Welcome to visit the homepage of our paper entitled "A Post-Training Framework for Improving the Performance of Deep Leaning Models via Model Transformation". The homepage contains the source code of our fairness improvement approach FMT and other existing fairness improvement methods that we use in our paper, as well as the intermediate results, the installation instructions, and a replication guideline. Particularly, the implementation and experimental results of our paper regarding improving the robustness of DL models can be downloaded at Dare.

We use Python 3.7 for our experiments. We use the IBM AI Fairness 360 (AIF360) toolkit for implementing fairness improvement methods and computing fairness metrics.

Installation instructions for Python 3.7 and AIF360 can be found on https://github.com/Trusted-AI/AIF360. That page provides several ways for the installation. We recommend creating a virtual environment for it (as shown below), because AIF360 requires specific versions of many Python packages which may conflict with other projects on your system. If you would like to try other installation ways or encounter any errors during the installation proces, please refer to the page (https://github.com/Trusted-AI/AIF360) for help.

Conda is recommended for all configurations. Miniconda is sufficient if you do not already have conda installed.

Then, to create a new Python 3.7 environment, run:

conda create --name fmt python=3.7
conda activate fmt
pip install -r requirements.txt
# pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

We use the 4 default datasets supported by the AIF360 toolkit. When running the scripts that invoke these datasets, you will be prompted how to download these datasets and in which folders they need to be placed. You can also refer to https://github.com/Trusted-AI/AIF360/tree/master/aif360/data for the raw data files.

Dare uses half of the training set to train the model, and our data partitioning is stored in the data folder, while the original model used is stored in the models folder.

Dare needs to use the trained historical model to perform the data augmentation process and collect training data for fine-tuning. However, due to the large number of historical models, it is not possible to upload all of them. Therefore, we provide the intermediate results of data augmentation, 'dataset_model_modify' is used to mark the selected training data, 'dataset_model_rs' is the standard output collected for training data. In addition, another time-consuming part is generating adversarial samples, so we store adversarial data in the adv folder. You can refer to https://mega.nz/folder/1alCBThJ#JLh9CC6lY0FpOIP6icgh8w for all the data files used for DARE.

For FMT, we present the p-value results of our calculations using the Wilcoxon signed-rank test against all baseline methods, as shown in the following table:

The repository contains the following files and folders:

• baselines/ contains code for implementing four state-of-the-art fairness improvement approaches (i.e., REW, ADV, ROC and CARE).
• DARE/ contains code for the implementation of our DARE method.
• data/ contains original datasets and our processed data.
• datasets/ contains code for processing raw datasets.
• examples/ contains script files for running the code of all methods.
• features/ contains json files for feature mapping when processing datasets.
• models/ contains our trained source model used for the following fairness improvement.
• plot/ contains code for drawing plots and pictures.
• data_helper.py is the code for generating processed data and mutation samples.
• fmt.py is the code for our method Fairness improvement by Model Transformation (FMT).
• fmt_variants.py is the code for the three variants of our FMT.
• mutation.py is the code for mutant generation.
• requirements.txt is the list of required packages and corresponding versions for environment setup.
• train.py is the code for training source models used for fairness improvement.
• utils.py is the code for some utility functions used by our methods.
• wilcoxon_test_dare.py, wilcoxon_test_fmt.py contain the code for calculating the Wilcoxon signed-rank test.
• RQ5_results/, RQ6&7_results/ contain the raw results of the models after applying different fairness improvement methods/strategies. These folders also contain some plots.

You can reproduce the results for our research questions (RQs) based on the intermediate results provided by us.

The trained source models are saved in models directory. You can also reproduce them as follows:

python train.py -d adult
python train.py -d bank
python train.py -d compas
python train.py -d german

This RQ evaluates the effectiveness of FMT over four datasets involving seven different privileged attributes, and compares the results with four state-of-the-art fairness improvement approaches. To answer this RQ, we need to run the code as follows.

(1) We obtain the ML performance and fairness metric values obtained by our approach FMT.

fmt.py supports three arguments: -d configures the dataset; -p configures the protected attribute; -m configures the training data construction strategy.

python fmt.py -d adult -p race -m single
python fmt.py -d adult -p race -m dual
python fmt.py -d adult -p sex -m single
python fmt.py -d adult -p sex -m dual

python fmt.py -d bank -p age -m single
python fmt.py -d bank -p age -m dual

python fmt.py -d compas -p race -m single
python fmt.py -d compas -p race -m dual
python fmt.py -d compas -p sex -m single
python fmt.py -d compas -p sex -m dual

python fmt.py -d german -p age -m single
python fmt.py -d german -p age -m dual
python fmt.py -d german -p sex -m single
python fmt.py -d german -p sex -m dual

As a result, we can obtain the results of FMT. The result for each combination is included in the RQ5_results/fmt/ folder.

(2) We obtain the ML performance and fairness metric values obtained by existing bias mitigation methods in the ML community: REW (baselines/Fair360/rew.py), ROC (baselines/Fair360/roc.py), and ADV (baselines/Fair360/adv.py). The three methods support three arguments: -d configures the dataset; -p configures the protected attribute. We take the REW method as an example to show how to run the code.

python baselines/Fair360/rew.py -d adult -p race
python baselines/Fair360/rew.py -d adult -p sex

python baselines/Fair360/rew.py -d bank -p age

python baselines/Fair360/rew.py -d compas -p race
python baselines/Fair360/rew.py -d compas -p sex

python baselines/Fair360/rew.py -d german -p age
python baselines/Fair360/rew.py -d german -p sex

This RQ investigates how much the model transformation in our framework contributes to the effectiveness of FMT. We conduct an ablation study with the variants of $FMT^{-s}$ and $FMT^{−sl}$, which are implemented as fmt_variants.py. The results are included in the RQ6_results/ folder.

python fmt_variants.py -d adult -p race -m single -v remove_s
python fmt_variants.py -d adult -p race -m single -v remove_sl
python fmt_variants.py -d adult -p race -m dual -v remove_s
python fmt_variants.py -d adult -p race -m dual -v remove_sl

python fmt_variants.py -d adult -p sex -m single -v remove_s
python fmt_variants.py -d adult -p sex -m single -v remove_sl
python fmt_variants.py -d adult -p sex -m dual -v remove_s
python fmt_variants.py -d adult -p sex -m dual -v remove_sl

python fmt_variants.py -d bank -p age -m single -v remove_s
python fmt_variants.py -d bank -p age -m single -v remove_sl
python fmt_variants.py -d bank -p age -m dual -v remove_s
python fmt_variants.py -d bank -p age -m dual -v remove_sl

python fmt_variants.py -d compas -p race -m single -v remove_s
python fmt_variants.py -d compas -p race -m single -v remove_sl
python fmt_variants.py -d compas -p race -m dual -v remove_s
python fmt_variants.py -d compas -p race -m dual -v remove_sl

python fmt_variants.py -d compas -p sex -m single -v remove_s
python fmt_variants.py -d compas -p sex -m single -v remove_sl
python fmt_variants.py -d compas -p sex -m dual -v remove_s
python fmt_variants.py -d compas -p sex -m dual -v remove_sl

python fmt_variants.py -d german -p age -m single -v remove_s
python fmt_variants.py -d german -p age -m single -v remove_sl
python fmt_variants.py -d german -p age -m dual -v remove_s
python fmt_variants.py -d german -p age -m dual -v remove_sl

python fmt_variants.py -d german -p sex -m single -v remove_s
python fmt_variants.py -d german -p sex -m single -v remove_sl
python fmt_variants.py -d german -p sex -m dual -v remove_s
python fmt_variants.py -d german -p sex -m dual -v remove_sl

This RQ investigates the performance of our specially designed data augmentation algorithm. We compare the results of FMT with the variant of $FMT^{rand}$, which is implemented as fmt_variants.py. The results are included in the RQ7_results/ folder.

python fmt_variants.py -d adult -p race -m single -v rand
python fmt_variants.py -d adult -p race -m dual -v rand
python fmt_variants.py -d adult -p sex -m single -v rand
python fmt_variants.py -d adult -p sex -m dual -v rand

python fmt_variants.py -d bank -p age -m single -v rand
python fmt_variants.py -d bank -p age -m dual -v rand

python fmt_variants.py -d compas -p race -m single -v rand
python fmt_variants.py -d compas -p race -m dual -v rand
python fmt_variants.py -d compas -p sex -m single -v rand
python fmt_variants.py -d compas -p sex -m dual -v rand

python fmt_variants.py -d german -p age -m single -v rand
python fmt_variants.py -d german -p age -m dual -v rand
python fmt_variants.py -d german -p sex -m single -v rand
python fmt_variants.py -d german -p sex -m dual -v rand

Thanks to the authors of existing fairness improvement methods for open source, to facilitate our implementation of this paper. Therefore, when using our code or data for your work, please also consider their repositories, including AIF360, DARE and MAAT.