On Disentangling Spoof Trace for Generic Face Anti-Spoofing

• Face spoofing presents faces from manufactured media such as photographs, screens, masks, or makeup, threatening mobile unlock, building access, and transportation security.Source: §1 · pp. 1–2
• Prior deep-learning methods often address only limited spoof types and return a classification score without clarifying the live–spoof differences underlying their decisions.Source: §1 · p. 2
• The paper defines a spoof trace as patterns distinguishing a spoof face from its live counterpart and seeks to disentangle traces despite their diversity and missing ground truth.Source: §1 · p. 2
• The authors propose trace disentanglement as a basis for classification, visual explanation, and synthesizing samples for minority spoof types.They state that traces may reveal attack properties such as severity and methodology.Source: §1 · pp. 2–3
• The paper concerns monocular face anti-spoofing methods.Source: §1 · p. 2
• STDN uses binary live/spoof labels to disentangle traces, reconstruct live counterparts, and synthesize new spoofs with geometric correction and multiscale discriminators.Synthesized samples also train the generator in a supervised fashion.Source: §1 · pp. 2–3

Face Anti-Spoofing

• Face anti-spoofing methods evolved from motion cues to handcrafted texture features and deep binary classifiers, with some methods adding depth-map or rPPG supervision.Recent work also explores few-shot or zero-shot detection and domain adaptation.Source: §2 · p. 3

Disentanglement Learning

• Prior GAN-based trace methods modeled traces as low-intensity noise, limiting them to print and replay attacks and producing unconvincing visual results.Source: §2 · p. 4
• This work represents spoof traces at multiple scales, including geometric information, using outputs from different layers, multiscale discriminators, and a warping layer.Source: §2 · p. 4

3.1 Problem Formulation

• STDN seeks the minimum change that transfers an input face to the live domain, representing spoof traces as global, smooth low-level, and high-level patterns.The representation uses color-range and balance biases, smooth content, and high-level texture.Source: §3.1 · pp. 4–5

3.2 Disentanglement Generator

• An Early Spoof Regressor predicts a live/spoof map from bottleneck features, and testing combines its output with spoof-trace intensity for classification.The ESR map has size 16×16; the generator processes both spoof and live inputs.Source: §3.2 · pp. 5–6

3.3 Reconstruction and Synthesis

• Shape-dependent traces can misalign across faces, so online differentiable warping interpolates sparse vertex offsets into dense offsets before traces are added to target live faces.The method uses selected 3D face vertices to compute sparse offsets, then triangulation interpolation to obtain dense offsets.Source: §3.3 · pp. 6–7, Figure 4 · p. 7
• Synthesized spoofs provide ground-truth traces for supervised generator training; weakening one trace element creates harder spoof samples that may encourage reliance on other elements.The method randomly zeros one of {s, b, C, T} during synthesis.Source: §3.3 · pp. 6–7

3.4 Multi-scale Discriminators

• Three discriminators operate at 256, 128, and 64 resolution, emphasizing fine texture, content patterns, and global elements, respectively.Each discriminator compares reconstructed live or synthesized spoof samples with real samples from the corresponding domain.Source: §3.4 · pp. 7–8

3.5 Training Steps and Loss Functions

• Each minibatch uses generator and discriminator updates followed by extra ESR and pixel supervision on live and synthesized-spoof samples.Generator updates use adversarial, ESR, and regularizer losses; the extra supervision step applies pixel and ESR losses.Source: §3.5 · pp. 8–9

4.1 Experimental Setup

• Experiments use Oulu-NPU, SiW, and SiW-M, with face-region crops and EER, ACER, and TDR at FDR=0.5%; SiW-M is split 60%/40% for training and testing.SiW-M includes 13 spoof types.Source: §4.1 · pp. 9–10

4.2 Anti-Spoofing for Known Spoof Types

• 30% and 32% ACER reductions over the best prior work on Oulu-NPU protocols 1 and 4 accompany the best performance across all four protocols.The comparison is shown in Table 1(a).Source: §4.2 · pp. 10–11, Table 1 · p. 10
• On SiW, STDN outperforms prior work on protocols 1–2 and is competitive on protocol 3; on SiW-M Protocol I, it outperforms across all spoof types and overall.SiW protocol 3 tests training on print or replay and testing on the other attack.Source: Table 1 · p. 10, §4.2 · p. 11

4.3 Anti-Spoofing for Unknown Spoof Types

• 24.8% overall EER and 15.5% overall ACER improvements over the previous best are reported for SiW-M unknown spoof detection.TDR is 35.7% at FDR=0.5%, which the authors say remains far from practical for unknown spoof types.Source: Table 2 · p. 11, §4.3 · p. 12

4.4 Spoof Traces Classification

• 92.0% (3-class) and 83.3% (5-class) trace-based accuracy improve over by 10% and 29%; original-image CNN accuracy is 86.3% and 80.6%, respectively.The results are compared in the left (3-class) and right (5-class) confusion matrices of Table 3.Source: §4.4 · pp. 12–13, Table 3 · p. 12

4.5 Ablation Study

• ACER falls from 2.6% for ESR to 2.1% with GAN and 1.6% with D-GAN; good-quality synthetic samples further reduce it to 1.1%, while bad-quality samples raise error to 4.5%.The authors also report higher-quality traces from ESR+D-GAN than ESR+GAN (Fig. 6).Source: Table 4 · p. 12, Figure 6 · p. 12, §4.5 · p. 13

4.6 Visualization

• Disentangled traces visualize color distortion and Moiré in print/replay, makeup details, and paper-mask folds, though some funny-eyeglasses traces are not convincing.For those difficult examples, the model still focuses on the correct region and disentangles parts of the traces.Source: Figure 7 · p. 13, §4.6 · p. 14
• Spoof traces transfer to a new face without changing its identity, and the 3D warping layer corrects geometric discrepancy during synthesis.Figure 8 shows source spoof samples, disentangled traces, target live faces, and synthesized spoofs.Source: §4.6 · p. 14, Figure 8 · p. 14

• STDN improves state-of-the-art anti-spoofing for both known and unknown attacks while providing visual evidence for decisions.Source: §5 · p. 14
• STDN disentangles spoof traces from faces, then uses them to reconstruct live faces and synthesize new spoofs.Source: §5 · p. 14
• A 3D warping layer corrects geometric discrepancy when synthesizing spoofs.Source: §5 · p. 14