Counterfeiting of currency globally remains a significant problem to this day. According to the authorities, a large portion of this fake currency is produced by Small Office or Home Office inkjet printers. In this paper, we explain why a previously developed machine learning based
Printer Identification System works with high accuracy and we investigate to improve the stability and generalization of the classifier.
We study the print patterns of 8 inkjet printers from different manufacturers. We look at the features of the data by reducing its dimensions using
Principle Component Analysis. This shows significant separation between printers which implies that the Deep Neural Network was able to pick up on key differences. The results are also comparable with that of reducing dimensions with Linear Discriminant Analysis. The model however does have
some limitations regarding ink density and print media. It always classifies an image amongst the trained printers and does not show anomalies. For this, we consider the Gaussian distribution of target printers to see how the probabilities fared when trained and fitted separately for each
printer, thus having a set of images that do not fit in with any of the printers for which the classifier was trained. The results acquired from these methods have contributed to making a more real-world implementation of our classifier, named LAPIS (machine-Learning Applied Printer Identification
System), for printer forensics.