Authentication of printed documents using high resolution 2D codes relies on the fact that the printing process is considered as a Physical Unclonable Function used to guaranty the security of the authentication system. The 2D code is corrupted by the printing process in a non-invertible way by inducing decoding errors, and the gap between the bit error rate generated after the first and second printing processes enables to perform the authentication of the document. In this context, the adversary's goal is to minimize the amount of decoding errors obtained from the printed code in order to generate a forgery which can be considered as original. The goal of this paper is to maximize the decoding performance of the adversary by inferring the original code observing the printed one. After presenting the different kinds of features that can be derided from the 2D code (the scanner outputs, statistical moments, features derived from Principal Component Analysis and Partial Least Squares), we present the different classifiers that have been evaluated and show that the bit error rate decreases from 32% using the baseline decoding to 22% using appropriated features and classifiers.