GANs (Generative Adversarial Networks) are widely employed in many domains of science and technology. They have produced high-resolution, photo-realistic faces that appear real to the human eye. However, recognizing these images is becoming increasingly difficult. This study introduces a new GAN ensemble model that is intended to improve GAN training issues (Mode Collapse and Convergence) and generate knowledge from diverse input samples. The proposed model is built using multiple CNN discriminators architecture based on the voting ensemble technique. It utilizes a modified diversity loss function designed with the goal of minimizing the distance between the generated and original distributions. It is proven as a robust technique for forgery detection, achieving 98.31% accuracy values. One of the major findings of the study is that the novel method outperforms existing GAN models based on a small dataset of "Face Mask Lite"(193 Unmasked images), using quantifiable parameters such as Inception score (IS), Fréchet Inception distance (FID), SSIM (Structural Similarity Index Metric) and Total Computational Time Function.

GAN（生成对抗网络）广泛应用于许多科学技术领域。他们制作了高分辨率、逼真的面孔，在人眼看来是真实的。然而，识别这些图像变得越来越困难。本研究引入了一种新的 GAN 集成模型，旨在改善 GAN 训练问题（模式崩溃和收敛）并从不同的输入样本中生成知识。所提出的模型是使用基于投票集成技术的多个 CNN 判别器架构构建的。它利用改进的多样性损失函数，其设计目的是最小化生成的分布和原始分布之间的距离。它被证明是一种强大的伪造检测技术，准确率高达 98.31%。该研究的主要发现之一是，该新方法优于基于“Face Mask Lite”小数据集（193 张 Unmasked 图像）的现有 GAN 模型，使用可量化参数，例如 Inception Score (IS)、Fréchet Inception distance (FID) ）、SSIM（结构相似性指数度量）和总计算时间函数。