Physical unclonable function anti-counterfeiting labels with deep learning authentication

Abstract

Physical Unclonable Function (PUF) is a recently developed anti-counterfeiting technique. the ability to generate strong anti-counterfeiting tags is the main reason for its vast development. Mostly, Convolutional Neural Network is used to authenticate these anti-counterfeiting tags due to its ability to automatically extract input image features. However, a very deep convolutional neural network must deal with overfitting. In the PUF authentication process, the main cause of overfitting is the minor alteration of PUF tags by a flow of time. In this project, the Resnet-feature extraction pair model is proposed to deal with the overfitting problem. The Resnet-feature extraction pair model combined extracted features from a convolution neural network and extracted features from the mathematical computation. Subsequently, these features are used to fit the Support Vector Machine. To evaluate its compatibility, the Resnet-feature extraction pair model is implemented in PUF authentication process by using 8 true PUF tags and 356 fake PUF tags. As a result, the Resnet-feature extraction pair model achieved a 15% accuracy improvement. Hence, it can be concluded that the Resnet-feature extraction pair model is a considerable tool for PUF authentication.