Stabilized Performance Maximization for GAN-based Real-Time Authentication Image Generation over Internet

Providing the completely automated public test to tell computers and humans apart (CAPTCHA) services that are not vulnerable to learning is an important issue. Image-based CAPTCHA services have strengthened authentication by taking advantage of the fact that it is more difficult for computers to understand images, but the rapid growth of deep learning has made it possible to break and attack authentication. Image-based CAPTCHA uses pre-stored data, which enables learning and results in vulnerability. This paper presents an adaptive generative adversarial network (GAN) selection scheme using time-average image quality maximization subject to system/buffer stability. By using GAN, it can generate and provide new images for CAPTCHA authentication every time, preventing deep learning from learning images and enhancing security. In the image generation process, the trade-off exists between image quality and generation time, and in consideration of this trade-off, delay aware image-based authentication Lyapunov-based algorithm is proposed for stable and maximized performance. Moreover, through the performance evaluation, we investigate and show the existence of trade-off between generation time and generated image quality in the image generation process in both quantitative and qualitative manner.