Context-Predictive Low-Rank Adaptation for Face Presentation Attack Detection
Authors: Aminu Yusuf, Shakeel Muhammad Ibrahim, Sulaimon Oyeniyi Adebayo, Muhammad Sohail Ibrahim, Muhammad Shahid Jabbar, Shujaat Khan
Organizations: Department of Computer Engineering, College of Computing and Mathematics, King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia. · Geomexsoft Co. Ltd., Chuncheon-si, Republic of Korea. · Interdisciplinary Research Center for Intelligent Secure Systems, King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia. · SDAIA-KFUPM Joint Research Center for Artificial Intelligence, King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia.
Face presentation attack detection requires representations that distinguish physical spoofing cues across changes in acquisition devices, subjects, and attack media. Large visual encoders offer transferable representations, but conventional binary fine-tuning supervises their final decisions without explicitly constraining how local facial evidence is represented under partial observation. We present CP-JEPA, a context-predictive adaptation framework that combines low-rank tuning of a pretrained CLIP image encoder with a joint-embedding predictive objective. A student predicts complete-image teacher features at masked spatial locations, while unmasked classification, masked classification, and prediction consistency retain the connection to the live/spoof task. An exponential-moving-average teacher supplies evolving latent targets. The auxiliary components are used only during training; inference requires only the LoRA-adapted CLIP visual encoder and binary classifier. {We evaluate CP-JEPA under single-source and multi-source OCIM transfer, additional CelebA-Spoof training, transfer to CelebA-Spoof, two-source training, and unseen-attack evaluation on SiW-Mv2.} On the standard OCIM benchmark, CP-JEPA achieves 3.74% mean HTER and 98.73% mean AUC. {Single-source evaluation achieves 11.36% mean HTER across twelve source--target combinations.} With CelebA-Spoof added to the three OCIM training sources, CP-JEPA reduces mean HTER from 3.45% to 2.35% and increases mean AUC from 98.68% to 99.32%. On SiW-Mv2, it achieves 3.36% mean HTER across the held-out attack categories. Controlled loss ablations examine the contributions of latent prediction, masked classification, and prediction consistency. The results show that context-predictive adaptation can improve cross-dataset PAD performance, although the gains vary with the training sources and evaluation setting.
Figures & tables
Fig. 1: Overview of CP-JEPA training and inference. The student learns from masked spatial predictions and binary classification, using an EMA teacher to provide latent targets. The teacher and predictor are used only during training; inference uses the adapted CLIP encoder and its native binary classifier.
Setting
Configuration
Architecture
Visual encoder
CLIP ViT-B/16 (frozen base)
LoRA
Q/V, all 12 blocks; r=8 , α=8
Spatial projector
Linear, 512→192 , followed by LayerNorm
Predictor
2 layers; 4 heads; hidden 192; FFN 768
Patch masking
Uniform random; 30% of 196 patches
TABLE I: Main CP-JEPA architecture and training configuration.
Method
OCI → M
OMI → C
OCM → I
ICM → O
Avg. HTER(%) ↓
Avg. AUC(%) ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
CIFAS [ 33 ] ( 2022a )
5.95
96.32
10.66
95.30
8.50
97.24
13.17
93.44
9.57
95.58
IADG [ 4 ] ( 2023 )
5.41
98.19
8.70
96.44
10.62
94.50
8.86
97.14
8.40
96.57
DRDN [ 10 ] ( 2023 )
10.00
96.20
20.90
83.80
11.40
94.50
15.60
90.90
14.50
91.30
DiffFAS-R [ 34 ] ( 2024 )
5.90
98.10
7.32
97.40
5.66
98.69
12.38
94.33
7.82
97.13
DiffFAS-V [ 34 ] ( 2024 )
2.86
98.41
10.11
96.32
6.36
97.89
8.11
97.27
6.86
97.47
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 .
Method
Train C
Train I
Train M
Train O
Avg.
I
M
O
C
M
O
C
I
O
C
I
M
DRDN-ii [ 10 ] ( 2023 )
27.10
17.10
24.50
39.10
23.10
37.10
27.80
13.00
21.70
20.10
19.00
19.90
24.13
CF-PAD [ 39 ] ( 2024 )
24.80
17.14
19.43
34.00
24.76
31.70
14.44
15.90
25.34
20.33
21.50
15.00
22.03
FoundPAD (ViT-B) [ 8 ] ( 2025 )
16.40
24.52
15.14
17.00
18.57
13.38
20.00
17.10
19.41
7.89
8.95
23.33
16.81
FoundPAD (ViT-L) [ 8 ] ( 2025 )
14.05
21.43
11.00
10.22
19.29
16.94
12.00
14.55
20.93
7.22
14.40
23.81
15.49
CP-JEPA (Ours)
13.13
20.00
20.10
4.63
8.33
10.76
12.96
6.88
15.07
8.15
5.00
11.25
11.36
TABLE III: Single-source cross-dataset HTER (%) on OCIM. Each method is trained on one OCIM source dataset and evaluated separately on the remaining three. Lower HTER is better. The best and second-best results are shown in bold and underlined , respectively.
Model
ICM → O
OMI → C
OCM → I
OCI → M
Average
FLIP-MCL [ 15 ] ( 2023 )
2.31 / 99.63
0.54 / 99.98
4.25/ 99.07
4.95/98.11
3.01/99.20
FSFM ViT-B [ 22 ] ( 2025 )
7.68/97.11
3.16/99.41
4.63/99.03
3.78/ 99.15
4.81/98.68
CCPE [ 37 ] ( 2026 )
3.64/ 99.67
1.30/ 99.98
4.15 / 99.31
2.86 / 99.24
2.99/ 99.55
CP-JEPA (Ours)
2.43 /99.54
0.74 / 99.73
2.88 /98.97
3.33 /99.04
2.35 / 99.32
TABLE IV: Performance on the four OCIM leave-one-domain-out protocols with CelebA-Spoof as an additional FacePAD training source. Results are reported as HTER ↓ /AUC ↑ .
Fig. 2: Cross-domain HTER, AUC, APCER, and BPCER for CE adaptation and CP-JEPA with CelebA-Spoof added to the source training data.
Model
O
C
I
M
Avg.
HTER (%) ↓
CE
3.65
3.33
4.75
4.17
3.97
Full
3.89
3.33
5.25
2.50
3.74
− Pred
3.19
5.00
5.25
2.50
3.99
− Mask
4.31
4.81
5.25
5.83
5.05
− Cons
4.31
3.52
5.38
6.67
4.97
TABLE V: Ablation study of the auxiliary objectives in CP-JEPA across the four OCIM protocols.
Attack
MD-FAS (SRENet)
DINO (ViT-B-14)
CLIP-B16
CLIP-B32
CLIP-L14
DINO-S14
DINO-L14
CP-JEPA
[ 40 ] ( 2022 )
[ 41 ] ( 2025 )
[ 41 ] ( 2025 )
[ 41 ] ( 2025 )
[ 41 ] ( 2025 )
[ 41 ] ( 2025 )
[ 41 ] ( 2025 )
(Ours)
FunE.
29.50
8.52
13.46
19.24
9.57
18.76
21.62
11.57
PEye
2.70
0.19
0.19
0.19
0.19
0.19
0.19
0.19
PMouth
1.10
0.19
0.39
3.65
0.19
0.19
0.19
0.19
PaperG.
11.90
0.19
1.62
2.47
0.19
1.04
0.19
2.30
Ob.
1.30
18.16
8.60
10.14
12.80
9.76
7.68
12.49
TABLE VI: HTER (%) on the SiW-Mv2 leave-one-out protocol. Lower values indicate better performance. The best and second-best results for each attack type are highlighted in bold and underlined , respectively.
Model
HTER ↓
AUC ↑
Panwar et al. [ 42 ] ( 2021 )
26.10
80.00
CIFAS [ 33 ] ( 2022a )
24.60
83.20
SourceOnly [ 43 ] ( 2022b )
29.70
77.50
SDA-FAS [ 43 ] ( 2022b )
18.90
90.90
SourceOnly++ [ 44 ] ( 2024a )
23.87
83.67
SDA-FAS++ [ 44 ] ( 2024a )
13.74
93.70
TABLE VII: Cross-dataset evaluation on CelebA-Spoof using O, C, and M as training sources.
Model
M+I → C
M+I → O
Average
HTER ↓
AUC ↑
HTER ↓
AUC ↑
HTER ↓
AUC ↑
SSDG-M [ 2 ] ( 2020 )
31.89
71.29
36.01
66.88
33.95
69.09
D 2 AM [ 45 ] ( 2021 )
32.65
72.04
27.70
75.36
30.18
73.70
SSAN-R [ 46 ] ( 2022 )
25.56
83.89
24.44
82.86
25.00
83.38
HFN+MP [ 47 ] ( 2022 )
30.89
72.48
20.94
86.71
25.92
79.60
MADG [ 48 ] ( 2023 )
23.33
82.32
26.29
81.50
24.81
81.91
TABLE VIII: Cross-dataset evaluation with M and I as training sources and C and O as held-out targets.
Face Presentation Attack Detection identifies spoofing attempts, such as printed, replayed, masked, or digitally altered faces, targeting face recognition systems.
Department of Computer Engineering, College of Computing and Mathematics, King Fahd University of Petroleum & Minerals, Dhahran, 31261, Saudi Arabia · SDAIA–KFUPM Joint Research Center for Artificial Intelligence, King Fahd University of Petroleum & Minerals, Dhahran, 31261, Saudi Arabia