IMAGE ANONYMIZATION DETECTION WITH DEEP HANDCRAFTED FEATURES

This is the original code implementation of the paper:

N. Bonettini, D. Güera, L. Bondi, P. Bestagini, E.J. Delp, S. Tubaro, "Image Anonymization Detection With Deep Handcrafted Features", IEEE International Conference on Image Processing (ICIP), 2019

Please, cite this if you use this code for your research.

Clone the repository

PRNU functions are contained into the submodule prnu. In order to clone the repository with the submodule you need to add the --recurse-submodules flag to the clone command, for instance:

git clone --recurse-submodules https://github.com/polimi-ispl/prnu-anonymization-detector.git

Alternatively, you can clone the prnu repository into the empty prnu folder.

Prerequisites

Python 3.6
Pytorch
Numpy
Scipy
Pandas
PIL
Scikit-image
Scikit-learn
tqdm
TensorboardX (optional)
Jupyter (optional)

Configuration

Change root_path in params.py according to your project root folder.

All the command below are meant to be run in your project root folder.

Build db

python generate_db.py

This will generate several .csv files in data/db, containing the paths of the images in data/dataset. For convenience, we included a very reduced subset of the images we considered for this work, alongside their matching PRNUs. The original images are from Dresden dataset.

If you want to use your own set of images, just put them in dataset and run generate_db.py.

Finetuning (training) a model

python finetuning.py --gpu 0 --db D1 --model resnet --num_workers 24 --batch_size 24 --transform_pre wv_fft_wiener2

This will generate a run folder under the folder specified in runs_path variable of params.py. The run name is composed of all the train parameters value and a 6-digits alphanumeric code, $RUN from now on.

Cross dataset testing

python test.py --gpu 0 --db TestOS --model resnet --num_workers 1 --batch_size 1 --transform_pre wv_fft_wiener2 --runs $RUN

Leave one out testing

python test.py --gpu 0 --db D3_no_D200_0 --model resnet --num_workers 1 --batch_size 1 --transform_pre wv_fft_wiener2 --test_on_train --runs $RUN

Transformation testing

python test.py --gpu 0 --db TestOS --model resnet --num_workers 1 --batch_size 1 --transform_pre wv_fft_wiener2 --transform_test --runs $RUN

Paper plots

jupyter notebook notebook/paper_figures.ipynb