GaitSpoofNet: advanced spatio-temporal architectures for vision-based presentation attack detection.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Gait recognition is a promising biometric, but vulnerable to presentation attacks (PAs).
• Effective gait anti-spoofing (Presentation Attack Detection - PAD) is crucial for secure gait-based authentication.
• Existing research lacks dedicated deep temporal models and standardized benchmarks for vision-based gait PAD.

Purpose of the Study:

• To address the lack of deep temporal models for vision-based gait PAD.
• To establish a standardized benchmark for gait anti-spoofing using the CASIA-B dataset.
• To comparatively evaluate advanced spatio-temporal architectures for gait PAD.

Main Methods:

• Repurposed the CASIA-B dataset to create a standardized vision-based PAD baseline.
• Evaluated Mamba Selective State Space Model (mamba-ssm), a custom Mamba architecture, GRU, and LSTM networks.
• Utilized a robust CNN backbone for all evaluated temporal models.

Main Results:

• All investigated temporal models significantly improved gait spoofing detection compared to baselines.
• In open-access environments, the GRU-based model achieved 0.9840 accuracy and 0.9983 ROC-AUC.
• Under restricted-access conditions, the LSTM-based model demonstrated the strongest performance.

Conclusions:

• Advanced temporal models offer significant improvements in gait anti-spoofing.
• GRU and LSTM networks are highly effective for vision-based gait PAD in different scenarios.
• The study provides a standardized benchmark and evaluation framework for future gait anti-spoofing research.