DOOMGAN: High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing

This repository contains the official PyTorch implementation and pre-trained models for DOOMGAN, a novel generative framework for creating high-fidelity morphed ocular images. This work was accepted at the IEEE International Joint Conference on Biometrics (IJCB) 2025 and is supported by the NSF award no. 2345561. NSF_Logo

Paper: DOOMGAN: High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing
Dataset: DOOMGAN-Ocular-Morphs
Interactive Demo: DOOMGAN Gradio Space
GitHub Code

DOOMGAN Teaser
Figure 1: Illustration of morphed ocular images generated by our proposed DOOMGAN. A morphed image combines images of two different identities such that the ensuing biometric template can match both contributing subjects, posing a serious security threat.

Abstract

Ocular biometrics in the visible spectrum have emerged as a prominent modality due to their high accuracy, resistance to spoofing, and non-invasive nature. Morphing attacks—where synthetic biometric traits are created by merging features from multiple individuals—pose a significant threat to the trustworthiness and integrity of biometric systems. While morphing attacks have been thoroughly investigated for the near-infrared iris and face biometrics, their implications for visible-spectrum ocular biometrics remain largely unexplored. Effectively simulating such attacks requires a sophisticated morphing generation model capable of handling the complexities of uncontrolled acquisition environments while preserving fine-grained ocular features. To address this gap, we introduce DOOMGAN, a novel framework that encompasses landmark-driven encoding, attention-guided generation, and dynamic weighting of multi-faceted losses for optimized convergence. We evaluated DOOMGAN against visible ocular recognition systems, achieving over 20% improvement in attack success rates over baseline models. Moreover, DOOMGAN shows a 20% boost in generating elliptical iris structures and a 30% boost in maintaining gaze consistency. Additionally, we provide the first comprehensive ocular morphing datasets to aid research in defending against such attacks.

Model Details

DOOMGAN is a novel Generative Adversarial Network (GAN) based framework designed specifically for synthesizing high-quality ocular morphs. It combines features from two different subjects to generate a synthetic biometric sample that can be matched to either individual.

The architecture uses several key components to achieve high-fidelity results:

Landmark-Driven Encoding: Utilizes precise ocular landmarks to guide the generation process, ensuring anatomical correctness.
Attention-Guided Generation: Employs attention mechanisms to focus on fine-grained ocular features.
Multi-Faceted Loss with Dynamic Weighting: A sophisticated loss function balances identity preservation, image quality, and anatomical consistency for stable and optimized training.

DOOMGAN Architecture
Figure 1: The overall architecture of the proposed DOOMGAN for visible spectrum ocular morph generation.

How to Use

The model can be run locally using the provided scripts. First, clone the repository and set up the environment as described in the README.md.

Command-Line Inference

To generate a morphed image from two source images, use the generate_morphs.py script. The script requires paths to two input images and will produce a morphed output.

# Clone the repository and install requirements first
# git clone https://github.com/Bharath-K3/DOOMGAN.git
# cd DOOMGAN
# pip install -r requirements.txt
# (Download pre-trained models as per the README instructions in GitHub)

Example command to generate a morph

Follow Code for detailed instructions.

python generate_morphs.py \
  --image1 "assets/1144_r_1.png" \
  --image2 "assets/1147_r_1.png" \
  --epoch 450 \
  --output "generated_morphs/1144_1147_morph.png"

Intended Use

This dataset is created to support the academic and research communities. The primary intended use is to train and evaluate Morph Attack Detection (MAD) algorithms to defend biometric systems against this vulnerability.

Citation

If you use this dataset or our model in your work, please cite our paper:

@article{krishnamurthy2025doomgan,
  title={DOOMGAN: High-Fidelity Dynamic Identity Obfuscation Ocular Generative Morphing},
  author={Krishnamurthy, Bharath and Rattani, Ajita},
  journal={arXiv preprint arXiv:2507.17158},
  year={2025}
}

Acknowledgements

This work is supported in part by the National Science Foundation (NSF), United States award no. 2345561.

Downloads last month: 1,674

BharathK333
/

DOOMGAN