ZICQ
中 Log in / Sign up
Newsroom Research & Papers #Face Presentation Attack Detection #Video PAD #Long-Range Benchmark #Biometric Security #EfficientNet

Long-Range Video Face PAD Benchmark Spans 1–100 Meters and Exposes Sensor-Resolution Limits

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:The paper introduces a long-range, video-based face presentation attack detection dataset and three-class benchmark covering 261 identities and more than 374,000 frames. It spans five capture distances from 1 to 100 meters, indoor and outdoor conditions, and both DSLR and mobile sensors. EfficientNet reports the best aggregate result at 98.48% accuracy and 2.01% ACER, although the lower-resolution mobile sensor degrades sharply beyond 50 meters, with MobileNetV3 reaching only 55% accuracy for th


Core Findings

The paper introduces a long-range, video-based face presentation attack detection dataset and three-class benchmark. It covers 261 identities and more than 374,000 frames across five capture distances from 1 to 100 meters, indoor and outdoor settings, and two imaging devices: a DSLR camera and a lower-resolution mobile sensor. The labels separate live subjects, spoofing attacks, and props or occlusions. Models are trained and evaluated on video clips rather than isolated frames.

Four CNN families—MobileNetV3, MobileNet, EfficientNet, and ResNet-50—are fine-tuned under the benchmark. EfficientNet reports the best aggregate result at 98.48% accuracy and 2.01% ACER. The higher-resolution DSLR retains a near-perfect spoof-classification rate at every tested distance. By contrast, the mobile sensor degrades sharply beyond 50 meters, with MobileNetV3 reaching only 55% accuracy for the props class at 75 meters. Generalization tests on unseen attack datasets report 89–94% accuracy, providing preliminary evidence of cross-domain robustness.

Why Distance Changes the PAD Task

Face PAD depends on fine-grained cues such as print texture, screen artifacts, edge behavior, reflectance, and temporal inconsistency. Before a classifier can recover those cues, the target must occupy enough sensor pixels. For a face of physical width W, camera focal length f, and capture distance R, its approximate projected width is:

p ≈ fW/R

With fixed optics, moving from 1 meter to 100 meters can reduce the face width to roughly one hundredth of its original size. No amount of classifier input resolution can reconstruct texture that was never captured. Mobile devices also tend to combine smaller sensors, stronger compression, and tighter compute budgets, which explains why their long-range degradation cannot be fixed merely by enlarging the neural-network input.

Video clips can potentially exploit temporal stability, motion, blink patterns, or replay inconsistencies. However, the abstract does not disclose whether temporal information is implemented with 3D convolutions, optical flow, recurrent aggregation, or late fusion across sampled frames. It also omits clip length and frame-sampling rates. If neighboring frames are highly correlated, more than 374,000 frames do not represent 374,000 statistically independent test cases.

The three-class formulation also creates an operational distinction that binary PAD does not expose. A prop or occlusion is not necessarily a presentation attack. A traveler holding a passport, phone, or screen might need a new capture or a different decision rather than an automatic spoof rejection. The operational mapping of the props class therefore matters as much as its measured accuracy.

Model and Engineering Trade-offs

| Approach | Reported result | Engineering implication | |---|---|---| | MobileNetV3 | Mobile-sensor props accuracy falls to 55% at 75 meters | Suitable for constrained edge inference, but vulnerable to missing long-range detail | | MobileNet | Evaluated on the same video-clip protocol | Provides a lightweight baseline; per-distance results and inference cost are not disclosed | | EfficientNet | Best overall result at 98.48% accuracy and 2.01% ACER | Best reported accuracy does not establish best deployment efficiency without FLOPs, latency, and input-size data | | ResNet-50 | Included as a residual-network baseline | Helps separate dataset effects from architecture-specific gains, but speed, memory, and ablations remain unavailable |

The DSLR-versus-mobile comparison illustrates a direct accuracy-versus-deployment-cost trade-off. Higher-resolution imaging preserves more distant facial evidence but increases bandwidth, storage, lens requirements, and compute cost. Mobile hardware scales more easily across a fleet but struggles at long range. A subtle implementation issue is that higher sensor resolution provides little benefit if the full frame is resized to a fixed network input before inference. A practical pipeline should detect and track the face first, allocate an explicit pixel budget to the ROI, and preserve source detail during crop, resize, and temporal sampling.

How to Interpret the Benchmark

The reported 98.48% accuracy is encouraging, but the abstract does not disclose whether identities are disjoint across splits, whether complete capture sequences can leak between training and test sets, the per-class sample counts, threshold calibration, or run-to-run variance. Leakage through adjacent frames or shared identities would inflate long-range generalization. The effective split unit should therefore be the identity and complete recording session, not the individual frame.

ACER generally averages attack-class and bona fide-class error rates, making it useful for balancing false acceptance and false rejection in binary PAD. The paper uses three labels, however, and the abstract does not explain how the props class is mapped into ACER. Aggregate accuracy can also hide class imbalance and local failures at particular distances or with particular sensors. A fuller evaluation should include APCER, BPCER, per-distance and per-sensor confusion matrices, confidence intervals, and attack-specific error rates.

The 89–94% results on unseen attack datasets are useful cross-domain evidence, but the abstract does not state whether identities, devices, and thresholds were reconfigured for those tests, or whether the unseen sets share environmental conditions with training. With only 261 identities, generalization across demographic groups remains uncertain. The summary also omits training cost, inference latency, memory use, code availability, and immediate dataset availability. These numbers should therefore be treated as preprint-reported results rather than independently reproduced benchmarks.

Deployment Guidance

  1. Route by distance and sensor class: Prefer mobile sensors for short-range capture and higher-resolution or telephoto imaging for distant subjects. Use an estimated distance to select the model and input path.
  2. Build an ROI and pixel-budget pipeline: Preserve native resolution before cropping, record the target's pixel dimensions, and detect blur, overexposure, and compression degradation. Low-quality frames should trigger recapture rather than forced classification.
  3. Decide from temporal evidence: Apply quality gating and temporal-consistency checks over consecutive frames, then verify that live, spoof, and props decisions each receive a measurable benefit from video rather than frame sampling alone.
  4. Calibrate thresholds by domain: Maintain separate acceptance thresholds for distance bands, sensors, attack media, and deployment environments, with an explicit unknown-attack or low-quality rejection path.
  5. Report security and fairness metrics: Add per-identity errors, attack-specific false acceptance, bona fide false rejection, confidence intervals, and performance across demographic groups.
  6. Treat face video as sensitive biometric data: The planned EULA release does not replace consent, retention, access-control, and deletion policies for face recordings.

Conclusion

The main contribution is not a new PAD architecture but a benchmark that treats capture distance, sensor resolution, and video sequence modeling as first-class evaluation variables. The results show that strong near-range accuracy cannot be extrapolated to 100 meters, and sensor differences may affect deployment outcomes more strongly than backbone choice. If the dataset, protocol, code, and independent reproductions become available, this benchmark could become a useful foundation for long-range biometric-security evaluation. Until then, it is best treated as a research instrument rather than a production deployment conclusion.


Source: ArXiv Machine Learning (cs.LG) (2026-11-20)

— END —

Tags: #Face Presentation Attack Detection #Video PAD #Long-Range Benchmark #Biometric Security #EfficientNet

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…