Greedy-DiM (IJCB 2024)

Greedy-DiM: Greedy Algorithms for Unreasonably Effective Face Morphs

Zander W. Blasingame   ·   Chen Liu

Clarkson University

News

• 2024.12.17: Dataset for morphs created on FRLL is now public!
• 2024.11.11: Camera ready version is now available on IEEE Xplore!
• 2024.09.26: We have released the main code. Configuration details for Greedy-DiM and past models (DiM and Fast-DiM) are found in configs
• 2024.09.17: We presented our work in Buffalo, NY. Thanks to everyone who stopped by our posters and listened to our talk!
• 2024.07.16: Congratulations on Greedy-DiM for being accepted by IJCB 2024! Our code base is under development, stay tuned for updates.

Introduction

The official PyTorch implementation of Greedy-DiM (IJCB 2024 Spotlight), DiM (IEEE TBIOM and IJCB 2024 Oral), and Fast-DiM (IEEE Security & Privacy).

Greedy-DiM is a simple yet unreasonably effective face morphing algorithm that far suprasses previous representation-based morphing algorithms and even beats landmark-based morphing algorithms. With a simple greedy guided algorithm Greedy-DiM is able to significantly improve the effectiveness of DiM while still retaining the high-visual fidelity that is characteristic of DiM. Moreover, we prove that the search space of the Greedy-DiM is well-posed and that it contains the optimal morphed face. Experimental results show that Greedy-DiM is currently the strongest face morphing algorithm available, pushing the SOTA.

The heart of the Greedy-DiM* algorithm.

Release Notes

This repository is an updated version of DiM, with several new features:

• Greedy guided generation for Diffusion Morphs (Greedy-DiM) (Greedy-DiM*, Greedy-DiM-S)
• Simplified codebase for switching between different DiM variants, i.e., DiM, Fast-DiM, and Greedy-DiM
• General optimization improvements: reduced memory usage, faster inference, &c.

Supported Models

In addition to Greedy-DiM* we also support the following models:

Model	Description
DiM	The original DiM model, named DiM-A in Table 5
Fast-DiM	Uses second-order multistep ODE solver to solve the PF ODE to reduce NFE
Fast-DiM-ode	Same as Fast-DiM plus an improved formulation to solve the PF ODE forwards in time, resulting in reduced NFE
Greedy-DiM-S	Greedy-DiM with a greedy search strategy results in improved MMPMR and uses less memory

with configurations files for these different models are found in the configs directory. Their efficacy measured in MMPMR (Mated Morphed Presentation Match Rate) can be observed in Table 5 from our paper shown below.

Dataset

We have released the morphed images used in our papers on the FRLL dataset at https://huggingface.co/datasets/zblasingame/DiM-FRLL-Morphs

Code Release

In accordance with the our agreement with CITeR we have encrypted our morphing code. To obtain access to it sign the code release form and send it to citer@clarkson.edu. Upon approval the passphrase requested by the installer will be sent over.

Installation

To install this repository first clone the repo using git. Then run the provided installation script install.sh. This script will setup the virtualenv and install all dependencies for the project. Lastly, the installation script will ask for passphrase to decrypt run_dim.py.gpg, the passphrase can be obtained by filling out the code release form.

Setting up external models

In this repository we make use of two pre-trained models from other projects.

1. The diffusion backbone used in this project is the Diffusion Autoencoder from the diffae repository. Install the ffhq256_autoenc/last.ckpt checkpoint and place in the checkpoints directory.

2. The identity loss makes use of the ArcFace FR system. To download the same version we used in our experiments navigate to the arcface repository and download the glint360k_cosface_r100_fp16_0.1 folder and place it in the arcface folder.

Usage

First activate the virtual environment with

source venv/bin/activate

Then run morphing script run_dim.py, which takes the following arguments:

usage: run_dim.py [-h] [-b BATCH_SIZE] morph_list config

positional arguments:
  morph_list            A CSV file containing paths to morph pairs and output directory, an example line:
                        "source_dir/id_a.png,source_dir/id_b.png,output_dir/morph_a_b.png"
  config                YAML file containing configuration for the run

optional arguments:
  -h, --help            show this help message and exit
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        The batch size for processing.

The CSV describes the images to be morphed together and consists of rows in the following format

path/to/source/id_a.png,path/to/source/id_b.png,path/to/output/morph_a_b.png
path/to/source/id_c.png,path/to/source/id_d.png,path/to/output/morph_c_d.png

N.B., the morphing script expects the images to already be aligned. We used the dlib alignment process used to create FFHQ.

Additionally, the morphing script expects a configuration file which details the structure of the experiment. For example the Greedy-DiM* config file looks like

# Configuration file for Greedy-DiM* algorithm. See Algorithm 1 in https://arxiv.org/abs/2404.06025
---
encoding_timesteps: 250
sampling_timesteps: 20
scheduler_kwargs:
  n_train_steps: 1000
  solver_order: 1
  prediction_type: epsilon
  algorithm_type: dpmsolver++
encoding_solver: diffae
greedy:
  type: opt
  kwargs:
    n_opt_steps: 50
    opt_stride: 1
    opt_kwargs:
      lr: 0.01
      betas: [0.5, 0.9]
loss_fn:
  type: zhang_identity_prior
  arcface_backbone: arcface/glint360k_cosface_r100_fp16_0.1/backbone.pth

We provide a configuration file for several DiM variants.

Citation

If this work is helpful for your research, we ask that you consider citing the relevant papers:

@INPROCEEDINGS{blasingame_greedy_dim,
  author={Blasingame, Zander W. and Liu, Chen},
  booktitle={2024 IEEE International Joint Conference on Biometrics (IJCB)}, 
  title={Greedy-DiM: Greedy Algorithms for Unreasonably Effective Face Morphs}, 
  year={2024},
  volume={},
  number={},
  pages={1-11},
  keywords={Greedy algorithms;Schedules;Face recognition;Biological system modeling;Closed box;Generative adversarial networks;Diffusion models;Iterative methods;Optimization},
  doi={10.1109/IJCB62174.2024.10744517}}


@article{blasingame_dim,
   title={Leveraging Diffusion for Strong and High Quality Face Morphing Attacks},
   volume={6},
   ISSN={2637-6407},
   url={http://dx.doi.org/10.1109/TBIOM.2024.3349857},
   DOI={10.1109/tbiom.2024.3349857},
   number={1},
   journal={IEEE Transactions on Biometrics, Behavior, and Identity Science},
   publisher={Institute of Electrical and Electronics Engineers (IEEE)},
   author={Blasingame, Zander W. and Liu, Chen},
   year={2024},
   month=jan, pages={118–131}}

@article{blasingame_fast_dim,
   title={Fast-DiM: Towards Fast Diffusion Morphs},
   volume={22},
   ISSN={1558-4046},
   url={http://dx.doi.org/10.1109/MSEC.2024.3410112},
   DOI={10.1109/msec.2024.3410112},
   number={4},
   journal={IEEE Security & Privacy},
   publisher={Institute of Electrical and Electronics Engineers (IEEE)},
   author={Blasingame, Zander W. and Liu, Chen},
   year={2024},
   month=jul, pages={103–114} }