Nowadays, counterfeiting and piracy are among the main problems for modern society. Counterfeiting of medication, food, cosmetics, mechanical parts and goods in general poses tremendous risks to public welfare and health, businesses and brand value reputation. Along with the fact that many traditional anti-counterfeiting technologies become quickly obsolete in view of the rapid technological progress that offers a wide range of modern high-tech tools and applications to the counterfeiters such as modern machine learning systems, high quality digital industrial printers and scanners. On the other hand, many new approaches to anti-counterfeiting appear thanks to the advancement of modern mobile technologies and machine learning algorithms.
As a new anti-counterfeit technology the printable graphical codes attract a lot of attention as a link between the physical and digital worlds and are broadly used on the Swiss market. However, the security of printable codes in terms of their reproducibility by unauthorized parties or clonability remains largely unexplored. In this respect Thesis addresses a problem of anti-counterfeiting of physical objects and aims at investigating the authentication aspects and the resistances to illegal copying of the modern printable graphical codes from machine learning perspectives. A special attention is paid to a reliable authentication on the modern mobile phones.
Most of modern copy detection technologies perform the detection of fakes based on a machine learning authentication that is a sub-class of general multi-class classification problem. However, the recent findings revealed vulnerability of classification techniques based on classical methods such as support vector machines and even deep learning classifiers. An active adversary can create various permutations to data known as adversarial examples targeting to trick the reliability of classification decision. These adversarial examples can be used against both digital and physical classification systems. For this reason, in Thesis, a particular attention is paid to the reliable classification in both digital and physical worlds under prior ambiguity that is related to both the lack of the labeled training data and the occurrence of the adversarial examples and their particular design and objectives behind.