy-preserving medical image analysis and contributes to advancing federated learning applications in the healthcare domain.