Open Access

ARTICLE

Generated Preserved Adversarial Federated Learning for Enhanced Image Analysis (GPAF)

Sanaa Lakrouni^*, Slimane Bah, Marouane Sebgui

Smart Communication Research Team, Mohamedia School of Engineers, University Mohammed V in Rabat, 10100, Morocco

* Corresponding Author: Sanaa Lakrouni. Email:

(This article belongs to the Special Issue: Multi-Modal Deep Learning for Advanced Medical Diagnostics)

Computers, Materials & Continua 2025, 85(3), 5555-5569. https://doi.org/10.32604/cmc.2025.067654

Received 09 May 2025; Accepted 26 August 2025; Issue published 23 October 2025

Abstract

Federated Learning (FL) has recently emerged as a promising paradigm that enables medical institutions to collaboratively train robust models without centralizing sensitive patient information. Data collected from different institutions represent distinct source domains. Consequently, discrepancies in feature distributions can significantly hinder a model’s generalization to unseen domains. While domain generalization (DG) methods have been proposed to address this challenge, many may compromise data privacy in FL by requiring clients to transmit their local feature representations to the server. Furthermore, existing adversarial training methods, commonly used to align marginal feature distributions, fail to ensure the consistency of conditional distributions. This consistency is often critical for accurate predictions in unseen domains. To address these limitations, we propose GPAF, a privacy-preserving federated learning (FL) framework that mitigates both domain and label shifts in healthcare applications. GPAF aligns conditional distributions across clients in the latent space and restricts communication to model parameters. This design preserves class semantics, enhances privacy, and improves communication efficiency. At the server, a global generator learns a conditional feature distribution from clients’ feedback. During local training, each client minimizes an adversarial loss to align its local conditional distribution with the global distribution, enabling the FL model to learn robust, domain-invariant representations across all source domains. To evaluate the effectiveness of our approach, experiments on a medical imaging benchmark demonstrate that GPAF outperforms four FL baselines, achieving up to 17% higher classification accuracy and 25% faster convergence in non-IID scenarios. These results highlight GPAF’s capability to generalize across domains while maintaining strict privacy, offering a robust solution for decentralized healthcare challenges.

---

Keywords

---