The security of CDP is based on the difficulty to reverse opera8on of PD process (i.e., to es8mate the original code from its printed version). Nevertheless, the advances of deep learning techniques have made possible to reverse the PD process. That is why, it is urgent to develop novel metrics or machine learning models able to differen8ate the es8mated codes from the original unclonable codes. It was shown that in the case of supervised classifica8on (when all possible fakes are known in the moment of training), the classical machine learning techniques can easily detect all the fakes. Nevertheless, when the fakes are unknown during the training stage or while the fakes are printed using the same device as the authen8c samples, the current detectors are incapable of separa8ng the originals from fakes. The recent work presents the first tenta8ve to consider the PD process and to detect the anomalies in the printable unclonable code. However, the current model cannot perfectly imitate the PD process and thus the anomalies were searched only in the pixels that are correctly imitated by the PD model (41-43% of code). In addi8on, current anomaly detectors are dependent to the training dataset (i.e., on the printer and acquisi8on device used). This PhD project aims to develop a robust AI-based authen8ca8on detector for CDP by studying similarity metric learning approaches [6] and forensic techniques. The PhD student will inves8gate how to combine forensic features with recent deep neural networks and train them to be robust against es8ma8on aTacks. The student will be also encouraged to follow the best prac8ces of reproducible research to share his/her advances with the research community.