Open-Set Product Authentication Based on Deep Texture Verification

Abstract

The authenticity of consumer products has a significant impact on the economic and social issues of countries around the world. Due to the recent advancements in machine learning, there emerges some authentication techniques, in a close-set setting, based on texture features extracted from the product surfaces. However, the existing techniques have suffered from the problem of low accuracies or inability to deal with unknown classes in open-set authentication. In this work, we build an anti-counterfeiting system that works for consumer products in an open-set scenario. Different from other anti-counterfeiting methods, this work considers the problem of product authentication as a simple texture verification process. It allows the authentication being conducted under both close-set and open-set scenarios by a distance comparison operation with some customized metrics in the embedding space. We evaluate our system with two state-of-the-art texture databases. Experimental results show that the proposed system achieves 89.91% open-set authentication accuracy for a feature of 256 dimensions in the Outex texture database.