Code repository for bachelor thesis

This repository contains code and scripts for the bachelor thesis "Detect manipulated face Images using deep learning tools". This repository will mostly contain code for training and testing for the different models in the thesis. The other part will contain various utility scripts for preprocessing, and modifying data.

Directory structure of repository

.
├── classifiers
├── dataset_utils
├── MobaiSVM
├── morph_utils
└── multiple_feature_shapes

This repository has 5 main directories: classifiers, dataset_utils, MobaiSVM, morph_utils and multiple_feature_shapes Classifiers is the directory for the different training, testing code for classifiers modified and used in the thesis.

The datset_utils folder is for utilities and scripts used for preparing dataset. This directory includes files like renaming scripts and dataset validation scripts.

The MobaiSVM directory is the original SVM received from MOBAI AS, this is whats been used as a base for the modified classifiers in the classifiers directory.

The morph utils directory contains various scripts used to make the morphing algorithms run with the desired format.

The multiple_feature_shapes directory is for code used to figure out the best featureshape.

Example usecases for modified SVM classifier

How to train with a new dataset

Step 1:

Run create_dataset.py in mobai_dataset_utils:

Example:

python create_dataset.py \
  --original datasets/unibo/ \
  --new datasets/mordiff/ \
  --output datasets/mordiff_runnable/

Step 2:

Run the Svm with the dataset created in previous step.

Usage:

python svm_training_pipeline.py --bonaFideFeatures BONAFIDEFEATURES \
                                --morphedAttackFeatures MORPHEDATTACKFEATURES \
                                --modelOutput MODELOUTPUT \

If assert errors occur try removing all the .DS_store files.

You can do this by moving to the desired folder and running:

trash **/.DS_store

How to test with another dataset

Step 1:

Copy the model folder, since the results will be overwritten.

Step 2:

Example usage:

python3 svm_testing_pipeline.py --bonaFideFeatures unibo/Feature_Bonafide \
	--morphedAttackFeatures  unibo/Feature_Morphed \
	--modelOutput mordiff_runnable/model

How to merge/combine datasets

Step 1:

If the datasets does not have any probe images in it run create_dataset.py from mobai_dataset_utils with unibo dataset as original.

Example:

python create_dataset.py \
  --original datasets/unibo/ \
  --new datasets/mordiff/ \
  --output datasets/mordiff_runnable/

Do this for each dataset that is gonna be merged/combined.

Step 2:

To merge the datasets whilst keeping the size proportional run.

Run mege_dataset.py in mobai_dataset_utils

Example usage:

python merge_dataset.py --datasets UNIBO MORDIFF MIPGAN --ouput MERGED

To combine the datasets fully

Run combine_dataset.py in mobai_dataset_utils

Example usage:

python combine_dataset.py --datasets UNIBO MORDIFF MIPGAN --ouput COMBINED

Citations

mordiff

This repository has utilities and code for running among other things the MORDIFF morphing algorithm.

The github repository for mordiff: mordiff

MORDIFF paper: mordiff_paper

Mordiff SYN-mad benchmark: syn-mad

MIPGAN

This repository has utilities and code for running among other things the MIPGAN morphing algorithm.

Mipgan github repository: mipgan