Multimodal Models Meet Presentation Attack Detection on ID Documents

Marina Villanueva; Juan M. Espin; Juan E. Tapia

cs.CVMar 31, 2026

Multimodal Models Meet Presentation Attack Detection on ID Documents

Authors: Marina Villanueva, Juan M. Espin, Juan E. Tapia

Organizations: Facephi Alicante, Spain · da/sec-Biometrics and Security Research Group Hochschule Darmstadt, Germany

Abstract

The integration of multimodal models into Presentation Attack Detection (PAD) for ID Documents represents a significant advancement in biometric security. Traditional PAD systems rely solely on visual features, which often fail to detect sophisticated spoofing attacks. This study explores the combination of visual and textual modalities by utilizing pre-trained multimodal models, such as Paligemma, Llava, and Qwen, to enhance the detection of presentation attacks on ID Documents. This approach merges deep visual embeddings with contextual metadata (e.g., document type, issuer, and date). However, experimental results indicate that these models struggle to accurately detect PAD on ID Documents.

Figures & tables

Fig. 2 : Scoring process based on model output logits applied to each model. Example for prompts with answer “Yes” or “No”.

Panel of Figure 3 — Fig. 3 : Attack examples for ID Cards and Passports.

Fig. 4 : Results of Exp. 1. Comparison of the use of technical terminology among models.

Fig. 5 : Visual keypoints of a spoof PVC ID Card.

Model	OCR	Security features	Material	Print quality	Total
PaliGemma	3	0	0	0	3 /12
LLaVA	3	2	1	H	6 /12
Qwen	3	H	H	0	3 /12

TABLE I : Results of Exp. 2. Visual Scores for PVC attack. Scores: 0 (Blindness), 1 (Mention), 2 (Guided detection), 3 (Spontaneous detection), H (Hallucination).

Prompt	Paligemma2 (%)		Llava1.6 (%)		QWEN2.5 (%)
Prompt	APCER	BPCER	APCER	BPCER	APCER	BPCER
Single_turn_1	100.0	0.0	0.0	100.0	98.0	0.0
Single_turn_2	100.0	0.0	4.0	73.0	94.0	4.0
Single_turn_3	100.0	0.0	0.0	100.0	58.0	10.0
Multi-turn 1	N/A	N/A	2.0	98.0	0.0	100.0
Multi-turn 2	N/A	N/A	0.0	100.0	0.0	100.0

TABLE II : Results of Exp. 3. N/A: Represents a null value because PaliGemma2 is not a multi-turn Chatbot. The results represents de [email protected] and [email protected].

Prompt	Paligemma2 (%)			Llava1.6 (%)			QWEN2.5 (%)
Prompt	EER	B10	B20	EER	B10	B20	EER	B10	B20
Single_turn_1	66.0	95.0	98.0	46.5	79.0	80.0	42.0	73.0	82.0
Single_turn_2	44.0	90.0	98.0	31.0	53.0	65.0	40.5	75.0	86.0
Single_turn_3	53.0	93.0	97.0	35.5	62.0	74.0	35.0	85.0	96.0
Multi turn_1	N/A	N/A	N/A	64.0	93.0	97.0	49.5	87.0	90.0
Multi turn_2	N/A	N/A	N/A	59.0	99.0	100.0	43.0	82.0	87.0

TABLE III : Results of Exp. 3 evaluated at fixed BPCER operating points (BPCER 10 and BPCER 20 ). N/A: Represents a null value because PaliGemma2 is not a multi-turn Chatbot. B10 and B20 represent the BPCER 10 and BPCER 20 respectively.

Panel of Figure 9 — Fig. 6 : DET performance achieved for the best prompts on the multimodal models (Left to Right): Single_turn_3 , Simple_8 , and With_background_3 .

Biometric Identification

Biometric identification matches or verifies people and animals using distinctive physical, physiological, or behavioral signals.

See all 129 papers

From Vision to Text: A Compact Multimodal Approach for Robust, Cross-Domain Presentation Attack Detection on ID Cards
Jun 5, 2026Qingwen Zeng, Juan E. Tapia, Sneha Das +1Multimodal Robustness Domain Generalization
Technical University of Denmark (DTU), Denmark · da/sec-Biometrics and Security Research Group, Hochschule Darmstadt, Germany
Towards Zero-Shot Domain Generalization for ID Cards Presentation Attack Detection
Aug 17, 2026Mario Nieto-Hidalgo, Juan M. Espin, Juan E. TapiaDomain Generalization Few-Shot Learning
Facephi, R&D Department, Spain. · da/sec-Biometrics and Internet Security Research Group. Hochschule Darmstadt, Germany.
Foundation and Multimodal Large Language Models for Face Presentation and Morph Attack Detection
Aug 30, 2026Hatef Otroshi Shahreza, Asif Hussain Khan, Peter Lorenz +2Image Forgery Detection Vision Foundation Model Adaptation
Idiap Research Institute, Switzerland · Université de Lausanne, Switzerland

Jun 5, 2026J/K move · Enter open · S save

From Vision to Text: A Compact Multimodal Approach for Robust, Cross-Domain Presentation Attack Detection on ID Cards

Qingwen Zeng, Juan E. Tapia, Sneha Das, Christoph Busch

Technical University of Denmark (DTU), Denmark · da/sec-Biometrics and Security Research Group, Hochschule Darmstadt, Germany

Multimodal Robustness Domain Generalization Multimodal Model Evaluation Synthetic Data Evaluation

Open paper AskPDF arXiv

Cross-domain shifts challenge Presentation Attack Detection (PAD) on ID Cards, given the restricted data available due to privacy concerns. This work proposes a compact multimodal model, based on new generative and discriminative blocks, which combines visual and textual data for PAD on genuine and synthetic ID images. While multimodal models exhibit strong generalisation after supervised fine-tuning, they fail in zero-shot settings. Our findings underscore that model capacity and real-world data are essential for reliable PAD, while existing synthetic datasets may not reflect real-world challenges. We argue for a re-evaluation of synthetic data as a benchmark and emphasise the need for more realistic, diverse datasets to advance PAD research.

Abstract

Biometric Identification

Explore similar work

From Vision to Text: A Compact Multimodal Approach for Robust, Cross-Domain Presentation Attack Detection on ID Cards

Towards Zero-Shot Domain Generalization for ID Cards Presentation Attack Detection

Foundation and Multimodal Large Language Models for Face Presentation and Morph Attack Detection