FoundPAD: Foundation Models Reloaded for Face Presentation Attack Detection

This repository contains the official implementation of the paper "FoundPAD: Foundation Models Reloaded for Face Presentation Attack Detection", accepted at WACV2025 Workshops.

Overview

FoundPAD leverages foundation models with LoRA adaptation to tackle the challenges of face presentation attack detection (PAD). It achieves state-of-the-art generalization across unseen domains and performs well under various data availability scenarios, including synthetic data.

Features:

• Foundation model adaptation with LoRA for PAD tasks.
• Generalization benchmarks on diverse datasets.
• Training pipelines for limited and synthetic data scenarios.

Figure 1: Complete pipeline of FoundPAD. The proposed PAD model consists of an FM followed by a binary fully-connected classification layer. During training, the FM's feature space is adapted due to the training of the LoRA weights, while the classification layer is simultaneously trained to predict the PAD labels. It is 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, FoundPAD, 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 - ViT-B/16

				epoch = 40
FoundPAD	Train		Test	AUC	HTER
ViT-B/16	I	Table 5	C	90,43%	17,00%	Table 5
			M	89,26%	18,57%	Avg. AUC
			O	93,74%	13,38%	90,75%
	C		I	91,50%	16,40%	Avg. HTER
			M	82,14%	24,52%	16,81%
			O	92,38%	15,14%	Std. Dev. HTER
	M		C	88,03%	20,00%	5,03%
			I	91,05%	17,10%
			O	88,63%	19,41%	Table 2
	O		C	97,11%	7,89%	Avg. AUC
			M	87,98%	23,33%	95,52%
			I	96,77%	8,95%	Avg. HTER
	M&I	Table 4	C	93,97%	13,22%	10,62%
			O	96,69%	9,31%	Std. Dev. HTER
	O&C&M	Table 3	CA	90,98%	15,65%	7,29%
		Table 2	I	95,80%	10,45%
	O&C&I		M	89,88%	20,95%	Table 6
	O&M&I		C	98,08%	4,89%	Avg. AUC
	I&C&M		O	98,31%	6,19%	78,39%
	SynthASpoof	Table 6	M	66,18%	47,14%	Avg. HTER
			C	83,03%	27,33%	30,94%
			I	90,79%	16,15%
			O	73,56%	33,12%
	Total Average			89,84%	17,66%

				epoch = 40
ViT-FE	Train		Test	AUC	HTER
ViT-B/16	I	Table 5	C	67,57%	38,56%	Table 5
			M	68,59%	35,71%	Avg. AUC
			O	64,66%	40,52%	74,86%
	C		I	73,46%	32,95%	Avg. HTER
			M	71,28%	35,24%	31,89%
			O	76,96%	30,63%	Std. Dev. HTER
	M		C	82,91%	25,33%	6,29%
			I	75,47%	30,45%
			O	77,66%	28,86%	Table 2
	O		C	91,07%	16,89%	Avg. AUC
			M	76,80%	32,86%	80,32%
			I	71,88%	34,65%	Avg. HTER
	M&I	Table 4	C	79,94%	27,22%	28,14%
			O	72,70%	33,57%	Std. Dev. HTER
	O&C&M	Table 3	CA	84,24%	23,66%	7,32%
		Table 2	I	72,71%	36,10%
	O&C&I		M	77,50%	30,71%	Table 6
	O&M&I		C	90,33%	18,67%	Avg. AUC
	I&C&M		O	80,74%	27,07%	72,21%
	SynthASpoof	Table 6	M	59,27%	47,14%	Avg. HTER
			C	78,81%	28,11%	33,76%
			I	87,08%	19,50%
			O	63,66%	40,28%
	Total Average			75,88%	31,07%

				epoch = 40
ViT-FS	Train		Test	AUC	HTER
ViT-B/16	I	Table 5	C	82,05%	24,33%	Table 5
			M	89,29%	22,38%	Avg. AUC
			O	89,10%	18,43%	91,33%
	C		I	83,51%	26,05%	Avg. HTER
			M	97,14%	8,33%	15,88%
			O	90,91%	17,79%	Std. Dev. HTER
	M		C	99,25%	4,00%	7,07%
			I	92,05%	15,05%
			O	98,10%	6,68%	Table 2
	O		C	95,33%	11,56%	Avg. AUC
			M	92,18%	15,48%	95,99%
			I	87,00%	20,45%	Avg. HTER
	M&I	Table 4	C	92,97%	14,00%	10,37%
			O	97,88%	7,11%	Std. Dev. HTER
	O&C&M	Table 3	CA	89,07%	16,01%	3,87%
		Table 2	I	93,59%	14,90%
	O&C&I		M	96,09%	11,19%	Table 6
	O&M&I		C	95,67%	9,89%	Avg. AUC
	I&C&M		O	98,60%	5,52%	63,91%
	SynthASpoof	Table 6	M	58,61%	50,24%	Avg. HTER
			C	59,46%	44,44%	41,40%
			I	81,48%	24,40%
			O	56,08%	46,53%
	Total Average			87,63%	18,90%

Key Results - ViT-L/14

				epoch = 40
FoundPAD	Train		Test	AUC	HTER
ViT-L/14	I	Table 5	C	96,14%	10,22%	Table 5
			M	89,52%	19,29%	Avg. AUC
			O	90,79%	16,94%	92,38%
	C		I	93,12%	14,05%	Avg. HTER
			M	88,48%	21,43%	15,49%
			O	95,19%	11,00%	Std. Dev. HTER
	M		C	94,22%	12,00%	5,07%
			I	92,62%	14,55%
			O	87,37%	20,93%	Table 2
	O		C	97,64%	7,22%	Avg. AUC
			M	89,01%	23,81%	96,60%
			I	94,48%	14,40%	Avg. HTER
	M&I	Table 4	C	99,22%	4,67%	9,67%
			O	95,58%	10,23%	Std. Dev. HTER
	O&C&M	Table 3	CA	59,66%	42,99%	5,17%
		Table 2	I	96,07%	9,90%
	O&C&I		M	93,18%	16,90%	Table 6
	O&M&I		C	98,72%	6,00%	Avg. AUC
	I&C&M		O	98,41%	5,87%	85,01%
	SynthASpoof	Table 6	M	69,76%	45,71%	Avg. HTER
			C	96,03%	9,89%	23,51%
			I	98,58%	6,40%
			O	75,69%	32,05%
	Total Average			90,85%	16,37%

				epoch = 40
ViT-FE	Train		Test	AUC	HTER
ViT-L/14	I	Table 5	C	87,85%	19,78%	Table 5
			M	82,10%	26,90%	Avg. AUC
			O	74,33%	31,83%	82,91%
	C		I	81,37%	25,70%	Avg. HTER
			M	77,99%	30,24%	24,43%
			O	81,89%	25,06%	Std. Dev. HTER
	M		C	92,41%	15,44%	6,21%
			I	88,90%	19,00%
			O	79,36%	28,37%	Table 2
	O		C	93,24%	13,78%	Avg. AUC
			M	73,31%	32,62%	88,52%
			I	82,18%	24,40%	Avg. HTER
	M&I	Table 4	C	94,57%	11,33%	18,76%
			O	81,31%	26,19%	Std. Dev. HTER
	O&C&M	Table 3	CA	58,25%	43,86%	6,51%
		Table 2	I	86,27%	22,05%
	O&C&I		M	86,87%	21,67%	Table 6
	O&M&I		C	96,10%	9,00%	Avg. AUC
	I&C&M		O	84,85%	22,32%	78,41%
	SynthASpoof	Table 6	M	55,76%	52,62%	Avg. HTER
			C	92,82%	13,89%	29,15%
			I	87,94%	20,50%
			O	77,13%	29,58%
	Total Average			82,47%	24,61%

				epoch = 40		batch size = 412
ViT-FS	Train		Test	AUC	HTER
ViT-L/14	I	Table 5	C	83,28%	25,00%	Table 5
			M	87,37%	20,48%	Avg. AUC
			O	84,11%	23,62%	87,42%
	C		I	85,94%	22,60%	Avg. HTER
			M	97,74%	5,71%	19,84%
			O	68,58%	37,07%	Std. Dev. HTER
	M		C	96,39%	7,89%	8,69%
			I	94,67%	15,00%
			O	86,00%	22,93%	Table 2
	O		C	86,42%	20,33%	Avg. AUC
			M	95,72%	11,43%	84,49%
			I	82,87%	26,05%	Avg. HTER
	M&I	Table 4	C	85,66%	25,22%	23,04%
			O	96,41%	9,07%	Std. Dev. HTER
	O&C&M	Table 3	CA	52,29%	48,23%	11,74%
		Table 2	I	87,29%	21,55%
	O&C&I		M	98,12%	8,10%	Table 6
	O&M&I		C	82,97%	26,11%	Avg. AUC
	I&C&M		O	69,57%	36,40%	59,48%
	SynthASpoof	Table 6	M	55,94%	50,00%	Avg. HTER
			C	58,14%	47,11%	45,19%
			I	73,20%	33,60%
			O	50,63%	50,04%
	Total Average			80,84%	25,81%

Please see the data preparation, and the evaluation protocol in CF-MAD.

Citation

@inproceedings{ozgur2025foundpad,
  title={FoundPAD: Foundation Models Reloaded for Face Presentation Attack Detection},
  author={Ozgur, Guray and Caldeira, Eduarda and Chettaoui, Tahar and Boutros, Fadi and Ramachandra, Raghavendra and Damer, Naser},
  booktitle={WACV AI4MFDD Workshop},
  year={2025},
  institution={Fraunhofer IGD, TU Darmstadt, Norwegian University of Science and Technology},
  url={https://arxiv.org/abs/2501.02892}
}

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.