DINOFAS: Bonafide-Centric Representation Learning for Face Anti-Spoofing
Authors: Shakeel Muhammad Ibrahim, Muhammad Sohail Ibrahim, Muhammad Shahid Jabbar, Shujaat Khan
Organizations: Geomexsoft Co. Ltd., Chuncheon-si, Republic of Korea · Interdisciplinary Research Center for Intelligent Secure Systems (IRC-ISS), King Fahd University of Petroleum & Minerals, Dhahran, 31261, Saudi Arabia · SDAIA-KFUPM Joint Research Center for Artificial Intelligence (JRC-AI), King Fahd University of Petroleum & Minerals, Dhahran, 31261, Saudi Arabia · Department of Computer Engineering, College of Computing and Mathematics, King Fahd University of Petroleum & Minerals, Dhahran, 31261, Saudi Arabia
Face anti-spoofing (FAS) models often learn attack- and dataset-specific cues that degrade performance under unseen acquisition conditions and presentation attacks. We present DINOFAS, a DINOv3-based framework that formulates FAS as bonafide-centric representation learning. The proposed framework combines parameter-efficient DINOv3 adaptation using low-rank adaptation (LoRA), an attention style-content disentanglement (ASCD) module for reducing domain-specific appearance variations, and a dual-branch architecture that separates prototype-based liveness learning from attack-semantic supervision. The model is jointly optimized using a bonafide-centric prototype objective, an attack-semantic classification loss applied to detached features, and orthogonality regularization. DINOFAS achieves an average HTER of 9.09% under the standard OCIM leave-one-domain-out protocol, where training is performed on three source datasets and evaluation is conducted on the held-out unseen domain. We further evaluate synthetic-to-real generalization by training exclusively on SynthASpoof and testing on four unseen real-world datasets, where DINOFAS achieves the lowest average HTER of 21.41% and the highest average AUC of 86.14% among the compared methods. Evaluations on OULU-NPU and SiW-Mv2 additionally demonstrate robustness to variations in acquisition conditions, devices, and unseen presentation attacks.
Figures & tables
Fig. 1: Conceptual overview of DINOFAS. The proposed framework replaces conventional binary classification with bonafide-centric representation learning, where a prototype-based liveness branch and a detached attack-semantic branch jointly learn robust and complementary representations.
Fig. 2: Modified DINOv3 transformer with LoRA adaptation and the proposed attention style-content disentanglement (ASCD) module. The lower panel illustrates the ASCD pipeline, comprising gated style-content decomposition, label-aware style borrowing, and adaptive feature recombination.
Fig. 3: Task-specific branches of the proposed framework. The upper branch performs bonafide-centric prototype learning for binary liveness estimation, while the lower branch learns attack semantics through a detached classifier to prevent attack-specific supervision from contaminating the shared representation.
Method
M
C
I
O
Average
HTER(%) ↓
AUC(%) ↑
HTER(%) ↓
AUC(%) ↑
HTER(%) ↓
AUC(%) ↑
HTER(%) ↓
AUC(%) ↑
HTER(%) ↓
AUC(%) ↑
ResNet [ 43 ]
25.48
79.54
39.22
62.00
8.90
96.96
34.23
71.48
26.96
77.50
PixBiS [ 43 ]
38.33
63.87
38.44
64.79
7.50
96.88
35.77
63.50
30.74
72.26
ViT-SIDE B [ 44 ]
36.67
69.78
33.33
75.21
9.80
96.67
13.26
94.04
23.27
83.93
SynFace Co-Former A [ 44 ]
18.57
90.76
41.11
64.49
16.30
92.31
21.67
86.44
24.41
83.50
SynFace Co-Former B [ 44 ]
16.67
91.61
40.00
63.05
18.80
88.20
25.35
82.02
25.21
81.22
TABLE I: Single-source cross-dataset evaluation when training on the SynthASpoof dataset. Lower HTER and higher AUC indicate better performance. The best and second-best results are highlighted in bold and underlined , respectively.
Method
OCI → M
OMI → C
OCM → I
ICM → O
Avg. HTER(%) ↓
Avg. AUC(%) ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
NAS-FAS [ 20 ]
19.53
88.63
16.54
90.18
14.51
93.84
13.80
93.43
16.10
91.52
LMFD [ 17 ]
10.48
94.55
12.50
94.17
18.49
84.72
12.41
94.95
13.47
92.10
MADDG [ 26 ]
17.69
88.06
24.50
84.51
22.19
84.99
27.89
80.02
23.07
84.40
RFM [ 21 ]
17.30
90.48
13.89
93.98
20.27
88.16
16.45
91.16
16.98
90.95
SSDG-R [ 27 ]
7.38
97.17
10.44
95.94
11.71
96.59
15.61
91.54
11.29
95.31
TABLE II: Cross-domain Face Presentation Attack Detection results on the OCIM benchmark. Lower HTER and higher AUC indicate better performance. The best results are shown in bold , while the second-best results are underlined .
Fig. 4: t-SNE visualization of DINOFAS representations under the four OCIM cross-domain protocols.
P
Method
ACER
APCER
BPCER
1
LMFD-PAD [ 17 ]
1.50
1.40
1.60
PatchSwap [ 14 ]
0.60
0.40
0.80
MAX (DINO, CLIP) [ 39 ]
2.64
2.50
2.77
DINOFAS (Proposed)
1.67
1.67
1.67
2
LMFD-PAD [ 17 ]
2.00
3.10
0.80
PatchSwap [ 14 ]
1.80
2.50
1.10
TABLE III: Comparison of Face Presentation Attack Detection performance on the OULU-NPU dataset under the four standard evaluation protocols. Lower ACER, APCER, and BPCER indicate better performance. The best and second-best results for each protocol are highlighted in bold and underlined , respectively.
Attack
MD-FAS
Meta
DINO (ViT-B-14)
DINOFAS
[ 50 ]
[ 51 ]
[ 39 ]
(Proposed)
FunE.
29.50
14.80
8.52
19.62
PEye
2.70
–
0.19
0.38
PMouth
1.10
10.30
0.19
0.19
PaperG.
11.90
15.20
0.19
5.44
Ob.
1.30
45.30
18.16
7.47
TABLE IV: HTER (%) on the SiW-Mv2 leave-one-out protocol. Lower values indicate better performance. The best and second result for each attack type are highlighted in bold and underlined , respectively.
Method
OCI → M
OMI → C
OCM → I
ICM → O
Avg. HTER ↓
Avg. AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
DINOv2 (ViT-B/16) (Frozen)
22.50
83.22
20.55
87.55
16.75
86.95
30.72
74.83
22.63
83.14
DINOv2 (ViT-B/16) + LoRA
12.50
95.77
11.85
95.26
13.37
91.40
14.68
91.44
13.10
93.47
DINOv3 (ViT-B/16) (Frozen)
23.75
81.02
23.14
83.62
25.62
80.38
25.97
80.53
24.62
81.39
DINOv3 (ViT-B/16) + LoRA
13.33
94.22
11.85
95.15
8.87
94.71
9.34
95.27
10.85
94.84
DINOFAS (ViT-S/16)
17.08
90.41
12.40
94.74
20.06
83.03
15.52
91.72
16.26
89.97
TABLE V: Effect of foundation-model backbone and LoRA adaptation on the OCIM cross-dataset benchmark. The DINOv2/DINOv3 baseline variants use a conventional linear binary head with BCE, whereas DINOFAS rows use the complete proposed framework. The best and second-best results are highlighted in bold and underlined , respectively.
Cfg.
Lbin
Lattack
Lorth
StopGrad
Avg. HTER (%) ↓
A
✓
×
×
–
14.16
B
✓
✓
×
×
13.95
C
✓
×
✓
✓
14.03
D
✓
✓
×
✓
13.45
E
✓
✓
✓
×
13.30
Proposed
✓
✓
✓
✓
9.09
TABLE VI: Ablation study of the proposed learning strategy on the OCIM cross-dataset benchmark. ASCD is enabled in all configurations, while Lbin , Lattack , Lorth , and the stop-gradient mechanism are varied to evaluate their contributions.
Component
Configuration
Avg. HTER (%) ↓
Avg. AUC (%) ↑
Style Borrowing
Class-agnostic
13.36
92.00
Class-aware (Proposed)
9.09
94.40
Gating Mechanism
Without gating
12.14
93.52
With gating (Proposed)
9.09
94.40
ASCD Placement
None
13.32
91.87
Early (Blocks 1,2)
13.67
91.84
TABLE VII: Ablation study of the proposed ASCD module on the OCIM cross-dataset benchmark. Lower HTER and higher AUC indicate better performance.
Face Anti-Spoofing (FAS) remains challenging due to the requirement for robust domain generalization across unseen environments. While recent trends leverage Vision-Language Models (VLMs) for semantic supervision, these multimodal approaches often demand prohibitive computational resources and exhibit high inference latency. Furthermore, their efficacy is inherently limited by the quality of the underlying visual features. This paper revisits the potential of vision-only foundation models to establish a highly efficient and robust baseline for FAS. We conduct a systematic benchmarking of 15 pre-trained models, such as supervised CNNs, supervised ViTs, and self-supervised ViTs, under severe cross-domain scenarios including the MICO and Limited Source Domains (LSD) protocols. Our comprehensive analysis reveals that self-supervised vision models, particularly DINOv2 with Registers, significantly suppress attention artifacts and capture critical, fine-grained spoofing cues. Combined with Face Anti-Spoofing Data Augmentation (FAS-Aug), Patch-wise Data Augmentation (PDA) and Attention-weighted Patch Loss (APL), our proposed vision-only baseline achieves state-of-the-art performance in the MICO protocol. This baseline outperforms existing methods under the data-constrained LSD protocol while maintaining superior computational efficiency. This work provides a definitive vision-only baseline for FAS, demonstrating that optimized self-supervised vision transformers can serve as a backbone for both vision-only and future multimodal FAS systems. The project page is available at: https://gsisaoki.github.io/FAS-VFMbenchmark-CVPRW2026/ .
Mika Feng, Pierre Gallin-Martel, Koichi Ito +1
Graduate School of Information Sciences, Tohoku University, Japan
Face anti-spoofing (FAS) is increasingly expected to provide not only bona fide/spoof decisions, but also attack semantics and image-grounded evidence for human inspection. Existing discriminative FAS models remain largely label-centric, while recent MLLM-based methods offer structured outputs but still rely mainly on supervised fine-tuning, often producing template-like rationales and weak optimization for difficult attacks. We propose FAS-R1, a two-stage reasoning-oriented MLLM framework for unified FAS prediction, covering authenticity classification, attack-type recognition and spoof-region localization. FAS-R1 first uses FAS-R1-23K, a high-quality long-CoT dataset, for cold-start supervised fine-tuning, and then performs FAS-specific GRPO post-training. Degradation-Simulated Augmentation (DSA) encourages stable spoof-cue reasoning across visual-quality shifts, while Difficulty-Aware GRPO (DA-GRPO) mitigates easy-sample dominance that may leave difficult task--attack groups under-optimized, especially for subtle or ambiguous attacks such as makeup and mask attacks. The main 3B FAS-R1 model achieves 98.75% authenticity accuracy, 93.33% attack-type accuracy, and 96.30/94.73% AP@40/AP@50 in-domain. It also outperforms the compared systems in cross-domain authenticity generalization and answer-and-rationale quality. Experiments with different base models further show favorable scaling behavior. The code will be released soon.
Hongyang Wang, Yichen Shi, Hongrui Li +3
Shijiazhuang Tiedao University · Fudan University · Shanghai Jiao Tong University +2
With the increasing diversity of spoofing attacks, there is a growing demand for unified Face Anti-Spoofing (FAS) models capable of detecting both physical and digital threats. While existing Vision-Language Models (VLMs) demonstrate high generalization in this context, they heavily rely on complex multimodal fusion and external text encoders. In this paper, we propose DINO-VPT, a lightweight, vision-only framework leveraging hierarchical visual prompt tuning. By dynamically injecting prompts conditioned on input features via a Prompt Routing Network (PRN), our method effectively disentangles diverse spoofing artifacts without requiring multimodal fusion. Evaluations on the UniAttackData benchmark demonstrate that DINO-VPT achieves higher accuracy than state-of-the-art VLM-based methods. Our results indicate that a properly structured vision-only architecture can achieve state-of-the-art performance in unified FAS without the need for multimodal supervision.
Pierre Gallin-Martel, Mika Feng, Koichi Ito +1
Graduate School of Information Sciences, Tohoku University, Japan