Generative Adversarial Networks (GANs) have become a powerful tool for generating synthetic data, which is valuable in domains like credit card fraud detection, where the availability of real data may be limited. However, GANs are prone to mode collapse, a phenomenon where the model produces repetitive samples with limited diversity, reducing the quality and utility of the generated data. Early detection and mitigation of mode collapse is crucial to maintaining sample diversity, conserving computational resources, and improving model robustness. Despite various strategies for addressing this issue, identifying the collapse as it begins remains a significant challenge. Very few studies have focused on detecting mode collapse at the earliest possible stage, though increased attention in this area is essential for improving GANs’ performance. This paper proposes a novel approach to detect mode collapse early by monitoring fluctuations in the generator and discriminator loss values throughout training. Our approach initiates mode collapse detection dynamically after the model stabilizes during training, making it adaptable to any GAN architecture. By identifying collapse at its onset, our method allows for prompt intervention through targeted mitigation strategies, reducing wasted computational effort. As a result, this approach enhances GAN stability, improving its applicability and reliability across various real-world scenarios, from image synthesis to fraud detection.