MADation: Face Morphing Attack Detection with Foundation Models

This repository contains the official implementation of the paper "MADation: Face Morphing Attack Detection with Foundation Models", accepted at WACV2025 Workshops.

Overview

Despite significant advancements in face recognition algorithms, morphing attacks pose serious threats. MADation leverages foundation models, specifically the CLIP architecture, adapted with LoRA weights, to effectively detect morphing attacks. Our framework achieves competitive results against existing MAD solutions and is released to encourage reproducibility and further research.

Features:

• Foundation model adaptation with LoRA for the MAD task.
• Competitive performance on diverse datasets.
• Comprehensive training and evaluation pipelines for reproducibility.

Figure 1: Morphing attack generation and MADation's pipeline. The left side of the figure depicts a morphing sample and the two bona-fide identities that were morphed to generate it. Keep in mind that attackers commonly choose to morph faces with similar features for higher success. The right side represents MADation's pipeline, consisting of an adapted FM followed by a binary fully connected classification layer. The embedding space of the FM is adapted by fine-tuning the LoRA parameters and the classification layer is simultaneously trained to produce the MAD predictions. Better visualized in colour.

Figure 2: Integration of LoRA trainable weights (orange boxes) in a standard multi-head self-attention block, whose weights are kept frozen (blue boxes). In the proposed framework, MADation, the LoRA adaptation is limited to the q and v matrices, leaving k and o unaltered. Better visualized in colour.

How to replicate

• Create a virtual environment by using requirements.txt

conda create -n env_name python=3.9
pip install -r requirements.txt

• Adjust config file in /src/config.py
• Start training with ./train.sh

Pre-trained Models and Training Logs

All pre-trained models and their respective training logs are available here. To get access, please share your name, affiliation, and email in the request form.

Key Results

Method		Test data	EER (%)	APCER (%) @ BPCER (%)			BPCER (%) @ APCER (%)
				1.00	10.00	20.00	1.00	10.00	20.00
ViT-B	TI	FaceMorph	51.50	98.40	88.20	81.40	99.51	93.63	85.29
		MIPGAN I	36.40	99.80	81.10	65.30	86.76	55.88	46.57
		MIPGAN II	33.40	99.60	76.00	55.30	80.39	49.02	43.63
		OpenCV	47.15	99.90	83.74	74.90	98.04	81.37	70.10
		WebMorph	35.60	98.20	70.20	57.20	86.76	61.27	48.53
		MorDIFF	51.90	100.00	92.60	86.70	99.02	92.65	85.29
		Average	42.66	99.32	81.97	70.13	91.75	72.30	63.24
		Worst	51.90	100.00	92.60	86.70	99.51	93.63	85.29
	ViT-FS	FaceMorph	5.38	8.77	2.49	0.90	20.98	0.49	0.00
		MIPGAN I	32.87	85.66	61.35	47.41	100.00	49.02	49.02
		MIPGAN II	27.19	94.92	64.94	44.42	100.00	57.84	30.88
		OpenCV	16.30	50.40	26.42	14.27	100.00	56.31	34.47
		WebMorph	22.80	83.60	58.00	44.40	100.00	52.94	32.35
		MorDIFF	28.14	84.73	52.00	35.93	100.00	56.31	34.37
		Average	22.13	68.01	44.20	31.22	86.83	40.68	26.41
		Worst	32.87	94.92	64.94	47.41	100.00	57.84	49.02
	FE	FaceMorph	2.89	4.89	1.30	0.20	11.22	0.49	0.49
		MIPGAN I	26.00	83.27	55.68	36.06	77.94	50.98	32.84
		MIPGAN II	34.26	91.43	74.70	57.27	84.80	65.20	51.96
		OpenCV	14.88	39.98	20.34	9.21	61.27	18.63	10.78
		WebMorph	32.80	91.40	71.40	49.80	84.80	66.18	52.94
		MorDIFF	17.86	50.90	27.05	13.77	59.22	24.27	12.62
		Average	21.45	60.31	41.74	27.72	63.21	37.62	26.94
		Worst	34.26	91.43	74.70	57.27	84.80	66.18	52.94
	MADation (ours)	FaceMorph	0.00	0.00	0.00	0.00	0.00	0.00	0.00
		MIPGAN I	33.37	82.97	55.18	43.92	94.12	72.55	52.94
		MIPGAN II	22.21	79.98	34.66	24.30	84.80	47.55	26.47
		OpenCV	3.85	11.64	1.82	1.11	23.53	0.98	0.00
		WebMorph	10.80	60.00	11.40	5.00	51.47	11.76	4.41
		MorDIFF	1.10	1.60	0.00	0.00	1.94	0.00	0.00
		Average	11.89	39.36	17.18	12.39	42.64	22.14	13.97
		Worst	33.37	82.97	55.18	43.92	94.12	72.55	52.94
ViT-L	TI	FaceMorph	44.60	98.40	79.70	63.60	99.02	87.25	76.96
		MIPGAN I	18.90	71.80	32.20	17.80	69.61	33.82	18.14
		MIPGAN II	12.80	56.70	17.00	8.90	59.31	17.16	8.33
		OpenCV	35.47	96.24	77.54	63.11	96.08	73.53	55.39
		WebMorph	25.20	94.80	52.00	30.20	87.75	50.98	32.35
		MorDIFF	42.60	97.80	79.60	69.50	97.06	83.33	68.63
		Average	29.93	85.96	56.34	42.19	84.81	57.68	43.30
		Worst	44.60	98.40	79.70	69.50	99.02	87.25	76.96
	ViT-FS	FaceMorph	22.63	75.17	38.68	24.93	88.29	40.98	24.88
		MIPGAN I	23.80	79.08	42.93	25.50	91.18	46.57	28.43
		MIPGAN II	21.81	80.28	36.65	25.40	91.67	25.00	40.69
		OpenCV	30.47	84.72	59.92	44.23	94.12	60.29	42.16
		WebMorph	33.60	91.60	59.80	48.60	100.00	75.49	52.45
		MorDIFF	40.92	94.51	77.94	67.86	100.00	81.55	67.96
		Average	28.87	84.23	52.65	39.42	94.21	57.59	40.15
		Worst	40.92	94.51	77.94	67.86	100.00	81.55	67.96
	FE	FaceMorph	9.77	44.17	9.77	4.09	35.12	10.24	5.37
		MIPGAN I	23.51	88.84	55.28	31.37	71.57	40.69	27.45
		MIPGAN II	21.81	82.37	45.42	25.10	69.61	32.84	23.53
		OpenCV	15.89	55.77	25.40	10.83	48.53	22.06	12.75
		WebMorph	26.40	86.60	56.80	37.80	68.63	41.67	29.90
		MorDIFF	22.85	87.03	50.70	29.14	67.48	35.92	24.27
		Average	20.04	74.13	40.56	23.06	60.16	30.57	20.54
		Worst	26.40	88.84	56.80	37.80	71.57	41.67	29.90
	MADation (ours)	FaceMorph	0.40	0.40	0.00	0.00	0.49	0.00	0.00
		MIPGAN I	20.32	55.88	29.08	20.32	79.41	35.78	15.69
		MIPGAN II	9.06	19.42	9.06	5.58	100.00	5.39	0.98
		OpenCV	2.23	3.74	1.32	0.71	15.69	0.00	0.00
		WebMorph	20.40	47.60	20.40	20.40	82.35	37.25	13.24
		MorDIFF	19.26	48.40	24.45	19.26	84.47	34.95	15.53
		Average	11.94	29.24	14.05	11.04	60.40	18.90	7.57
		Worst	20.40	55.88	29.08	20.40	100.00	37.25	15.69

Please see the evaluation protocol in SYN-MAD-2022.

Citation

@InProceedings{Caldeira_2025_WACV,
    author    = {Caldeira, Eduarda and Ozgur, Guray and Chettaoui, Tahar and Ivanovska, Marija and Peer, Peter and Boutros, Fadi and Struc, Vitomir and Damer, Naser},
    title     = {MADation: Face Morphing Attack Detection with Foundation Models},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {February},
    year      = {2025},
    pages     = {1650-1660}
}

License

This project is licensed under the terms of the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.
Copyright (c) 2025 Fraunhofer Institute for Computer Graphics Research IGD Darmstadt
For more details, please take a look at the LICENSE file.