Face Presentation Attack Detection by Excavating Causal Clues and Adapting Embedding Statistics

Meiling Fang, Naser DamerWACV 2024

TL;DR

This paper tackles the challenge of face presentation attack detection (PAD) in real-world scenarios where the model encounters unseen data distributions. They propose a novel approach, CF-PAD, that leverages causal inference and a class-guided style mixing technique to improve generalization without requiring extra trainable parameters or complex architectures. This leads to more robust and efficient PAD systems.

What Problem Does This Paper Address?

Face recognition systems are vulnerable to "presentation attacks" (PAs), where attackers use fake faces (e.g., printed photos, replayed videos, 3D masks) to fool the system. Face Presentation Attack Detection (PAD) is crucial for securing these systems. While deep learning has improved PAD performance, it often degrades significantly when faced with data from different sources (e.g., different cameras, lighting conditions, or attack types) than what it was trained on. Existing methods like domain adaptation (DA) require access to unlabeled target data, which is often unavailable, while domain generalization (DG) methods typically rely on prior knowledge of domain labels. Furthermore, many current DA/DG PAD methods are computationally intensive due to complex model architectures or multi-stage training. This paper aims to develop a more efficient and generalizable PAD method that doesn't require prior domain knowledge or target data.

Methods

The core of CF-PAD lies in two novel components:

  • Class-guided MixStyle (CGMixStyle): Inspired by the MixStyle technique, this method enhances the diversity of training data by mixing feature statistics. However, to avoid confusing real and fake faces, the shuffling of features is restricted within the same class (bona fide or attack). This ensures that style augmentation enriches the feature distribution without compromising the discriminative information needed for PAD.
  • Counterfactual Intervention (CI): This component frames PAD as a causal reasoning problem. The authors propose to uncover the "causal factors" in high-level features that truly contribute to the PAD decision, disentangling them from spurious or non-causal factors (like identity or capture environment). They achieve this by introducing three simple interventions (random zeroing, random shuffling, and random replacement of features) during training. By comparing the model's prediction with and without these interventions, the model learns to rely on causal features, thus improving generalization.

Crucially, both CGMixStyle and CI are designed to introduce no extra trainable parameters and negligible computational overhead during training, and are removed entirely during inference, making the method highly efficient.

Datasets

The paper evaluates CF-PAD on five publicly available face PAD datasets:

  • MSU-MFSD (M)
  • CASIA-FASD (C)
  • Idiap Replay-Attack (I)
  • OULU-NPU (O)
  • CelebA-Spoof (CA)

The experiments are conducted across various cross-dataset scenarios, including multi-source, limited-source, and single-source evaluations, following established protocols in the field. No new datasets were introduced.

Key Insights and Ideas

  • Causality for Generalization: The paper highlights that focusing on causal relationships, rather than just statistical correlations, is key to achieving robust generalization in PAD. By intervening on features and observing the impact, the model can learn to ignore spurious correlations.
  • Class-Guided Augmentation is Crucial: Simply mixing styles (like in standard MixStyle) can be detrimental to PAD because it might mix bona fide and attack features. Restricting this mixing to within the same class (CGMixStyle) is a critical insight for effective data augmentation in this domain.
  • Efficiency Through Simplicity: The proposed CGMixStyle and CI components are remarkably efficient, requiring no additional trainable parameters and minimal computational cost during training, and are completely removed during inference. This is a significant advantage over many existing complex PAD methods.
  • Approximating Intervention Risk Minimization: The theoretical framework of minimizing intervention risk is complex. The paper shows that enriching data diversity (CGMixStyle) and performing simple interventions (CI) effectively approximates this ideal, leading to better generalization.

Key Results

  • State-of-the-Art Performance: CF-PAD achieves competitive and often superior performance compared to existing state-of-the-art PAD methods across various cross-dataset evaluation protocols, particularly in challenging scenarios. For instance, in the I&C&M → O case, CF-PAD achieved an HTER of 9.87% and an AUC of 95.13%.
  • Significant Improvement over Baseline: The full CF-PAD method (combining CGMixStyle and CI) significantly outperforms a baseline model (without these components), showing a drop of 5.84 percentage points in HTER and an enhancement of 5.35 percentage points in AUC in multi-source evaluations.
  • Computational Efficiency: CF-PAD demonstrates the lowest computational complexity (FLOPs) among the compared methods while maintaining high performance, as illustrated in Figure 1 and Table 8.
  • Robustness: Ablation studies show that both CGMixStyle and CI individually contribute to improved performance, but their combination in CF-PAD yields the best results. The diversity of counterfactual interventions also proves beneficial.

Limitations and Future Work

  • Practicality of Causal Factors: The paper acknowledges that generating "true" counterfactual features is impractical due to the unobservable and unformulated nature of causal factors. The proposed interventions are approximations.
  • Hyperparameter Sensitivity: While the paper investigates the sensitivity of CI hyperparameters, further exploration might be beneficial.
  • Future Work: The authors suggest that their work provides a theoretical basis for future research in causal inference for PAD. The effectiveness of CGMixStyle and CI components in other computer vision tasks could also be explored.

Summary updated

BibTeX
@inproceedings{fang2024face,
  author = {Meiling Fang and Naser Damer},
  title = {{Face Presentation Attack Detection by Excavating Causal Clues and Adapting Embedding Statistics}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure