PUF authentication

Introduction

This is the implementation for our paper Random fractal-enabled physical unclonable functions with dynamic AI authentication.

It contains the preprocessing of the PUF patterns, training of ResNet50 based classification model, update of the model with newly added PUFs, testing dataset generation, and the test of the model on the testing dataset. Besides, the algorithms for the computation of the histogram of the PUFs and similarities between PUFs are also listed here.

Dataset

The dataset is located in the folder ./data, where

all_rgb: original images of the PUF patterns

all: grayscale images,

all_render_bg: images after the preprocessing

annotations: json files for training, adding, and testing the method

features: the CNN features generated for the base dataset

similarity: images used for the computation of similarity

Running the code

1. Create the environment using the file puf_authentication.yml

2. Training the initial base model

python train_class_only_init.py

3. Update the model with newly added PUF patterns

python train_class_only_add.py --a_n_classes 200 --acc_thre 95 --batch_size 2000

4. Generate the test dataset before testing

python generate_test_set.py

5. Test the model

python test_all_pipeline.py

For cite of this work:

@article{sun2022random,
	title	=	{Random fractal-enabled physical unclonable functions with dynamic AI authentication},
	author	=	{Sun, Ningfei and Chen, Ziyu and Wang, Yanke and Wang, Shu and Xie, Yong and Liu, Qian},
	year	=	{2022}
}