SNF Project: Information-theoretic analysis of deep

identification systems


1. Lead


The vast majority of consumer goods consimed in our society is targeted by counterfeiters. The advent of machine learning, including the deep neural networks, has encouraged the appearance of powerful simulation technologies capable to reproduce and generate complex images and textures. These tools are well known and publicly available. Therefore, it is important to develop technologies allowing to accurately detect and possibly identify counterfeited products without special equipment.

2. Aims of the research project

The goal of the project is to extend and further develop a new theoretic framework for physical object identification using deep representations and apply it to practical systems. This goal will be achieved based on a novel interpretation of links between machine learning, variational inference, information-theoretic coding, models of the production process, attacks, physics of uncloneble features and particularities of identification taking the specifics of mobile imaging. Finally, the field also currently lacks public datasets, which are needed to gain reproducible results and benchmarking. 

3. Keywords

Security; anticounterfeiting; adversarial attacks; clonability; deep identification; information-theoretical analysis; database; unsupervised learning; mobile imaging.