Anomaly Detection-Based Unknown Face Presentation Attack Detection

This text is study for arxiv.org/pdf/2007.05856.pdf

Overview

• Implement face presentation attack detection based on anomaly detection
• Propose a novel end-to-end trainable face presentation attack detection model based on one-class CNN

Comparision with OC-CNN

• OC-CNN
  
  • Use samples from Gaussian distribution centered at the origin with a small standard deviation as the pseudo-negative class
  • Anomaly data is visually distinct as compared to the training class
• Proposed model
  • Attacked samples during testing are very similar to the non-attacked samples used during training
  • Propose an adaptive mean estimation strategy to generate pseudo-negative data for training
    • Make sure that pseudo-negative samples lie in the proximity of bonafide presentation class features

 

Proposed model

• Feature Extractor
• One Class Classifier
  • Utilize an adaptive startegy to estimate the mean of pseudo-negative Gaussian distribution
    • Define a pseudo-negative distribution N(μ',σ)
    • Consider displacement of sample feature mean of bonafide presentation data across two iterations when defining μ'
    • μ_new be the mean of featue vectors, μ_old be the mean of features of the previous batch
    • Mean of pseudo-negative class μ' is calculated as [1]
  • Training 
    • Use cross-entropy loss defined as [2]
    • Utilize pairwise confusion (PC) loss [3]
      • Eliminate the identity information from pre-trained feature spaces
    • Overall loss function is [4]
      • Model can be trained as end-to-end manner

[1]

[2]

[3]

[4]

Datasets and Evaluation Protocols

• Replay Attack
  • fPAD dataset of 1300 videos consisting of photo and replay attacks
  • Extract 30-40 frames per video with a gap of atleast 10 frames in between to carry out out experiments
• Rose-Youtu