Organizations: School of Information Science and Technology, Shijiazhuang Tiedao University, Shijiazhuang, China. · School of Integrated Circuit, Shanghai Jiao Tong University, Shanghai, China · Eastern Institute of Technology, Ningbo, Zhejiang, China. · School of Computing and Information Technology, Great Bay University, Dongguan, China. · the School of Integrated Circuits and Micro-Nano Electronics, Fudan University, Shanghai, China. · Tencent Youtu Lab, Shanghai, China.
Multimodal face presentation attack detection (fPAD) improves robustness using complementary RGB, depth, and infrared (IR) cues, but deployment remains challenging when modalities are unavailable or incomplete. Existing flexible-modality methods mainly use zero-filling or modality omission, potentially discarding cross-modal information and introducing modality-dependent bias. We propose the Flexible Modality Adaptation Network (FMA-Net), a unified framework that compensates for missing modalities through instruction-guided latent-space translation. The Face Modality Instruct Translator (FMIT) estimates target modality observations from RGB inputs, while context-aware category enhancement uses learnable text prompts as modality-invariant semantic anchors. A Facial Feature Adapter (F2A) integrates global facial structure and local spoof-relevant cues to enhance discriminative representations. Synthetic-to-Real Face Domain Alignment (SR-FDA) reduces feature discrepancies between translated and real modalities, with benefits depending on the missing-modality pattern and data domain. Experiments on WMCA and PADISI-Face validate the framework. On WMCA with a 70% modality missing ratio, FMA-Net achieves a mean ACER of 3.22% across four intra-domain missing-modality protocols, compared with 4.24% for the strongest baseline by the same criterion. When both depth and IR are entirely unavailable, it achieves 5.23% ACER, compared with 5.99% for the best competing result. Cross-domain experiments also reveal limitations of synthetic-to-real alignment under dataset shifts. These results support FMA-Net as a practical approach to fPAD with flexible and incomplete modalities.
Figures & tables
Fig. 1: Comparison between our method and existing flexible-modality face Presentation Attack Detection (fPAD) approaches under a missing infrared (IR) modality. (a) Existing methods address missing modalities via skipping or zero-padding. (b) FMA-Net employs a modality-translation module to reconstruct the missing modality, enabling all modalities to be used by the fPAD model for robust inference. (c) Comparative analysis of task settings: unimodal, multimodal, and flexible-modality, in terms of input-modality structure, data-acquisition complexity, and model capabilities. The source face images are from the WMCA dataset [ 1 ] .
Fig. 2: Overview of the proposed FMA-Net. (a) Learnable context and the text encoder capture face-related semantic features. (b) FMIT estimates missing modality observations in the latent space. (c) The F 2 A module extracts visual features from both real and translated multimodal inputs. (d) SR-FDA mitigates domain shift by encouraging domain-invariant feature learning. The source face images are from the WMCA dataset [ 1 ] .
Fig. 3: Overview of the modality translation instruction dataset construction and the face missing-modality compensation process. Steps 1–2 illustrate dataset creation through instruction design and paired data generation, while step 3 shows modality translation by FMIT for estimating missing modality observations. The source face images are from the WMCA dataset [ 1 ] .
Fig. 4: Illustration of the proposed Synthetic-to-Real Face Domain Alignment (SR-FDA). By leveraging adversarial learning between real and synthetic data domains, SR-FDA extracts domain-invariant features to alleviate domain shift introduced by generated data, while aligning visual and textual features for robust real vs. spoof classification. The source face images are from the WMCA dataset [ 1 ] .
Fig. 5: Illustration of flexible-modality translation and completion protocols: (a) depth-only missing, (b) IR-only missing, (c) symmetric missing of both depth and IR modalities, and (d) asymmetric missing of depth and IR modalities.
Method
Seen
Unseen
Mean ± Std
A1
A2
A3
A4
A5
A6
A7
MC-PixBiS [ 33 ]
1.80
49.70
3.70
0.70
0.10
16.00
0.20
3.40
10.54 ± 18.14
MCCNN-OCCL-GMM [ 34 ]
3.30
22.80
31.40
1.90
30.00
50.00
4.80
18.30
22.74 ± 16.57
CMFL [ 6 ]
1.70
12.40
1.00
2.50
0.70
33.50
1.80
1.70
7.66 ± 12.11
ViT [ 12 ]
2.71
11.95
1.44
3.78
0.00
18.02
0.58
4.43
5.74 ± 6.75
MA-ViT [ 3 ]
1.45
9.76
0.93
0.55
0.00
14.00
0.00
1.46
3.81 ± 5.67
TABLE I: Comparative experimental results of different attack protocols on the WMCA dataset (ACER,%)
Modality
Method
Seen
Unseen
Mean ± Std
A1
A2
A3
A4
A5
A6
A7
A8
A9
R
ViT [ 36 ]
0.28
22.33
0.27
0.54
0.54
0.27
9.49
41.18
3.99
1.80
8.93 ± 14.08
ViT-H [ 9 ]
0.14
24.66
0.27
0.41
0.27
0.27
13.12
32.36
3.04
1.67
8.45 ± 12.23
ViT-H-H [ 9 ]
0.14
18.04
0.27
0.54
0.14
0.14
9.91
10.17
5.14
2.07
5.16 ± 6.31
RD
ViT-E-E [ 9 ]
0.55
24.43
0.27
0.27
0.41
0.14
16.2
14.31
3.18
0.41
6.62 ± 9.22
ViT-E-H [ 9 ]
0.14
20.54
0.00
0.41
0.41
0.14
8.37
9.50
5.14
0.27
4.98 ± 6.95
TABLE II: Comparative experimental results of different attack protocols on the PADISI-Face dataset (ACER,%)
Method
P1 (RI)
P2 (RD)
P3 (R)
P4 (R)
α=70%
ViT [ 34 ]
10.61
5.86
8.70
6.37
Prompt [ 40 ]
8.70
5.52
5.84
7.21
CMFL [ 6 ]
6.53
4.95
5.08
5.37
CrossAtten [ 2 ]
8.47
5.29
8.33
7.15
VP-FAS [ 32 ]
3.67
3.91
4.54
4.83
TABLE III: Comparison results for the flexible-modality translation completion protocol on the WMCA dataset (ACER,%)
Method
P1 (RI)
P2 (RD)
P3 (R)
P4 (R)
α=70%
CoCoOp [ 37 ]
0.80
0.91
0.69
0.90
CoOp [ 38 ]
0.55
1.36
0.85
1.24
FMA-Net(ours)
0.48
0.74
0.74
1.43
α=100%
CoCoOp [ 37 ]
0.44
2.20
0.93
TABLE IV: Comparison results for the flexible-modality translation completion protocol on the PADISI-Face dataset (ACER,%)
Method
W-P
P-W
P1
P2
P3
P4
P1
P2
P3
P4
CoCoOp [ 37 ]
10.99
28.75
9.86
16.07
18.93
22.11
18.12
17.25
CoOp [ 38 ]
14.25
20.30
11.29
12.95
19.47
18.56
19.25
19.36
FMA-Net w/o SR-FDA (ours)
9.95
19.80
9.72
11.29
19.46
15.79
17.89
16.63
FMA-Net w/ SR-FDA (ours)
8.99
21.84
22.01
20.93
19.80
19.75
18.87
17.94
TABLE V: The cross-domain comparison results for flexible-modality translation completion with α=70% (ACER,%)
Method
W-P
P-W
P1
P2
P3&P4
P1
P2
P3&P4
ViT-E-E [ 9 ]
11.78
11.90
-
13.14
15.39
-
ViT-E-H [ 9 ]
11.29
8.50
-
12.55
12.05
-
MC-PixBiS-E-E [ 9 ]
20.06
16.68
-
15.90
15.51
-
MC-PixBiS-E-H [ 9 ]
13.82
15.37
-
15.63
14.20
-
CoCoOp [ 37 ]
10.87
28.75
16.07
18.93
19.12
17.89
TABLE VI: The cross-domain comparison results for flexible-modality translation completion with α=100% (ACER,%)
Prompt
F 2 A
SR-FDA
P1 (RI)
P2 (RD)
P3 (R)
P4 (R)
✓
3.93
2.92
5.79
5.88
✓
✓
2.86
3.16
4.46
4.41
✓
✓
✓
2.15
2.78
3.88
4.05
TABLE VII: Ablation study of Prompt Learning, F 2 A, and SR-FDA under the flexible-modality translation completion protocol with α=70% (ACER,%)
Prompt
F 2 A
SR-FDA
P1 (RI)
P2 (RD)
P3&P4 (R)
✓
2.15
2.65
5.99
✓
✓
2.11
2.03
5.87
✓
✓
✓
2.55
1.70
5.23
TABLE VIII: Ablation study of Prompt Learning, F 2 A, and SR-FDA under the flexible-modality translation completion protocol with α=100% (ACER,%)
Fig. 6: Effect of the test-time completion percentage β on WMCA when 70% of test samples have missing auxiliary modalities. The full detector, including SR-FDA, is fixed. β denotes the percentage of incomplete samples whose missing modalities are completed; β=0 uses only available modalities and β=100 completes all missing modalities in that subset.
Fig. 7: Comparison of CLIP and FMA-Net feature-attention heatmaps. Top is CLIP and bottom is FMA-Net. FMA-Net exploits a richer set of discriminative cues. The source face images are from the WMCA dataset [ 1 ] .
Fig. 8: Examples of modality translation. Column 1: ground-truth RGB face images, Columns 2 and 4: corresponding ground-truth IR and depth images, Columns 3 and 5: IR and depth images synthesized by FMIT from the RGB input. The source face images are from the WMCA dataset [ 1 ] .
Face Presentation Attack Detection identifies spoofing attempts, such as printed, replayed, masked, or digitally altered faces, targeting face recognition systems.