Official implementation of Protected Attribute Suppression System, ICCV 2021

Introduction

This repository contains the source code for training PASS-g and PASS-s using features from a pre-trained model.

• Dhar, Prithviraj and Gleason, Joshua et al. "PASS: Protected Attribute Suppression System for Mitigating Bias in Face Recognition." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.

BibTeX:

@InProceedings{Dhar_Gleason_2021_ICCV,
    author    = {Dhar, Prithviraj and Gleason, Joshua and Roy, Aniket and Castillo, Carlos D. and Chellappa, Rama},
    title     = {{PASS}: Protected Attribute Suppression System for Mitigating Bias in Face Recognition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {15087-15096}
}

Running The Code

Requirements are defined in requirements.txt and may be installed in a new virtual environment using

pip install -r requirements.txt

An example configuration is defined in config/config_template.yaml.

In the config file set TYPE:'race' for PASS-s or TYPE:'gender' for PASS-g.

Required Input Files

Training features (train.py)

This file should be provided in the TRAIN_BIN_FEATS and VAL_BIN_FEATS config entries. Must be a binary file. Given a numpy array of N 512-dimensional features you can create this file using the following snippet (note we assume binary file created with same byte order as system used to train)

import numpy as np
import struct

# feat = ... (load features into np.ndarray of shape [N, 512])
# ...

with open('input_features.bin', 'wb') as f:
    f.write(struct.pack('i', np.int32(N)))
    f.write(struct.pack('i', np.int32(512)))
    np.ascontiguousarray(feat).astype(np.float32).tofile(f)

Training metadata (train.py)

This file should be provided in the TRAIN_META and VAL_META config entries. This CSV file must contain information about each training feature (one-to-one corresponding) and must contain the following columns:

SUBJECT_ID,FILENAME,RACE,PR_MALE

• SUBJECT_ID is an integer corresponding to subject
• FILENAME is original filename that feature was extracted from (not used currently)
• RACE is an integer representing a BUPT class label between 0 and 3 with {0: asian, 1: caucasian, 2: african, 3: indian}
• PR_MALE is a float between 0 and 1 representing probability that subject is male

Note that for PASS-g RACE may be omitted and for PASS-s PR_MALE may be omitted.

Test features (inference.py)

CSV file containing features to perform debiasing on after training is finished with following columns:

SUBJECT_ID,FILENAME,DEEPFEATURE_1,...,DEEPFEATURE_512

where DEEPFEATURE_* contains the value of the input feature at the specified dimension.

To run PASS training execute

python train.py

To generate debiased features, select the desired checkpoint file and update CHECKPOINT_FILE in the config then run

python inference.py