Controlled Approaching Face Protocol

• The protocol instructs users to slowly move their frontal face toward the camera, capturing three key frames to reveal facial volumetric changes.
• It leverages RAFT-based optical flow and a dual ResNet18 architecture to fuse motion cues with RGB appearance for robust liveness detection.
• Strict face-size checkpoints and background motion clipping standardize the capture process, improving discrimination between real faces and presentation attacks.

Searching arXiv for the primary paper and closely related face liveness detection work. The Controlled Approaching Face Protocol is a cooperative video capture protocol for face liveness detection in which a participant is instructed to slowly move a frontal-oriented face closer to the camera while the system extracts a small set of key frames and analyzes the resulting optical flow together with RGB appearance cues. In the published literature, the term is explicitly introduced as the core interaction scenario of a video-based presentation attack detection method that seeks to make facial volumetric change observable in the motion field, thereby improving discrimination between genuine faces and presentation attacks such as printed photos, screen displays, masks, and video replays (Sokolov et al., 14 Aug 2025). The protocol is defined operationally by a constrained trajectory along the camera’s optical axis, checkpointed face-size targets within the image frame, and a restart rule when the motion violates the specified progression. A broader implication is that the protocol converts liveness detection from a largely passive appearance-classification problem into a partially standardized spatiotemporal inference problem, in which motion itself is instrumented as evidence of three-dimensional structure (Sokolov et al., 14 Aug 2025).

1. Definition and scope

In its established usage, the Controlled Approaching Face Protocol denotes a cooperative user interaction scenario for video-based face liveness detection rather than a general face-generation or geometric-control framework. The defining instruction is that participants are “instructed to slowly move their frontal-oriented face closer to the camera”, with the camera interface guiding the movement using two target squares and a real-time face detector (Sokolov et al., 14 Aug 2025). The start condition is a face height of approximately 50% of the frame, the end condition is approximately 75%, and an intermediate checkpoint at approximately 62.5% is used to form a three-frame representation of the sequence: $f_1$, $f_2$, and $f_3$ (Sokolov et al., 14 Aug 2025).

This protocol is not a generic synonym for any method involving controlled faces. The literature supplied here contains other face-related uses of “controlled” settings—such as attribute-controlled synthesis from line drawings (Guo et al., 2017) and identity-preserved personalization framed in control terms (Yu et al., 2024)—but these are distinct research problems. Likewise, one supplied systems paper discusses a “controlled approach” to geometric “faces” in a catastrophe-theoretic software architecture, yet it explicitly does not use the exact phrase “Controlled Approaching Face Protocol” (Niznik, 2014). This distinction matters because the protocol’s technical meaning is tied specifically to presentation attack detection and to the motion geometry of a subject approaching a camera (Sokolov et al., 14 Aug 2025).

2. Interaction protocol and capture mechanics

The protocol’s user interface overlays two target squares on the camera view and displays a blue bounding box from a real-time face detector (Sokolov et al., 14 Aug 2025). The participant is instructed to align the detected face to the red reference square at 50% of the frame height, then to slowly move the face closer to the camera until it fills an enlarged target square at 75% of the frame height. If the face moves away from the camera or disappears, the recording restarts from the beginning (Sokolov et al., 14 Aug 2025).

The motion trajectory is explicitly constrained to lie along the camera’s optical axis, and the paper specifies frontal orientation throughout the interaction (Sokolov et al., 14 Aug 2025). No numeric speed thresholds are imposed; instead, speed variability is handled through data augmentation during training. The method relies only on the three extracted frames, not on a fixed frame rate or a reported capture duration. Environmental conditions are intentionally broad: the private dataset spans diverse devices and lighting conditions, and the only explicit practical requirement is maintaining a visible frontal face (Sokolov et al., 14 Aug 2025).

The key design choice is standardization. By forcing the face to move in a controlled axial approach rather than allowing unconstrained motion, the protocol makes the expected motion field more predictable and more directly linked to 3D facial structure. This suggests that the protocol functions simultaneously as a data-acquisition constraint and as a prior over the dynamics that the classifier is expected to observe.

3. Optical-flow rationale and volumetric cue extraction

The protocol is motivated by the claim that an actively approaching face generates a characteristic radial expansion optical-flow pattern tied to true 3D structure, whereas planar or fixed-depth presentation attack instruments (PAIs) do not exhibit the same depth-dependent expansion (Sokolov et al., 14 Aug 2025). The paper gives the standard brightness-constancy formulation

$I(x, y, t) = I(x + u, y + v, t + \Delta t)$

and the classical optical-flow constraint

$I_x u + I_y v + I_t = 0.$

For a perspective camera with image coordinates

$x = fX/Z,\qquad y = fY/Z,$

and rigid approach along the $z$-axis, the induced image motion is written as

$$u = \dot{x} = -\frac{\dot{Z}}{Z}x,\qquad v = \dot{y} = -\frac{\dot{Z}}{Z}y,$$

yielding positive divergence under approach:

$$\nabla \cdot \mathbf{u} = \frac{\partial u}{\partial x} + \frac{\partial v}{\partial y} = -2\frac{\dot{Z}}{Z}.$$

The stated interpretation is that a genuine 3D face exhibits fine-grained, spatially varying flow magnitudes across facial features and smooth face-background boundaries, while a strictly planar PAI produces motion dominated by global affine expansion/translation with less intra-face variation and sharper edges at boundaries after stabilization (Sokolov et al., 14 Aug 2025).

In implementation, the protocol uses RAFT to estimate optical flow between preprocessed $f_1$ and $f_2$0 (Sokolov et al., 14 Aug 2025). The flow is converted to a single-channel magnitude, defined in the paper as

$f_2$1

and values greater than 20% of the input crop side are clipped:

$f_2$2

For $f_2$3 inputs, the clipping threshold is therefore 51.2 pixels (Sokolov et al., 14 Aug 2025). The stated purpose of clipping is to suppress background motion, which the protocol treats as non-informative.

4. Pipeline architecture and information flow

The end-to-end pipeline consists of frame capture, preprocessing, optical flow estimation, flow processing, and binary classification (Sokolov et al., 14 Aug 2025). During capture, the system extracts $f_2$4 at approximately 0.500 face height, $f_2$5 at approximately 0.625, and $f_2$6 at approximately 0.750. These are the only frames used downstream (Sokolov et al., 14 Aug 2025).

Preprocessing has four steps for $f_2$7 and $f_2$8: detect facial key-points, align by shift and rotation, crop the face region with a 10% margin around the bounding box, and resize all crops to $f_2$9 (Sokolov et al., 14 Aug 2025). The frame $f_3$0 undergoes identical processing except that the alignment step is omitted. Optical flow is then estimated with RAFT between the preprocessed $f_3$1 and $f_3$2, using $f_3$3 resolution and 3 refinement iterations as the selected speed-quality trade-off (Sokolov et al., 14 Aug 2025).

Classification uses a dual-backbone ResNet18 architecture with a fully connected fusion layer (Sokolov et al., 14 Aug 2025). The flow backbone consumes the clipped flow magnitude as a $f_3$4, 1-channel input, while the RGB backbone consumes the preprocessed $f_3$5 RGB frame as a $f_3$6, 3-channel input. Their outputs are fused for a binary real/spoof prediction (Sokolov et al., 14 Aug 2025).

Training uses joint optimization of both backbones, with the option that the RGB backbone can be pre-trained separately on larger single-shot datasets. Three optical-flow-specific augmentations are reported: random frame, multi-resolution, and perspective augmentation (Sokolov et al., 14 Aug 2025). The paper does not explicitly specify the training loss. It notes that binary cross-entropy is the conventional choice for a two-class PAD classifier, but does not report it as the implemented loss (Sokolov et al., 14 Aug 2025).

5. Presentation attack detection behavior

The protocol is evaluated against five listed presentation-attack categories: Screen Photos, Printed Photos, Printed Masks, Dynamic Videos, and Static Videos (Sokolov et al., 14 Aug 2025). Its central claim is that the controlled approach amplifies discriminative motion cues that are weak or ambiguous under passive capture.

The paper describes the observable flow differences qualitatively. Real faces and dynamic replays exhibit higher, spatially structured magnitudes in the head region, smooth face-background boundaries, and 3D face patterns. Flat masks exhibit near-zero flow in the face region and high background magnitudes. Screen/printed photos exhibit near-zero magnitude within a square region around the face, interpreted as evidence of negligible depth-driven expansion after stabilization (Sokolov et al., 14 Aug 2025). The decision itself is not rule-based: it is learned by the dual-stream classifier, which combines textural cues in $f_3$7 with motion cues in the flow magnitude map.

The most difficult attack type is explicitly reported to be Dynamic Videos, because they mimic the same approaching motion scenario (Sokolov et al., 14 Aug 2025). This is an important clarification against a possible misconception that any motion-based PAD protocol automatically defeats replay attacks. The paper instead shows that optical flow alone is insufficient for strong replay robustness, and that the system’s high performance depends on fusion of flow and RGB rather than motion analysis in isolation (Sokolov et al., 14 Aug 2025).

6. Quantitative results, ablations, and runtime

The dataset is a private dataset collected under the approaching face scenario, with train/test splits reported as Real: 1860/291, Screen Photos: 4623/217, Printed Photos: 1679/85, Printed Masks: 1699/93, Dynamic Videos: 870/170, and Static Videos: 765/218 (Sokolov et al., 14 Aug 2025). The evaluation metric is ROC AUC per spoof subset vs. real (Sokolov et al., 14 Aug 2025).

The paper reports several ablations showing how the protocol interacts with preprocessing and model design.

Setting	Key result
Optical flow processing	Clipped OF magnitude improves Dynamic Videos from 0.670 to 0.784 relative to unclipped magnitude
Architecture comparison	Dual ResNet18 fusion achieves Screen 0.999, Printed 1.000, Masks 0.993, Dynamic 0.994, Static 0.999
Comparison methods	Proposed method reports runtime 0.55 s/sample without preprocessing

For optical flow processing, the reported AUCs are: Raw OF — Screen 0.992, Printed 0.990, Masks 0.953, Dynamic 0.668, Static 0.958; OF magnitude — Screen 0.991, Printed 0.991, Masks 0.955, Dynamic 0.670, Static 0.961; Clipped OF magnitude — Screen 0.994, Printed 0.993, Masks 0.965, Dynamic 0.784, Static 0.965 (Sokolov et al., 14 Aug 2025). The paper attributes the Dynamic-Videos improvement to background suppression.

For optical flow augmentations, the reported AUCs are: No augmentations — Screen 0.994, Printed 0.993, Masks 0.965, Dynamic 0.784, Static 0.965; Random frame — Screen 0.998, Printed 0.997, Masks 0.972, Dynamic 0.666, Static 0.972; Random frame + multi-resolution — Screen 0.998, Printed 0.996, Masks 0.972, Dynamic 0.675, Static 0.972 (Sokolov et al., 14 Aug 2025). The paper states that random frame yields large gains and multi-resolution provides stability, though the Dynamic-Videos figures remain lower than in the best fused setting.

For architecture comparison, the results are more decisive. OF only; ResNet18 gives Screen 0.998, Printed 0.996, Masks 0.972, Dynamic 0.675, Static 0.972. OF + RGB (stacked) single ResNet18 gives Screen 0.996, Printed 0.994, Masks 0.985, Dynamic 0.991, Static 0.998. OF + RGB (separately) dual ResNet18 gives Screen 0.999, Printed 1.000, Masks 0.993, Dynamic 0.994, Static 0.999 (Sokolov et al., 14 Aug 2025). This is the core empirical result supporting the protocol’s practical formulation: the controlled motion is most effective when paired with a separate RGB pathway rather than used alone.

The paper also compares the method with alternative aggregators. Single-shot ResNet18 ($f_3$8) reports Screen 0.983, Printed 0.963, Masks 0.970, Dynamic 0.991, Static 0.985; runtime 0.05 s/sample. Stabilized average (five frames; improved key-point stabilization) reports Screen 0.999, Printed 0.980, Masks 0.922, Dynamic 1.000, Static 1.000; runtime 0.05 s/sample. Rank pooling + OF reports Screen 0.993, Printed 0.982, Masks 0.904, Dynamic 0.921, Static 0.976; runtime 0.75 s/sample. The proposed method (dual ResNet18 with RAFT OF) reports Screen 1.000, Printed 1.000, Masks 0.994, Dynamic 0.996, Static 0.999; runtime 0.55 s/sample (without preprocessing) (Sokolov et al., 14 Aug 2025).

7. Robustness, limitations, and deployment considerations

The protocol improves robustness to some nuisance factors, but the paper also documents clear limitations. Head rotations are handled through key-point stabilization and perspective augmentation, and the reported AUCs with Perspective augmentations are Screen 1.000, Printed 1.000, Masks 0.994, Dynamic 0.996, Static 0.999 (Sokolov et al., 14 Aug 2025). However, the test split is stated to contain few strong rotations, so the rotation robustness evidence is limited. Motion blur is a more significant weakness: at High blur, AUC drops to Screen 0.937, Printed 0.883, Masks 0.833, Dynamic 0.846, Static 0.857 (Sokolov et al., 14 Aug 2025).

The protocol also has a nontrivial cooperation burden. It requires the participant to follow instructions, maintain frontal orientation, remain visible, and perform a short forward motion. The paper notes possible accessibility concerns for users with mobility constraints, neck issues, or devices with fixed mounting (Sokolov et al., 14 Aug 2025). Occlusions are not specifically quantified, but the method relies on face key-points and on flow extraction, so degradation is plausible when large parts of the face are obscured. This suggests that the protocol is not simply a drop-in replacement for passive PAD, but part of a design trade-off between user effort and stronger volumetric evidence.

From a security standpoint, the paper treats Dynamic Videos and high-quality 3D masks as the strongest threats (Sokolov et al., 14 Aug 2025). It explicitly evaluates Printed Masks, not molded 3D silicone masks. It also notes that robotic rigs simulating approach or high-quality displays and synchronized camera/display motion could narrow the margin between genuine and spoof signatures. The countermeasures proposed are procedural and architectural: strict enforcement of the controlled protocol, dual-stream fusion, stabilization via key-points and perspective augmentation, and background suppression via magnitude clipping (Sokolov et al., 14 Aug 2025).

For deployment, the recommended parameters are concrete: begin at approximately 50% face height, end at approximately 75%, capture $f_3$9, $I(x, y, t) = I(x + u, y + v, t + \Delta t)$0, $I(x, y, t) = I(x + u, y + v, t + \Delta t)$1 at approximately 0.500, 0.625, 0.750, preprocess with 10% margin and $I(x, y, t) = I(x + u, y + v, t + \Delta t)$2 resizing, run RAFT with 3 refinement iterations, convert to magnitude and clip above 51.2 px, then classify with dual ResNet18 fusion (Sokolov et al., 14 Aug 2025). Runtime is reported as approximately 0.55 s/sample on an Intel i5-11600K CPU, without preprocessing (Sokolov et al., 14 Aug 2025). A plausible implication is that the protocol is operationally feasible for interactive verification, but only when the application can tolerate modest latency and explicit user guidance.

In current usage, then, the Controlled Approaching Face Protocol is best understood as a cooperative PAD protocol that standardizes forward facial motion to expose 3D volumetric cues in optical flow, then combines those cues with RGB appearance for robust real/spoof discrimination (Sokolov et al., 14 Aug 2025). Its empirical value lies not in motion alone, but in the joint design of interaction protocol, motion geometry, stabilized preprocessing, and dual-stream classification.