FLGAN is a GAN-based approach designed to address the challenges of federated learning (FL) under non-IID (non-Independent and Identically Distributed) data settings. FLGAN mitigates local biases caused by non-IID data by generating synthetic samples using a federated GAN, while preserving user-level privacy. The method employs a divide-and-conquer strategy to avoid model collapse in non-IID scenarios and uses Fully Homomorphic Encryption (FHE) to ensure privacy during GAN training. FLGAN is implemented with a federated GAN that generates global IID data, enabling unbiased FL aggregation. The system model includes a central server and multiple users, each training a local GAN model. FLGAN ensures user-level privacy by using FHE to protect local data and parameters. The method is tested against two state-of-the-art FL baselines (FedAvg and FedSGD) under different non-IID settings, achieving up to 60% accuracy improvement. FLGAN's FHE-based privacy guarantees add only 0.53% overhead. The proposed scheme addresses the challenges of non-IID FL by generating synthetic data to balance local distributions, ensuring unbiased FL performance while maintaining user privacy. Theoretical analysis confirms the security and convergence of FLGAN under non-IID settings.FLGAN is a GAN-based approach designed to address the challenges of federated learning (FL) under non-IID (non-Independent and Identically Distributed) data settings. FLGAN mitigates local biases caused by non-IID data by generating synthetic samples using a federated GAN, while preserving user-level privacy. The method employs a divide-and-conquer strategy to avoid model collapse in non-IID scenarios and uses Fully Homomorphic Encryption (FHE) to ensure privacy during GAN training. FLGAN is implemented with a federated GAN that generates global IID data, enabling unbiased FL aggregation. The system model includes a central server and multiple users, each training a local GAN model. FLGAN ensures user-level privacy by using FHE to protect local data and parameters. The method is tested against two state-of-the-art FL baselines (FedAvg and FedSGD) under different non-IID settings, achieving up to 60% accuracy improvement. FLGAN's FHE-based privacy guarantees add only 0.53% overhead. The proposed scheme addresses the challenges of non-IID FL by generating synthetic data to balance local distributions, ensuring unbiased FL performance while maintaining user privacy. Theoretical analysis confirms the security and convergence of FLGAN under non-IID settings.