Recent papers point out the vulnerability of Copy Sensitive Graphical Codes (CSGC) while an opponent uses a neural network approach to estimate a pattern then prints it as an original one: such a fake can successfully pass the authentication test. Here, we show that a GAN-like network can be even more powerful. A SRGAN-based architecture including superresolution can tolerate a lower scanner resolution and decode efficiently. Besides, the use of such a decoding technique to perform the authentication test can improve the resistance of CSGC to estimation attacks.